Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• Meta expanded its Richland Parish campus in Louisiana to 5GW of compute capacity, the largest capacity figure among July’s five regional anchor deals. Meta is recorded as its own anchor customer, and the campus is under construction rather than proposed. Committed investment exceeds $50 billion, with a further sum for roads, water and wastewater held outside that figure.

• SK Telecom committed $91.5 billion, the equivalent of KRW140 trillion, to more than 2GW of AI data center capacity across Yeongnam in South Korea. SK Telecom describes the programme as the first major stage of a national platform targeting 15GW by 2035. No anchor customer is disclosed, and investment size, timing and financing structure remain subject to refinement.

• Nscale and Nordkraft formed Nordscale Operations AS to run the 230MW Narvik campus, the only one of July’s five regional anchor deals that names a generation source. Nscale Norway holds 51 percent of the operating company and Nordkraft holds 49 percent. The underlying Kvandal facility is designed on hydropower, with OpenAI recorded as the initial offtaker.

Top 5 Deals — Deep Dives

Five transactions anchor the month, with Entergy’s power arrangement among the most significant. Its supply portfolio combines new natural gas generation, nuclear uprates, batteries and purchased electricity to serve large, steady loads without relying on a single technology. The nuclear up rates are particularly important, adding firm capacity without new plant siting. Taken together they describe five different answers to the same question, which is how you get compute built where the power actually is.

Meta Richland Parish 5GW Expansion (North America)

Meta expanded the scale and the investment commitment of its Louisiana campus this month, taking Richland Parish to 5GW of compute capacity with more than $50 billion committed.

The status matters as much as the number. This is recorded as under construction, not proposed, which separates it from most of what sits above 1GW in the July log.

The power arrangement is the key feature.

Entergy’s supply portfolio combines new natural gas generation, nuclear uprates, batteries and purchased electricity to serve a large, steady load without relying on a single technology.

That is why the tracker records the source as grid rather than a specific fuel.

The nuclear uprates are particularly important because they add firm capacity without requiring new plant siting, a meaningful advantage for a 5GW load.

On capital structure, Meta is its own anchor customer.

There is no third-party tenant, no lease, no offtake contract to underwrite.

The company is funding the campus against its own demand, and the more than $1 billion of roads, water and wastewater work sits outside the headline figure.

That combination, self-anchored demand and a regulated utility building the supply portfolio around it, is available to a very short list of companies.

What Richland Parish reveals about North America is that the region’s largest credible capacity now moves through utility planning rather than through developer speculation.

The campus is not competing for grid headroom that already exists. Entergy is building the headroom to fit it.

Nscale–Nordkraft Nordscale Operations, Narvik (Europe)

The Narvik transaction is the smallest of the five and the most instructive.

Nscale and Nordkraft created Nordscale Operations AS, a jointly owned company that will manage Nscale’s AI data center facilities in the Narvik region, with Nscale Norway holding 51 percent and Nordkraft 49 percent.

The company was announced in early July and is expected to begin operations during 2026. No deal value was disclosed.

The underlying asset is a 230MW hydropower-backed campus at Kvandal, with OpenAI as the initial offtaker. A further 290MW of potential expansion is excluded.

The key feature is the structure: Nscale gave regional utility Nordkraft 49 percent of the operating company, aligning the power provider directly with the facility’s performance.

In a European market where interconnection queues and local opposition decide which projects survive, that is a structural move rather than an administrative one.

Nscale did not buy power. It bought alignment with the entity that supplies it, and it paid in ownership.

What Narvik reveals about Europe is that the projects clearing the bar are the ones that solved power before they solved anything else.

It is also the only anchor deal this month that names its fuel, which says something uncomfortable about the disclosure standard everywhere else.

SK Telecom $91.5 Billion Yeongnam Buildout (APAC)

SK Telecom outlined a plan to build more than 2GW of AI data center capacity across Yeongnam, with $91.5 billion attached, the equivalent of KRW140 trillion.

The programme starts with a 100MW facility in Ulsan, adds a 900MW expansion and then a further 1GW elsewhere in the region.

SK Telecom frames it as the first major stage of a national platform targeting 15GW by 2035.

The $91.5 billion capital figure requires context. Against 2GW of disclosed capacity, it implies roughly $45.75 million per MW, well above typical hyperscale construction costs.

The gap reflects both understated capacity and substantial spending on GPUs, high-bandwidth memory and other AI infrastructure beyond the facility shell.

The 5GW and 15GW targets are longer-term portfolio goals and are not directly tied to this capital figure.

No anchor customer is disclosed, and no power source is disclosed. For a programme of this size in a grid-constrained country, the second omission is the more significant one.

What Yeongnam reveals about APAC is that the region’s largest commitments are now national in framing and corporate in execution.

A telecoms operator is carrying a state-scale compute target on its own capital plan, and the announcement arrived before the power arrangement did.

Trinidad and Tobago–EY 300MW Development MOU (Latin America)

Trinidad and Tobago signed a development memorandum with EY and third-party developers for a 300MW data center project as part of a broader digital infrastructure push.

The 300MW represents peak power capacity rather than critical IT load.

A reported investment of more than $5 billion spans three national initiatives and cannot be attributed to this project, while its power source remains undisclosed.

The government also signed a separate memorandum with Hummingbird AI Holdings LLC on the same date for an initial 150MW with potential expansion to 500MW.

Two agreements, same week, same counterparty on the state side, different partners and different capacities.

An MOU with an advisory firm and unnamed developers is an early instrument. It commits nobody to spend anything.

What it does is signal that the state intends to package land, gas and policy for this use, and in a country with a long-established gas industry that is not an idle signal.

What Trinidad and Tobago reveals about Latin America is how thin the region’s disclosed pipeline is.

Three transactions across the entire month, and the region’s strongest entry carries no capital figure and no power source.

Any read on Latin America from this dataset is a read on three rows, and the report treats it at that weight rather than inflating it to match the other regions.

Crusoe $10 Billion Israel Lease Expansion to 150MW (Middle East & Africa)

Crusoe was reported to be expanding its leased footprint in Israel to approximately 150MW, with $10 billions of investment over roughly ten to fifteen years.

The expansion combines an additional 40MW on an existing Anan Group lease with a new 67.6MW agreement with MegaDC, across Afula, the Idan HaNegev Industrial Park and Haifa.

The status is Reported rather than announced, and the reason is specific.

Globes identified Crusoe as the customer, but neither MegaDC nor Crusoe had publicly confirmed the full programme, underscoring the distinction between reported and formally disclosed leased capacity.

The capital structure is unusual for a data center entry of this size, because Crusoe is not building.

It is leasing from third-party operators, and a substantial portion of the $10 billion is expected to fund GPU procurement rather than physical facilities.

That makes the figure incomparable with single-site construction capex and it makes the commitment far more reversible than an owned campus would be.

Power source is not disclosed for any of the sites.

What Israel reveals about the Middle East and Africa is that capacity in this region can now be assembled by contract rather than by construction.

Crusoe is acquiring one of its largest international positions without pouring concrete, which compresses the timeline and leaves the power question with the landlords.

Regional Scorecards

North America

Sixty-one transactions made this the deepest pipeline, led by long-term AI leases and power-backed projects.

TeraWulf, CleanSpark and Hut 8 secured 928MW of commitments, while major projects in Texas and Georgia paired data centers with dedicated power strategies.

Many Virginia and Pennsylvania proposals, however, lacked capacity, capital or power details.

The month was defined by hyperscalers and AI labs converting demand into firm commitments.

Europe

Fifty transactions showed a market dividing between projects with secured power and those still waiting for it.

Pure DC, Pantheon AI and Apto advanced large campuses through power and grid agreements, while others pursued alternatives such as offshore gas and wind.

Prologis’ $18.8 billion SEGRO proposal added a major consolidation signal. Europe’s defining constraint remains power access.

APAC

Thirty-three transactions showed the strongest concentration of state-adjacent ambition, led by major AI infrastructure programs in South Korea.

Australia delivered the region’s most complete power arrangement, while Southeast Asia advanced through large campus partnerships and financing.

India remained sovereign-led but modest in disclosed scale, while China surfaced mainly through early-stage initiatives.

Latin America

Just three transactions made July too thin for a confident regional read.

Trinidad and Tobago produced two government-backed memoranda totaling 450MW, while a nuclear-powered initiative emerged in Argentina and Brazil.

None disclosed capital commitments, and established markets such as Brazil, Chile and Mexico were largely absent. The month was defined by intent rather than execution.

Middle East & Africa

Seven transactions showed a region dominated by sovereign-led development.

The UAE, Saudi Arabia and Côte d’Ivoire advanced national AI and data center initiatives, while Jordan added a gas-backed campus proposal.

Commercial activity was limited, led by Crusoe’s reported Israel expansion and projects in Egypt and South Africa. Capital and power disclosures remained scarce.

Forward Indicators

The next 500MW to 1GW increments will land where secured land, power and demand converge. Few markets currently clear all three.

Texas is the strongest signal, with six independent developers assembling power-first positions, including MARA, Liberty Energy and PowerBridge, QTS, Crusoe, Nexus and Hut 8.

In Europe, Finland, Croatia and Poland stand out.

Seinäjoki has the full stack of permits, power, leasing and financing; Topusko has secured technical approval for 800MW; and Poland’s Gryfin Project offers a credible phased path toward 790MW.

Spain is also building a corridor of pre-secured power positions.

Australia leads APAC because large-scale power is already contracted, while South Korea’s major projects still lack disclosed power arrangements.

The UAE and Saudi Arabia are building sovereign delivery platforms ahead of capacity.

Latin America remains speculative, with Trinidad and Tobago’s memoranda still requiring developers and power contracts to convert.

Investor Takeaways

For institutional investors, July reinforced the separation of ownership from operations.

Meta, Nscale, Orange, TeraWulf and others brought outside capital into infrastructure while retaining operating or customer relationships.

These structures create new entry points, but should be priced on contracted demand, not announced capacity.

For sovereign investors, national ambition is not enough. The strongest projects pair policy support with a named utility or power counterparty from the outset.

For operators, contracted demand is now the entry ticket.

CleanSpark, TeraWulf and Hut 8 converted secured demand into long-term leases and financing, while speculative filings without tenants or power increasingly register as noise.

For utilities and policymakers, Richland Parish offers the clearest model: build new generation around future load rather than ration existing capacity.

Where utilities cannot do this, developers are increasingly pursuing off-grid and behind-the-meter alternatives.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• MTN Digital Infrastructure entered a strategic partnership with a UAE-based data center investment platform on August 27, 2026, operating through a new vehicle called Africa Data Hub Holding Limited. MTN disclosed no capital commitment, no equity split, and no financial close date. Ralph Mupita said phase one targets 150MW in South Africa and Nigeria and that MTN will hold a minority position.

• The party funding Africa Data Hub Holding Limited is Tarek Al Ashram’s own investment firm, not the KKR-backed Gulf Data Hub he founded. Bloomberg reported the distinction on August 27, 2026. Gulf Data Hub operated around 40 megawatts when KKR agreed to acquire a stake in January 2025, and Gulf Data Hub is not a party to the MTN transaction.

• Africa Data Hub Holding Limited has no disclosed jurisdiction, equity split, or committed capital, which leaves a co-investor or lender with nothing to underwrite. MTN contributes land, grid access, licenses, and Bayobab fiber, and takes the minority position. The equity split and the holding company domicile are the disclosures that convert the announcement into a transaction.

Part One: The Record

MTN Group announced on August 27, 2026, that MTN Digital Infrastructure has entered a strategic partnership with a UAE-based data center investment platform to accelerate the development of AI-ready digital infrastructure platforms across Africa.

The announcement disclosed no capital commitment, no equity split, and no financial close date.

The partners will operate through Africa Data Hub Holding Limited, a new vehicle that MTN says will serve as the platform through which future digital infrastructure opportunities are developed and scaled across key African markets, with an initial focus on South Africa and Nigeria.

Bayobab, MTN’s digital connectivity business, is a shareholder in the partnership and will provide open access connectivity and go-to-market support through its pan-African network.

The Partner And His Platform

MTN’s release names the founder of the UAE-based investment platform as Tarek Al Ashram and describes him as an experienced data center developer, operator and investor with an established track record across the Middle East.

The release does not name the platform itself. Al Ashram founded Gulf Data Hub, a data center company established in 2012 and headquartered in Dubai.

KKR announced in January 2025 that funds affiliated with it would acquire a stake in Gulf Data Hub through its Global Infrastructure strategy, and the two parties committed to support more than $5 billion of investment in Gulf data center capacity.

Gulf Data Hub operated around 40 megawatts across the United Arab Emirates and Saudi Arabia at that point.

Bloomberg reported on August 27, 2026, that Al Ashram will back Africa Data Hub Holding Limited through his own investment firm.

Scope, Markets And Capacity

Ralph Mupita, MTN Group chief executive, set out the scope of the first phase at the group’s interim results presentation.

Capacity reported him saying the partnership is looking at 150MW as phase one, in South Africa and Nigeria only, and that the partners will build out as demand requires.

Mupita also said MTN will hold a minority position in the venture and described the group’s approach to AI infrastructure as partner led. No percentage has been published.

Mazen Mroué, chief executive of MTN Group Digital Infrastructure, said the partnership brings investment, infrastructure expertise, connectivity and market access into a single platform capable of supporting hyperscalers, cloud providers, enterprises and technology innovators across Africa.

Part Two: The Read

The Credibility In This Deal Sits On Another Balance Sheet

Gulf Data Hub operated roughly 40 megawatts when KKR agreed to buy into it in January 2025. Africa Data Hub Holding Limited is targeting 150 megawatts in phase one.

Coverage of the announcement has repeatedly set the phrase KKR-backed Gulf Data Hub next to the new venture, while the Bloomberg report those same articles cite says something different: Al Ashram is funding Africa Data Hub through his own investment firm.

The track record carrying this announcement belongs to a company that is not a party to the transaction, and the balance sheet that is a party has not been described at all.

Gulf Data Hub Is Not A Party Here

The MTN release names Al Ashram. It does not name Gulf Data Hub, his investment firm, or KKR.

MTN’s own language describes a person with a track record, not an institution with committed capital.

That is an accurate description of what has been agreed, and it is a much narrower thing than the market is reading. The distinction is not pedantic.

Gulf Data Hub carries an institutional shareholder, a stated multi-billion-dollar investment programme, and an operating platform with revenue and lender relationships.

A founder’s personal investment vehicle carries whatever its founder chooses to put into it, disclosed to nobody.

Those two counterparties price differently in a credit committee, and only one of them has been confirmed here.

Watch for one event over the next four quarters.

If Gulf Data Hub or KKR is brought into Africa Data Hub Holding Limited in any contractual role, the venture becomes financeable on institutional terms almost immediately.

If neither appears, the vehicle has to raise third-party capital on a sponsor whose commitment nobody outside the deal has seen, into two markets where sponsor quality is the first question a lender asks.

MTN Contributes What Capital Cannot Buy And Takes The Minority

MTN brings land, grid interconnection, licenses, an existing data center footprint in both target markets, and through Bayobab the fiber that connects them.

The partner brings money and construction experience. MTN takes the minority position.

In South Africa and Nigeria, the binding constraint on data center development is not capital, it is power and the queue to get it.

MTN is contributing the scarce input and accepting the smaller share of the equity.

It gets paid twice for doing so: once on the minority stake, and once through Bayobab, which is a separate shareholder and sells connectivity into the facilities the venture builds.

The result is a structure that keeps 150 megawatts of construction risk off MTN’s balance sheet in the same month the group reported record margin and started buying back its own shares.

That is a coherent capital allocation policy.

It is also a template every other pan-African operator with land and fiber will now examine, and the ones that move first will trade their grid positions into ownership while grid positions still command a premium.

What Each Segment Prices Before The First Site

Private Capital should underwrite the sponsor rather than the track record.

Before indicative pricing, get four things in writing: the equity split, holding company jurisdiction, partner’s committed capital, and whether Gulf Data Hub has any contractual role.

Firms that price this off the Gulf Data Hub name are paying for a credit that has not been offered to them.

Public Markets holders of MTN are buying a minority interest in an unfunded vehicle, attached to a connectivity contract that is real and already earning. That is a reasonable thing to own.

It is not AI infrastructure exposure in the sense the share price will be asked to reflect, and the gap between those two readings widens every quarter the equity split stays unpublished.

Operators in South Africa and Nigeria now face a competitor with 150 megawatts of stated intent, existing sites in both markets, and its own fiber.

The scarce resource is the grid connection, not the capital.

Operators that have not secured their next interconnection agreement within the next two to three quarters will be negotiating for it against a party that already holds a queue position and a telecom licence in the same market.

The Equity Split Is The Next Real Disclosure

The megawatt number is the least informative thing MTN has said about this partnership.

A 150-megawatt phase one target from a sponsor whose only public operating record runs to roughly 40 megawatts is an ambition, and ambitions are free to state.

The disclosures that will actually move this from announcement to transaction are the equity split, the jurisdiction, and the first named site with a signed power agreement behind it.

Expect at least one of those inside two quarters, because a vehicle that intends to raise outside money cannot stay this quiet and still close.

The question to hold until then is simple.

When the terms arrive, whose balance sheet is standing behind the first 150 megawatts, and does that balance sheet have a name the market already trusts, or one it is being asked to trust for the first time.

Join Global Data Center Hub

elcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• Anthropic has reportedly agreed to pay Nscale roughly $45 billion over six years for 460 megawatts of computing capacity at the Monarch Compute Campus in Mason County, West Virginia. Bloomberg first reported the terms on August 26, 2026, citing people familiar with the matter. Neither Anthropic nor Nscale has confirmed the agreement, and no contract terms have been published.

• The reported terms imply $16.3 million per megawatt per year, or roughly 31 percent of allocated build cost returned annually. Reported Phase 1 economics of approximately $71 billion across 1.35 gigawatts put Anthropic’s allocated capital cost near $24 billion, against $45 billions of payments. A contract returning build cost 1.86 times in six years amortizes the NVIDIA Vera Rubin silicon inside the term.

• Nscale retains the residual value of a 460-megawatt accelerator fleet at the end of the term in roughly late 2033, and no public disclosure prices it. Anthropic pays for the hardware across the six years and holds nothing at expiry. The residual and renewal terms are most likely to surface in Nscale’s listing document rather than in any announcement.

Part One: The Record

Anthropic has reportedly agreed to pay Nscale roughly $45 billion over six years for artificial intelligence computing capacity in West Virginia.

Bloomberg first reported the agreement on August 26, 2026, citing people familiar with the matter, and CNBC and Reuters followed with the same terms from the same class of sources.

Neither Anthropic nor Nscale has confirmed the deal.

Nscale has not commented, no contract terms have been published, and every figure below is reported rather than announced.

The reported commitment covers approximately 460 megawatts at the Monarch Compute Campus in Mason County, West Virginia, according to Bloomberg.

Capacity is scheduled to begin coming online in late 2027. The reported term averages roughly $7.5 billion a year.

Capacity, Silicon, And Delivery Timeline

Nscale will deploy NVIDIA Vera Rubin systems at the site, per Bloomberg and CNBC.

The 460 megawatts cover the first of three planned buildings at Monarch, which Blockspace calculates at about 34 percent of the campus first phase.

Nscale has designed Phase 1 to hold 1.35 gigawatts of computing capacity and targets delivery by early 2028.

Reporting carried by TradingKey puts Phase 1 development cost at approximately $71 billion, including around $47 billion for chip purchases.

Nscale separately identifies more than $20 billion in planned capital investment.

Campus Scale And Onsite Generation

The Monarch Compute Campus occupies more than 2,250 acres, with a long-term expansion path exceeding 8 gigawatts.

The campus generates its own electricity rather than drawing from local utilities.

Caterpillar is supplying G3500-series natural gas generator sets, and Nscale targets 2 gigawatts of onsite generation by early 2028.

Nscale has said the development will use closed-loop cooling without municipal drinking water.

Planning documents cited by Bloomberg project 4,375 construction workers at peak, 645 direct operating jobs, and roughly 3,600 indirect jobs across the state.

Site History And Prior Tenants

Nscale acquired the Monarch site in March 2026 through its purchase of American Intelligence & Power Corporation.

Fidelis New Energy and 8090 Industries previously sponsored the development.

Nscale announced a letter of intent with Microsoft for up to 1.35 gigawatts at the campus.

Semafor reported that Google and Microsoft were both in talks for capacity at Monarch before the lease went to Anthropic.

Governor Patrick Morrisey issued a statement calling the reported agreement an extraordinary vote of confidence in West Virginia, and credited State Senator Glenn Jeffries for the state marketing effort.

Anthropic has signed other large compute agreements, including a $50 billion partnership with Fluidstack and capacity arrangements with Amazon Web Services, CoreWeave, and Akamai, according to Data Center Dynamics.

Nscale is preparing a public listing, with reports putting the target raise at around $3 billion.

Part Two: The Read

The Signal Is $16.3 Million Per Megawatt

The number that describes this contract is $16.3 million per megawatt per year.

That is what $45 billion over six years buys across 460 megawatts, and it is a rate of return on build cost that no coverage of the deal has calculated.

Every figure needed to run it has been public since August 26.

The consequence of leaving it unrun is that the market is reading a real estate lease when the payment profile describes something else entirely.

The Payment Profile Amortizes Chips Inside The Term

Take the reported Phase 1 economics at face value.

Approximately $71 billions of development cost across 1.35 gigawatts works out to roughly $52.6 million per megawatt, of which the reported $47 billion chip line accounts for about $34.8 million.

Anthropic’s 460 megawatts therefore carries an allocated build cost near $24 billion. Against that, six years of payments total $45 billion.

The ratio is 1.86 times. Annual rent runs at roughly 31 percent of allocated capital cost.

A stabilized wholesale data center lease does not return its build cost inside the term, because the building outlives the tenant and the landlord underwrites the reversion.

This contract returns the build cost nearly twice over in six years.

The silicon is being paid off inside the lease, and the shell and the power plant are riding along.

That distinction determines who carries technology risk.

Over the next four to six quarters, expect lenders to demand the split be visible in the documents, with the silicon component and the power-and-shell component priced separately.

The contracts that do not break it out will trade wider.

Nscale Owns The Residual And Nobody Has Priced It

Vera Rubin capacity comes online in late 2027 on the reported timeline.

A six-year term carries the contract to roughly late 2033.

Anthropic pays for the accelerators across that window and holds nothing at the end of it.

Nscale keeps a 460-megawatt fleet of six-year-old silicon, a building, and a gas plant.

What that fleet is worth in 2033 is the single largest open variable in the transaction, and it sits entirely on the Nscale side of the table.

If the residual holds value, Nscale has financed a campus on one contract and kept the asset. If it does not, Nscale has run a six-year amortization at a decent yield and Anthropic has bought compute at a fair clip. Both outcomes are live.

Neither is priced, because no one outside the two counterparties has read a renewal clause or a residual term.

The disclosure event is not a press release. It is Nscale’s listing document.

Watch whether the residual and renewal terms appear there, because the first neocloud contract whose back end becomes public sets the template every lender after it inherits.

What a 2033 buyer can do with that hardware was decided in civil engineering, back when the substations were sized and the floor loads were set. Those choices narrow the exit buyer pool years before anyone opens the contract to renew it.

Nscale’s Project Debt Underwrites Anthropic, Not Monarch

Forbes reported that the Anthropic contract functions as the collateral that lets Nscale finance a facility that does not yet exist.

Follow what that means for the lender.

The security is a six-year receivable from a private, unrated company with no published financials and major compute obligations, including the reported $50 billion Fluidstack partnership and deals with AWS, CoreWeave, and Akamai.

The tenant history sharpens the point.

Microsoft held a letter of intent for 1.35 gigawatts at Monarch. Semafor reported Google was in talks.

The anchor tenancy landed with the counterparty that has the largest forward obligations and the least public disclosure of the three.

Over the next twelve to twenty-four months, the pricing input that moves these facilities is not West Virginia gas or NVIDIA delivery schedules.

It is the first comprehensive disclosure of Anthropic’s aggregate compute commitments.

That disclosure reprices every project financing built on an Anthropic covenant on the same day.

What Each Segment Prices Before The First Filing

Private Capital underwriting Nscale project debt or pre-listing equity should treat the $71 billion Phase 1 figure as unverified reporting and rebuild the per-megawatt stack from the chip line up.

Demand the residual definition, the renewal mechanism, and the termination waterfall before pricing.

Firms that skip this step price a real estate risk and hold a technology residual, and they find out which one they bought in 2033.

Public Markets investors have no security to trade yet, which is the point.

The event to prepare for is Nscale’s listing document.

The question to bring to it is what share of contracted backlog this single contract represents.

A backlog concentrated in one unrated private counterparty is a different security from a diversified one at the same headline number, and it will not be presented that way.

Operators competing for the same tenants should read the microgrid, not the acreage.

Monarch’s differentiator is 2 gigawatts of behind-the-meter gas generation targeted for early 2028, which removes the interconnection queue from the delivery schedule.

Sites that depend on a utility connection are selling a later date to tenants who are paying for an earlier one.

Operators who have not secured generation by the end of 2027 compete for the tenants these campuses decline.

The Contract Nobody Has Read Sets The Benchmark

The reported $45 billion is not the durable output of this transaction.

The durable output is a per-megawatt clearing rate for chip-inclusive compute capacity, established by two counterparties who have said nothing on the record, and inherited by every neocloud contract negotiated after it.

Anthropic and Nscale have set a price without publishing one.

The question worth holding for the next two quarters is narrow.

When Nscale’s listing document arrives, does it disclose the residual and renewal terms, or does it report $45 billion as backlog and leave the back end unread?

The first version gives the market a benchmark. The second gives it a number, and a number without a residual clause behind it is the thing that gets mispriced.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• NVIDIA made a minority equity investment in Cloverleaf Infrastructure, a Houston data center site developer, under a strategic partnership announced on August 21, 2026. Neither company disclosed the investment size or the equity percentage. The announcement attaches no offtake, no chip supply agreement, and no power purchase agreement. J.P. Morgan Securities served as exclusive financial advisor to Cloverleaf Infrastructure and Kirkland & Ellis served as legal counsel.

• NVIDIA declined to acquire Cloverleaf Infrastructure outright and took a minority position instead. The Wall Street Journal reported that Cloverleaf Infrastructure was working with JPMorgan Chase on strategic alternatives that included a full sale, and that the company raised $300 million at founding in 2024. Control of a site developer carries the entitlement, permitting, and interconnection liabilities that a minority stake avoids.

• The NVIDIA DSX Platform enters Cloverleaf Infrastructure sites at the design phase, before utility agreements are signed. Substation sizing, floor loading, and coolant capacity are fixed during civil engineering, which moves accelerator switching cost out of procurement. Lenders underwriting powered land will need to price architecture neutrality as a separate variable.

Part One: The Record

NVIDIA made a minority equity investment in Cloverleaf Infrastructure, a privately held data center site developer, as part of a strategic partnership announced on August 21, 2026.

Neither company disclosed the size of the investment or the equity percentage acquired.

The stated purpose is to support the buildout of AI factories and to help Cloverleaf Infrastructure serve customers and utility partners meeting rising demand for accelerated computing.

Cloverleaf Infrastructure is headquartered in Houston.

Disclosed Terms And Transaction Advisors

The announcement records no price for the stake and no equity percentage.

It names no chip supply agreement and no power purchase agreement attached to the investment. It sets out no minimum customer commitment from either party.

J.P. Morgan Securities served as exclusive financial advisor to Cloverleaf Infrastructure, and Kirkland & Ellis served as legal counsel.

The Wall Street Journal reported before the announcement that NVIDIA was expected to invest several hundred million dollars.

The same report stated that Cloverleaf Infrastructure had been working with JPMorgan Chase in recent months on strategic alternatives that included a possible sale of the entire business.

Neither the figure nor the sale process appears in the announcement, and Reuters confirmed only that financial terms were undisclosed.

Founding Backers And Site Portfolio

Cloverleaf Infrastructure was formed in 2024 with an initial investment from Sandbrook Capital and NGP Energy Capital.

The company describes itself as a real estate developer working with investors, energy innovators, and utilities to deliver clean-powered, shovel-ready sites to technology companies that build and operate data centers.

Since founding it has delivered multiple gigawatt-scale projects to customers across North America.

The Wall Street Journal reported that the founding raise totaled $300 million, that Cloverleaf Infrastructure has sold projects amounting to more than 7 gigawatts of powered land, and that its pipeline exceeds 10 gigawatts.

The same report identified Wisconsin sites intended for Oracle and OpenAI. None of those figures appears in the announcement.

David Berry, Co-Founder and CEO of Cloverleaf Infrastructure, said the partnership strengthens the company’s ability to identify and develop high-quality sites.

Application Of The NVIDIA DSX Platform

Cloverleaf Infrastructure will apply the NVIDIA DSX Platform to bring site, power, cooling, computing, and facility decisions together earlier in the design phase.

NVIDIA states the platform helps customers weigh infrastructure tradeoffs and build AI factories that deliver more useful output within available power, water, and grid constraints.

Once sites are operational, NVIDIA DSX infrastructure and software are intended to optimize energy use and computing capacity.

Cloverleaf Infrastructure customers will be able to engage NVIDIA across accelerated computing, high-performance networking, infrastructure and platform software, and NVIDIA DSX.

Nico Caprez, Vice President of Global AI Infrastructure Growth at NVIDIA, said that land, power, and shell form the foundation of AI factories.

Neither company named a first joint site, a capacity target, or a delivery date.

Part Two: The Read

Three Hundred Million Is The Number That Prices This Deal

$300 million is the amount Cloverleaf Infrastructure raised at founding in 2024, according to The Wall Street Journal, and it sets the scale of what NVIDIA turned down.

The sale process and the minority structure have both been reported, and no coverage has put them together: NVIDIA had the whole company available and took a slice.

Every read treating this as NVIDIA moving upstream into land has the direction of travel backwards.

A Declined Acquisition Prices What NVIDIA Actually Wanted

NVIDIA has been writing far larger checks into the power layer.

Reporting places roughly $2 billion into Lancium and $1.5 billion into SB Energy, both ahead of this transaction, and both figures are secondary-sourced.

Against those, a company capitalized at $300 million two years ago was affordable at any structure NVIDIA preferred. It chose the smallest one available.

Control of a site developer means control of its liabilities.

Zoning, interconnection, transformer delays, water permits, and opposition become consolidated risks when a buyer takes the whole entity and a semiconductor balance sheet doesn’t make them disappear.

A minority position with a design integration attached delivers the outcome NVIDIA wants without the drag.

Watch the next site-layer investment. If it is also minority with platform integration, NVIDIA has a repeatable template, the acquisition path is closed, and site development stays fragmented for another cycle.

The Design Standard Lands Before The Interconnection Agreement

The NVIDIA DSX Platform enters at the design phase, which places NVIDIA’s density, cooling, and electrical assumptions inside the site before the utility agreement is signed.

Substation sizing, floor loading, coolant distribution capacity, and water strategy are fixed during civil engineering, and civil engineering is the least reversible layer of a data center.

That timing changes where switching costs live.

A site built around one accelerator’s power and thermal profile can require costly redesign, requalification, and utility renegotiation to switch. Procurement decisions cost little by comparison.

The choice is made years before anyone selects a chip.

Over the next twelve to twenty-four months, lenders underwriting powered land will need to price architecture neutrality as a distinct variable.

A site’s residual value depends on how many buyers can use it as entitled, and a portfolio designed to one standard has fewer of them.

Channel Access Without A Purchase Obligation

The announcement gives Cloverleaf Infrastructure customers access to NVIDIA across the full AI factory stack.

It attaches no offtake, no chip supply agreement, and no minimum commitment in either direction.

NVIDIA reaches gigawatt-scale projects years before any procurement process opens, and it owes nothing for the privilege.

Demand risk sits where it sat before, with the developers and hyperscalers who lease the land and buy the hardware.

Cloverleaf Infrastructure gains capital and a design partner, and gains no contracted revenue.

That distinction matters for anyone marking a site developer with a strategic investor on the register.

Nothing disclosed here guarantees Cloverleaf Infrastructure a dollar of NVIDIA-driven revenue.

Underwrite The Specification, Not The Stake

Private Capital. Infrastructure funds and family offices evaluating site developers should ask: when does a hardware vendor’s design standard enter the site, and what does removing it cost. Price re-specification as a line item.

Funds that skip this will buy powered land portfolios whose optionality was spent during entitlement, and they will find out at exit, when the buyer pool is narrower than the model assumed.

Public Markets. NVIDIA holders now have equity exposure to Lancium, SB Energy, and Cloverleaf Infrastructure, alongside a financing platform targeting over $500 billion in third-party capital. Each stake is individually immaterial, and none carries a disclosed size.

The aggregate is a portfolio of illiquid infrastructure exposures accreting inside a semiconductor income statement with no line item to track it. The disclosure that makes this measurable may take several quarters to arrive.

Operators. Developers and hyperscalers leasing Cloverleaf Infrastructure land inherit whatever specification the entitlement carries. Negotiate design neutrality into the site agreement now, while the pipeline is still being entitled and the terms are open.

Operators who take delivery of a finished site and then discover its constraints will pay for flexibility they could have written in for nothing.

The Standard Outlives The Stake

NVIDIA bought a position in how American AI sites get designed, and it paid a price that will look small against what the standard is worth.

The stake can be sold. The substations, the floor loads, and the coolant capacity built to one specification stay in the ground for thirty years.

The open question is whether any competing accelerator vendor can now enter the design phase at all.

Cloverleaf Infrastructure is one developer among several, and the model here is cheap to copy.

If the answer arrives as silence over the next four quarters, the design layer consolidated while the market was still counting gigawatts.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

In this week’s issue:

First, a utility gave up the one thing utilities never give up. Georgia Power won approval to serve OpenAI’s 3.2GW Effingham County campus and agreed to recover no shortfall from other customers if OpenAI leaves early. Nobody outside the room has read the contract, so who carries that cost instead?

Then, Nvidia kept buying the layer beneath its own customers. It took a minority stake in Cloverleaf Infrastructure, a powered-land developer founded in 2024 with a pipeline above 10GW. That is the third land and power position in a month, after Lancium and SB Energy.

Finally, Spain made cheap renewables the barrier to entry. A fast-tracked draft decree would make new data centers above 1MW match 80% of every hour’s consumption with new renewable generation. The penalty is loss of grid access.

Let us get into it.

What Georgia Sold to Get 3.2 Gigawatts

A regulated utility gave up guaranteed cost recovery and nobody can read the price. Georgia Power won staff approval on 26 August to serve OpenAI’s Project Camellia in Effingham County. The contract covers 3,200 megawatts over 25 years. OpenAI puts the campus investment at $20 billion.

The concession is specific. Georgia Power will not recover any revenue shortfall from other customers if OpenAI terminates early. It must tell the commission within fifteen days of any large-load cancellation. OpenAI pays the full cost of the infrastructure built to serve it.

Commission staff cleared the contract with no commissioner vote, under a rule the commission wrote in April 2025. The public version is entirely redacted. So are the six other data center contracts Georgia Power has filed under that rule.

So here is the discipline. Georgia Power’s protection is a promise, and the promise sits inside a document you cannot read. Before you treat a utility’s ratepayer pledge as protection, ask what the filing would have to disclose, then go and check whether it does.

What Georgia Power bought with it, and the filing that would show whether the trade holds, is below.

THIS WEEK BY REGION

The week’s biggest moves — what happened and what it signals.

North America

Termination risk became the term being negotiated. Georgia Power won approval for its OpenAI contract with no recovery from other customers if OpenAI leaves. Nvidia took a minority stake in Cloverleaf Infrastructure, its third land and power position this month.

nVent agreed to buy Maverick Power for $1.75 billion, with up to $550 million tied to 2027–28 performance. Anthropic reportedly committed $45 billion over six years for 460MW at Nscale’s West Virginia campus. Neither company confirmed the terms.

If you invest in public equities, the vendors are now buying their own demand, and that changes what a backlog is worth.

Europe

Europe is writing the entry rules before capital arrives. Spain agreed to fast-track a draft decree requiring data centers above 1MW to match 80% of hourly consumption with new renewables.

The penalty escalates to loss of grid access. Projects already in the permitting queue get six months to comply or forfeit their connection rights without compensation.

If you develop assets, a queue position in Spain is now conditional on generation you have not built.

Asia-Pacific

Capital in the region is moving toward power at source. Zerra DC lodged a development application for the Western Downs Digital Park, proposing four 360MW buildings for 1.44GW of IT capacity near the Braemar substation.

Cushman & Wakefield puts regional data center asset values above $950 billion by 2030, requiring over $280 billion in new capital. Japan accounts for nearly a quarter of that.

For infrastructure funds, the binding constraint is deliverable capacity, and Queensland is the test case for clearing it.

South America

Colombia’s expansion is carrying its own credit. Equinix raised Bogotá investment to $73 million, completing BG2 at 1,100 cabinets across 6MW to 7MW.

If you sit at a project finance institution, Colombia’s credit quality is settled, and the open question is whether anchor demand ever arrives.

PUBLISHED THIS WEEK

• NVIDIA Did Not Guarantee The Rent. It Guaranteed The Resale. — NVIDIA committed up to $105 billions of residual value guarantees on SB Energy’s Ohio AI data center campus on August 17. Inside the recovery waterfall and the undisclosed minimum value.

• The Texas Pause Is Not The Constraint. Docket 59220 Is. — Governor Greg Abbott froze 474 gigawatts of ERCOT data center interconnection requests on August 3. Inside the 525.5 MW curtailment order, the 30-minute shutdown rule, and the December 10 report.

• The Four-Year Wait for a Transformer (And Why It Decides How Fast AI Grows) — Every AI answer runs through a steel box that takes more than two years to order. That wait sets the pace of the AI buildout.

• Procurement Strategy: Equipment, EPC, and the Contracts That Set Your Timeline — Long-lead equipment, vendor concentration, EPC versus design-build, fixed-price and GMP risk allocation, liquidated damages.

Sponsored By: Global Data Center Hub

Your trusted source for global AI infrastructure analysis. Every week, we break down the trends shaping the future of data centers, cloud, and AI, from emerging markets to hyperscale battlegrounds.

📰 Subscribe to get expert insights delivered to your inbox.

What Georgia Power Bought With It

The utility took a curtailment right in exchange. OpenAI committed up to 1,000 megawatts of flexible demand response. That is close to a third of the load, released back to the grid on request. Georgia Power can serve peak demand without building generation for the full 3,200 megawatts.

The economics run on that trade. Georgia Power projects about $950 million a year of incremental large-load revenue. It projects $2.8 billion of customer savings between 2029 and 2031, and roughly $180 a year off residential bills from 2029.

Those are utility projections filed with the commission. They are not commitments. The bill reduction also depends on enough large-load contracts clearing the commission, and one contract does not deliver it.

What the trade cost is not disclosed. The tariff rate OpenAI pays, the demand-response compensation and the termination terms are redacted. Without those three numbers, nobody outside the commission can tell whether the curtailment right covers the recovery Georgia Power surrendered. A missing disclosure is a finding.

The marker to here watch here is Georgia Power’s 2028 rate adjustment. That filing has to show the large-load revenue actually collected against the $950 million projection. If the number lands short and the bill reduction does not arrive, the commission will have to defend a contract it never voted on.

NOTABLE TRANSACTIONS

Key structures and capital moves from this week’s deal tape.

nVent: $1.75 billion definitive agreement to acquire Maverick Power

nVent is acquiring McKinney switchgear maker Maverick Power for $1.75 billion in cash and new debt, with Bank of America providing the bridge. The deal values Maverick at about 11.5× 2026 adjusted EBITDA, plus a $550 million earnout.

If you underwrite industrials, the mechanism is the earnout, because it prices the buyer’s own doubt about how long the switchgear shortage lasts.

Cloverleaf Infrastructure: undisclosed minority investment from Nvidia

Nvidia took a minority stake in Cloverleaf and agreed to use its DSX design across a 10GW pipeline. The specification lock effectively turns the supplier relationship into a demand guarantee without balance-sheet consolidation.

For operators, the mechanism is standard setting bought with equity, and it decides whose racks fit the sites coming to market in 2028.

Anthropic: reported ~$45 billion, six-year capacity commitment at Nscale

Bloomberg reported a six-year, roughly 460MW commitment at Nscale’s West Virginia campus, later followed by CNBC and Reuters, though unconfirmed. The offtake contract effectively provides project credit to the developer.

If you buy neocloud paper, the mechanism is offtake standing in for balance sheet, and the term length decides whether it holds.

If you are new to Global Data Center Hub, visit our Start Here page for more information about the newsletter.

If you found this newsletter useful, share it with a colleague.

Have a great week.

— Obi

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

This article is the 11th article in the series: Ground to Grid: A Free 21-Lesson Guide to Mastering Data Center Development

Join Global Data Center Hub

TL;DR

• Large power transformers above 50 MVA now run 128 to 160 weeks, which makes procurement the binding constraint on data center delivery. Tier-1 manufacturers ABB, Siemens Energy, Hitachi Energy, and GE Vernova quote 48 to 60 months and hold backlogs into 2030. Tier-2 producers Efacec, SGB-SMIT, and CG Power quote 24 to 36 months, and medium-voltage switchgear runs 44 to 84 weeks.

• The manufacturing clock on data center equipment starts at Release to Manufacture, which can sit a full quarter behind the purchase order. A purchase order issued in January with Release to Manufacture held until April loses twelve weeks before fabrication begins. Hyperscale developers reserve factory slots with deposits of 10 to 30 percent of equipment cost at land close, six to twelve months before groundbreaking, using ratings frozen at roughly 60 percent single-line design completion.

• Guaranteed maximum price contracts signed against incomplete design have produced change orders of 8 to 14 percent on data center projects. Contracts where design is frozen at around 60 percent completion produce 3 to 5 percent. Delay liquidated damages in data center agreements run 0.1 to 0.5 percent of contract value per day with caps of 5 to 10 percent, against caps of 10 to 20 percent in the wider capital projects market.

Procurement sits between design and construction in the data center development sequence.

It is the stage where engineering specifications become purchase orders, vendor contracts, and delivery dates.

On most large projects it is also the stage that determines when the facility energizes.

The reason is straightforward. Construction speed has not changed much. Equipment lead times have.

A team can hold every civil and structural milestone and still wait a year for a transformer, and no amount of site productivity recovers that time.

The order date sets the delivery date before the first crew mobilizes.

The second half of this stage is contractual. The equipment schedule creates risk, and the contract decides who carries it.

Delivery models, pricing structures, and damages provisions allocate the consequences of a slip before anyone knows whether one will occur.

The four procurement decisions that set the timeline

Long-lead equipment. Transformers, switchgear, generators, and cooling systems often drive data center schedules.

Large power transformers above 50 MVA can take 128–160 weeks, while Tier-1 manufacturers quote 48–60 months, with backlogs into 2030. Tier-2 producers typically quote 24–36 months.

Medium-voltage switchgear takes 44–84 weeks, diesel generators 50–110 weeks, and behind-the-meter gas turbines 36–84 months.

Vendor strategy. The choice is between sole-source and multi-vendor procurement, and between bespoke and standardized designs across a portfolio.

Sole-sourcing improves commercial terms and removes the substitution option if that manufacturer slips.

Multi-vendor procurement requires performance-based specifications, written by voltage class, fault rating, and footprint, so several approved manufacturers can bid the scope.

The EPC contract. Design-build and EPC together account for roughly 70 percent of hyperscale project starts. EPC concentrates schedule and performance responsibility with one contractor.

Design-build keeps more design oversight with the owner and often leaves long-lead equipment for the owner to buy directly.

EPCM, more common internationally, makes the owner the direct counterparty to every vendor.

Contract structure and risk allocation. Fixed-price, cost-plus, and guaranteed maximum price contracts distribute cost and schedule exposure differently, and the choice usually follows design maturity.

Guaranteed maximum price contracts signed against incomplete design have produced change orders of 8 to 14 percent, against 3 to 5 percent where design is frozen at around 60 percent completion.

Why equipment lead time is the binding constraint

Three mechanics explain why the equipment schedule governs the project schedule.

The first is that the manufacturing clock frequently does not start when the purchase order is issued.

On many packages it starts at approved submittal or at Release to Manufacture.

A purchase order placed in January with Release to Manufacture held until April has lost twelve weeks before fabrication begins. A single blended lead-time figure hides this.

Purchase order date, submittal approval, Release to Manufacture, factory slot confirmation, factory acceptance test, ship date, and site-need date each need separate monitoring.

The second is that ordering can begin before design is complete.

High-voltage transformers and switchgear typically need ratings frozen only at roughly 60 percent single-line design completion.

Reserving a factory slot with a deposit of 10 to 30 percent of equipment cost at land close is now standard on hyperscale campuses, with the order date preceding groundbreaking by six to twelve months.

Without a reserved slot, nominal lead times extend by another eight to sixteen weeks.

The third is supply concentration. A small group of manufacturers dominates the categories that gate delivery.

Two to three global producers control the high-voltage transformer market, constrained by grain-oriented electrical steel supply, and there is no fast substitute.

Vertiv, Eaton, and Schneider Electric sit across power distribution, uninterruptible supply, switchgear, and cooling.

Qualifying a second manufacturer in every critical category, and writing specifications by configuration, preserves the option to move when a primary vendor slips.

Contract structures and who carries the risk

Two provisions do most of the work inside these structures.

Delay liquidated damages set a daily rate for late completion, commonly 0.1 to 0.5 percent of contract value per day in data center agreements, capped at 5 to 10 percent.

The wider capital projects market runs caps of 10 to 20 percent.

Delay damages and performance damages belong in separate provisions with separate caps, triggered by defined milestones such as mechanical completion.

Force majeure and change-in-law provisions allocate the delays nobody controls.

Current drafting practice enumerates the foreseeable categories explicitly, including equipment lead-time delays, interconnection delays, permitting delays, and tariffs.

These provisions need coordinating across the supply agreements, the construction contract, and the financing documents, so a declaration in one does not create a default in another.

Where the owner orders equipment early and novates the purchase order to the contractor, two mismatches appear.

The vendor damages cap is set against equipment value while the prime contract cap is set against project value.

Vendor warranties often begin at shipment and can expire before the equipment is energized.

Both are addressed by drafting the original purchase order for eventual novation.

What one-month costs

A one-month delay on a 60 MW AI data center has been costed at roughly $14.2 million dollars, covering lost lease revenue, idle labour, and penalty exposure.

That figure explains the behaviour on both sides of the table.

Owners press for wide liquidated damages coverage. Contractors press equally hard for caps.

It also explains why deposits placed years ahead of delivery are now treated as ordinary project cost.

Against a monthly delay cost in the tens of millions, a slot reservation deposit is cheap.

Strategic implications

For developers. Order the long-lead equipment first. The order goes out at land close, six to twelve months before groundbreaking, using ratings frozen at roughly 60 percent single-line design completion.

Specify by configuration so several manufacturers can bid the package and get the factory slot confirmed in writing.

For investors. Ask when the long-lead equipment was ordered and what the Release to Manufacture date is. A purchase order date alone proves little, because fabrication often begins a full quarter later.

A late order is a slipped delivery date the pro forma has not caught up to yet.

For operators. The vendors selected here define the maintenance regime, the spares inventory, and the reliability profile for the life of the asset.

Standardization cuts spares holding and simplifies training.

Multi-sourcing adds resilience at procurement and complexity at operation.

Key takeaway

Construction gets the attention, but procurement sets the date.

The developer who orders early controls the timeline. The one who orders late discovers it is already lost.

Equipment scarcity has moved the first meaningful milestone of a data center programme from groundbreaking to the purchase order.

Everything downstream, including the financing schedule, the lease commencement date, and the tenant fit-out plan, now depends on a manufacturing slot that was allocated years earlier.

The projects delivering in 2029 were ordered in 2025 and 2026.

Capacity beyond that is being allocated now.

On your projects, what is the long-lead item that most often controls the schedule and how early do you now commit to it?

Part 12 examines what a data center actually costs and the drivers that move it: What a Data Center Actually Costs: CapEx Breakdown and the Drivers That Move It

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• Large power transformers, the equipment that converts the very high voltage carried by long-distance power lines into a voltage a building can actually use, now average 128 weeks from order to delivery, according to a Wood Mackenzie survey of utility and developer orders. The largest units, which connect power plants to the grid, average 144 weeks. A data center building takes roughly two years to construct.

• Developers now order power equipment years before a project has a customer or a final go-ahead. Chevron ordered seven large gas turbines from GE Vernova, the power equipment company spun out of General Electric in 2024, roughly two years before signing the twenty-year electricity agreement with Microsoft that those turbines will serve.

• The published order books of power equipment makers show whether an announced AI buildout is real. GE Vernova reported $2.4 billion of data center electrical equipment orders in the first quarter of 2026, more than the company booked in all of 2025, and those figures are filed with regulators every quarter while project announcements carry no such requirement.

Somewhere near your house there is a grey drum on a wooden pole, or a green metal cabinet sitting on a concrete pad at the end of the street.

That is a transformer. It takes the very high voltage that travels down long-distance power lines and turns it into the much lower voltage your kitchen can handle.

Every building on earth has one of these standing between it and the grid.

So does every data center, except theirs can be the size of a small house and weigh several hundred tons.

That box is currently the thing the AI boom is waiting for.

Chips are shipping in volume. Money is arriving faster than anyone forecast. Land has been bought across half a dozen states.

The queue that all of it now stands behind runs through a handful of heavy-equipment factories, and it is measured in years.

This is worth a newcomer’s attention for a reason beyond engineering.

AI infrastructure stories often focus on technology companies or massive investments, but neither explains the pace.

The real constraint is less glamorous: factory schedules booked years in advance.

Once you see that layer, the delays and unusual announcements make sense.

Chevron ordered the turbines two years before it had a customer

In June 2026, Chevron announced a deal with Microsoft.

Chevron will build a power plant near Pecos, in West Texas, and sell all of its electricity directly to a Microsoft data center under a twenty-year agreement.

The plant is designed for roughly 2.67 gigawatts, which is about the electricity used by two million American homes. First power is expected in 2028.

Read as a headline, that is a story about an oil company walking into the AI business.

The detail that actually explains it sat in an interview.

Jeff Gustavson, who runs Chevron’s New Energies business, said the company began thinking about the project about three years earlier and committed to it about two years earlier.

What that commitment consisted of was placing the equipment order.

A gas turbine is essentially a truck-sized jet engine that burns natural gas to spin a generator.

Chevron ordered seven large ones from GE Vernova, the power equipment company spun out of General Electric in 2024.

It placed that order before it had a signed customer, before its own board had given the project a final approval, and before the design was settled.

That order was the deal.

By the time Microsoft was deciding where to put a very large data center, land was available in several places, natural gas was available in several places, and money was available almost everywhere.

A delivery slot for seven large turbines was available from almost nowhere. Chevron was holding one.

The project has a public cost attached to it.

The plant burns natural gas, and Microsoft has a standing pledge to remove more carbon from the atmosphere than it emits by 2030, a target this makes harder to reach.

Chevron has said the plant will run separately from the regional grid so that local households are not competing with it for power.

If you read that announcement as an oil company discovering technology, you missed the part that decided it, which is that the winner was whoever had joined the queue earliest.

The building takes two years. The equipment takes longer.

Here is the number that reorganizes everything.

Wood Mackenzie, an energy research firm, surveyed actual utility and developer orders and found that standard large power transformers average 128 weeks between order and delivery.

That is close to two and a half years. The largest units, the ones that connect a power plant to the grid, average 144 weeks.

Analysts at PwC have put the wait for the highest-capacity units as long as four years.

Now hold that against the building itself, which takes roughly two years to put up.

The reason is unglamorous and hard to fix.

A large power transformer needs a specialized electrical steel that only a few mills in the world produce, plus copper, plus custom engineering, plus months of testing before it ships.

Adding factory capacity takes years, because a transformer plant is itself a heavy industrial building.

And three separate waves of demand arrived at those same factories at once.

Data centers need them. Older grid equipment installed decades ago needs replacing.

Factories and vehicle fleets switching from fuel to electricity need them too. None of those three has any reason to slow down for the others.

So when you read that an AI project has been delayed, the cause is usually a delivery date agreed years ago rather than anything happening on the construction site.

You can now reserve a place in a factory line for power you have not designed yet

Because the wait got long enough, the industry invented a way to buy a spot in it.

GE Vernova calls these slot reservation agreements.

A customer pays to hold a place in the manufacturing schedule before knowing exactly what it will need, in the same way a person books a table months ahead and works out later who is coming.

The scale of this is easy to miss. In the first quarter of 2026 alone, GE Vernova signed 21 gigawatts of new gas equipment agreements.

Only 2 gigawatts of that were firm orders. The other 19 gigawatts were reservations.

The company’s reserved slots rose from 43 to 56 gigawatts in a single quarter, and it told investors it expects to reach at least 110 gigawatts of combined orders and reservations by the end of 2026.

For context, 110 gigawatts is roughly a tenth of all the electricity generating capacity in the United States, held on one manufacturer’s books as future work.

This changes how a headline should be read.

When a company announces gigawatts of new AI capacity, a portion of that number may be a place in a line rather than a machine that exists.

That is why announced capacity and delivered capacity have drifted so far apart.

The factory built to fix the shortage opens in 2028

The obvious response to a shortage is to build more factories, and that is happening.

In South Boston, Virginia, Hitachi Energy has made transformers since 1968. In 2026, it began a $457 million expansion to make the site the largest power transformer factory in the U.S.

The project adds around 825 jobs to a workforce of about 850, close to doubling the place.

It is part of a wider investment of more than a billion dollars in American grid manufacturing.

The plant is expected to begin operating in 2028.

Sit with that timeline. A factory announced to ease today’s shortage may take two years to start production, with transformers requiring another two years to build and deliver.

Relief may not arrive until the end of the decade.

Virginia is a fitting place for it. Northern Virginia holds the largest concentration of data centers on earth.

Data centers already account for roughly a quarter of the electricity sold by the state’s main utility.

The useful correction here is about time.

This shortage is shaped like a decade-long industrial rebuild, and if you read it as a news story that a strong quarter settles, almost every forecast you meet will look more optimistic than it should.

$2.4 billion in one quarter, more than all of last year

There is a practical skill buried in all of this, and it costs nothing to use.

Equipment makers are public companies, so they report what customers actually bought, every three months, in filings that carry legal consequences if they are wrong.

Announcements carry no such requirement.

In the first quarter of 2026, GE Vernova’s electrification division, which makes transformers, switchgear, and the substations that connect large buildings to the grid, booked $2.4 billion of orders tied to data centers.

That was more than the division booked from data centers in the whole of 2025.

Total orders across the company reached $18.3 billion for the quarter, up 71 percent on the year before, and its order backlog, meaning work sold but not yet delivered, stood at $163 billion.

Those numbers say something that a press release cannot. Somebody signed for the equipment, and somebody is paying for it.

So, the next time a very large AI project is announced, the fastest way to know whether it is real is to wait a quarter and look at what the equipment makers say they sold.

The slowest part of the fastest thing

The AI boom is described almost everywhere as a story about speed.

Models improve in months. Companies are built in a year. Announcements arrive weekly.

Underneath that sits an industry moving at the speed of heavy manufacturing. A decision made in 2026 shows up as electricity in 2030.

The scarcest thing in the whole chain is a place in a factory schedule, and somebody else may already be holding yours.

Every answer you get from an AI tool travelled through one of those grey boxes on its way to you.

The chain that ends in a chip begins in a steel mill, and the whole of it moves at the pace of the slowest link.

Before reading this, what did you assume was the slowest part of building a data center?

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• Meta launched the $1 billion Future Is For Everyone Fund on August 10, 2026, for the United States communities that host its data centers. Meta states the money will go to teachers, first responders, and energy and water infrastructure. Meta has not published an allocation formula, a per-community figure, a count of communities, or a spending period.

• The Richland Parish teacher bonuses Meta cites as its model came from a 1 percent sales tax on construction purchases created by a 1968 parish ordinance. Meta paid Richland Parish $22.4 million under that levy in May 2026. Construction on Meta’s Hyperion campus is expected to run until 2030, and the levy applies to purchases rather than to the finished facility.

• Meta confirmed the fund totals $1 billion, and Cheyenne Mayor Patrick Collins said a Meta representative told him the figure was $1 billion for each community. Collins put Meta’s footprint at 28 data centers across 23 states. The two accounts of the same announcement differ by a factor of 28.

Part One: The Record

Meta announced a $1 billion fund for the United States communities that host its data centers on August 10, 2026.

The commitment carries the name Future Is For Everyone Fund and was published inside a 6,500-word letter from Mark Zuckerberg titled The Future Is For Everyone: The Path To A Positive AI Future.

Meta states the money will support teachers, first responders, and energy and water infrastructure in the towns and parishes where it owns and operates data centers.

Meta spokesperson Francis Brennan confirmed the $1 billion figure and said the company will begin working with communities to design investments shaped by local needs.

The Fund Terms And Stated Uses

Zuckerberg described the arrangement as a community compact.

In the letter he wrote that sustainable infrastructure development means communities must benefit significantly from each project, and named high-paying local jobs, investment in schools and public services, holding energy prices flat, and environmental protection as the components.

He wrote that Meta helps keep electricity prices low by building its own energy-generating infrastructure in the places where it invests.

Meta released the model weights for Muse Glimmer, a 30-billion-parameter system, on the same day, and stated it will release open-source versions under the name Muse Spark 1.2.

The fund sits alongside the Data Center Community Action Grants program, which Meta has run since 2011 to fund schools and nonprofits in host regions.

In Fort Worth, Meta partnered with Ducks Unlimited and the Texas Parks and Wildlife Department to fund 206 acres of new wetlands in the Richland Creek Wildlife Management Area.

Richland Parish And The 1968 Levy

Zuckerberg named Richland Parish, Louisiana as the model case.

Meta broke ground on the Hyperion campus near Holly Ridge in December 2024. Construction is expected to continue until 2030.

Richland Parish collected $42.9 million in sales and use tax during the first nine months of the current fiscal year, against $21.3 million for all of fiscal year 2024-25.

Meta made a separate payment of $22.4 million to the parish in May, equal to 1 percent of qualifying purchases, with more than half flowing to the school district.

The Richland Parish School Board has collected more than $36.5 million this fiscal year, of which $16.2 million came from a one-cent sales tax created by a 1968 ordinance that funds employee bonuses.

Bonuses this year reached $50,935. The prior year maximum was $10,200.

Meta pays no state or local sales tax on data center equipment at the site for 20 years.

Once the facility receives its certificate of occupancy, a payment in lieu of taxes arrangement gives Meta an 80 percent property tax abatement for 30 years, conditioned on capital and job targets.

Meta must employ at least 500 people at the site continuously.

Undisclosed Terms And Site Count

Meta has not published an allocation formula for the fund, a figure for any individual community, a count of the communities that will receive money, or the period over which the money will be spent. Brennan said further details will follow.

Cheyenne Mayor Patrick Collins said a Meta representative told him on August 10 that the company plans to invest $1 billion in each community where it operates a data center, and that Meta operates 28 data centers across 23 states.

Meta has said the details will come in the coming months.

Part Two: The Read

The Proof Is A Construction Receipt

The number carrying Meta’s argument is $22.4 million, and it is a levy on construction purchases rather than a payment from the data center.

Coverage has carried the Richland Parish bonuses as evidence that a data center funds teachers, which inverts the mechanism, because the build funds the teachers and the build has an end date.

Every community negotiating against the Richland Parish precedent is therefore pricing a revenue stream scheduled to stop in 2030.

The Revenue Peaks Before The Load Does

Track the two tax streams separately.

The construction stream runs at 1 percent on materials, furniture and fixtures, and it scales with spending velocity.

It reached $22.4 million in a single payment against a parish that collected $21.3 million across the whole of the prior fiscal year.

The operating stream begins at the certificate of occupancy, and it arrives already reduced, at 20 percent of assessed property tax for 30 years.

The two curves move in opposite directions at the same moment.

Construction receipts fall as the last phase completes.

Power draw, water draw and road load reach their sustained maximum on the same date.

A parish that budgeted teacher bonuses against the first curve inherits the second one.

Over the next twelve to twenty-four months this becomes a drafting question rather than a political one.

Host jurisdictions negotiating in 2027 will have watched Richland Parish, and the terms they ask for will be operating-phase floors, escalators tied to megawatts rather than to construction spend, and clawbacks against the 500-job condition.

Developers who arrive with only a construction-phase benefit model will negotiate against a template that has already priced their omission.

An Unallocated Fund Behaves Like An Option

Compare the fund to the program it sits beside.

Data Center Community Action Grants runs through an external platform, publishes its criteria, and names its recipients. Applicants apply and are assessed.

The Future Is For Everyone Fund has none of that architecture.

There is no formula, no per-site figure, no community count, no spending window, and no stated audit mechanism.

That absence changes the relationship. A grant program is an application.

A discretionary pool negotiated directly with municipal government is a conversation that runs in parallel with a permit.

Discretion over the size and the timing sits with the payer, and it sits there during the period when the counterparty is voting.

The binding event in this program is the first published allocation.

Whichever community discloses a number first sets the benchmark every other mayor cites, and Meta has an interest in that number arriving late.

A disclosed ceiling sitting above an undisclosed floor: NVIDIA Did Not Guarantee The Rent. It Guaranteed The Resale.

Two Mayors Cannot Both Be Right

Brennan confirmed the fund at $1 billion in total.

Collins reported being told $1 billion for each community, across a footprint he put at 28 data centers in 23 states.

The first reading produces roughly $36 million per site.

The second produces $28 billion, which would be close to a fifth of Meta’s projected 2026 capital expenditure of $145 billion.

Both cannot hold, and the gap is the mechanism. An unallocated fund lets every host community model itself as the anchor.

That belief is worth more at a zoning hearing than any published figure, because a published figure tells 27 communities what they are not getting.

The gap closes the first time two mayors compare notes in public.

When it does, the fund’s negotiating value drops to the smallest confirmed allocation, and the communities that signed before the disclosure will have signed against the largest.

What Each Segment Now Prices

Private Capital. Underwriting a data center platform with host-community exposure now requires modeling the operating-phase tax base separately from the construction receipt.

The two are different instruments with different durations, and the comparable transactions in the market are quoting the first while disclosing the second.

Firms that price the construction curve into a 30-year asset will hold the political reaction when it inverts, and that reaction lands as permit delay on the next project in the same state.

Public Markets. The $1 billion is 0.7 percent of Meta’s 2026 capital expenditure and does not move the model.

Permit velocity does. Data Center Watch counted 75 projects worth $130 billion blocked or delayed in the first quarter of 2026 alone, matching the whole of 2025, and 71 percent of Americans told Gallup in May they oppose local data centers.

Holders tracking community spend are watching the wrong line. The line is time to energization.

Operators. Consent is now an input to be procured alongside power and land, and it has no published price.

Operators that bring an operating-phase benefit structure to the first council vote negotiate against Meta’s silence, which is the most favorable condition available.

Those that wait will negotiate against whatever formula Meta eventually publishes and will be asked to match it.

The First Published Allocation Sets The Price

Meta has established that a hyperscaler will pay for consent and has declined to say what consent costs.

That combination is unstable, and it resolves in one direction, because a fund with 28 possible claimants and no formula generates its own disclosure through the claimants themselves.

The question for the next two quarters is which community publishes first and what number it publishes.

If the first disclosed allocation lands near $36 million, the community compact is a rounding error on a $145 billion capital program and every subsequent negotiation starts from that anchor.

If it lands materially higher, Meta has repriced the cost of building in the United States for every operator behind it.

Watch the council minutes, not the press releases.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• NVIDIA disclosed multiple residual value guaranties with SB Energy on August 17, 2026, with an aggregate payment obligation cumulatively capped at $105 billion. The guaranties relate to leases covering approximately 4.25 gigawatts of IT load at the PORTS-Pike Technology Campus in Pike County, Ohio, where OpenAI is the tenant under a 20-year lease. NVIDIA is separately investing $1.5 billion in SB Energy.

• The $105 billion figure caps a shortfall, and the guaranteed minimum value that determines what NVIDIA owes is not disclosed anywhere in the public documents. NVIDIA pays the gap between a lease’s guaranteed minimum value and whatever SB Energy recovers through a replacement lease or a sale. The cap binds only where recovery approaches zero.

• OpenAI’s agreement to reimburse NVIDIA is triggered by the same insolvency that would make reimbursement uncollectible. The Trigger Events under the guaranties are OpenAI insolvency resulting in a lease default, or OpenAI failing to make lease payments. NVIDIA’s practical recovery is physical control of the leased capacity.

Part One: The Record

NVIDIA disclosed on August 17, 2026, that it had entered into multiple residual value guaranties with SB Energy, with an aggregate payment obligation cumulatively capped at $105 billion.

SB Energy will build, own and operate the data center under a 20-year lease to OpenAI, and NVIDIA will be the exclusive AI compute infrastructure provider at the site.

The guaranties relate to leases covering approximately 4.25 gigawatts of IT load at the PORTS-Pike Technology Campus in Pike County, Ohio.

NVIDIA filed the disclosure on a Form 8-K alongside a joint announcement with SB Energy and OpenAI.

The Guarantee Structure And Trigger Conditions

Each guaranty becomes effective when the applicable lease commences, and payment is conditioned on SB Energy satisfying ready-for-service conditions for the relevant premises, expected beginning in 2028.

The filing defines two Trigger Events: OpenAI insolvency resulting in a default under a lease, and OpenAI failing to make payments under a lease.

On a Trigger Event, NVIDIA pays an amount generally equal to any shortfall between the guaranteed minimum value of a lease and amounts recovered through a replacement lease or a sale.

NVIDIA holds five elections on a Trigger Event.

It may assume the lease, require SB Energy to relet the premises, initiate a sale, allow the lease to be terminated, or defer those remedies for up to one year while paying specified project agreement costs.

Obligations end upon the earliest of four events: the 20th lease anniversary, OpenAI’s lease termination, achieving a satisfactory credit rating, or other customary termination events.

OpenAI has agreed to reimburse NVIDIA for amounts paid under the guaranties.

Campus Scale, Power, And Grid Investment

The initial deployment is designed to support 4.25 IT-gigawatts of AI factory capacity.

NVIDIA holds an option to extend credit support to the remaining 3.75 IT-gigawatts, exercisable at its sole discretion.

The 8-K states the option covers approximately 3.8 additional gigawatts.

OpenAI is the customer for 8 IT-gigawatts across the campus.

SB Energy and SoftBank Group will build at least 10 GW of generation for 8 IT GW of capacity and invest at least $4.2 billion in regional grid infrastructure with AEP Ohio to protect ratepayers.

The campus spans private and federal land at the decommissioned Portsmouth Gaseous Diffusion Plant, developed with AEP Ohio, the U.S. Department of Energy and the U.S. Department of Commerce.

OpenAI has committed an incremental $40 million to SB Energy’s existing $40 million community benefits fund.

Equity Investment, Exclusivity, And Advisors

NVIDIA will invest $1.5 billion in SB Energy, joining existing investors SoftBank Group and OpenAI.

The investment is separate from the guaranty obligations.

The site will run NVIDIA’s full-stack DSX AI factory platform, including GPUs, CPUs and networking.

Goldman Sachs and J.P. Morgan served as financial advisors to SB Energy. Morgan Stanley served as financial advisor to NVIDIA.

The planned capacity is expected to come online in phases beginning in 2028.

Part Two: The Read

The Guaranteed Minimum Value Is The Number That Matters

Across 4.25 gigawatts of IT load, the figure that determines what NVIDIA owes is the guaranteed minimum value of each lease, and it appears in none of the public documents.

Coverage since August 17 has carried the $105 billion cap as NVIDIA’s exposure, while the filing describes it as a ceiling on a shortfall.

A lender pricing SB Energy’s project debt off the cap is pricing the top of a gap whose bottom is private, and the bottom decides whether the debt is money good.

The Disclosed Cap Sits Above An Undisclosed Floor

Work through the sequence the filing sets out. OpenAI defaults or becomes insolvent.

SB Energy seeks a replacement lease.

Failing that, SB Energy runs a sale. NVIDIA then pays the difference between the guaranteed minimum value and what those efforts recovered.

The $105 billion is reached only where recovery is close to nothing across every lease in the initial commitment at once.

That makes the cap a poor proxy for exposure in either direction.

Set the minimum value low and the guarantee is thin credit support that will not move a project finance committee. Set it high and NVIDIA has written a lease guarantee under a different name.

The value is fixed lease by lease as each lease commences, so it is not one number. It is a schedule.

The schedule leaks out one lease at a time over the next twelve to twenty-four months.

Before the first lease commences in 2028, the only place the aggregate can surface is NVIDIA’s own contingent liability disclosure.

Whether the company quantifies the guaranteed minimum values or republishes the $105 billion cap is the disclosure event that matters more than the announcement did.

OpenAI Reimburses NVIDIA Only While OpenAI Can Pay

OpenAI has agreed to reimburse NVIDIA for any amount NVIDIA pays under the guaranties.

Read that against the trigger. NVIDIA pays when OpenAI becomes insolvent and defaults, or when OpenAI stops paying rent.

In the first case the reimbursement claim ranks with general unsecured creditors of the entity that just failed. In the second case OpenAI is already missing payments on the same leases.

The indemnity is therefore worth the most where NVIDIA never needs it and the least in the scenario the guarantee was written for. What NVIDIA holds instead is physical.

It can assume the leases and run its own workloads. It can direct the relet. It can force a sale. It can wait a year while paying carrying costs. Those five elections are the recovery.

The equity position points the same way. NVIDIA is putting $1.5 billion into the landlord, not the tenant.

In a distress scenario NVIDIA sits inside the entity that owns the asset it would be paying to protect.

A Credit Rating Can End The Guarantee Early

The guaranty terminates on the earliest of four events, and one of them is OpenAI achieving a satisfactory credit rating.

That gives the credit support a variable life inside a fixed 20-year lease.

A project debt structure sized against the guarantee has to survive it going away years before maturity, at a moment set by a rating agency and a definition of satisfactory that sits in documents nobody outside the deal has read.

The direction depends on why the rating arrives. A rating that reflects genuine cash generation at OpenAI replaces NVIDIA’s support with something better.

A rating earned partly because the leases look creditworthy while NVIDIA stands behind them is circular, and it removes the support that produced it.

Watch the covenant package on SB Energy’s first project financing. If lenders accept a structure that steps down when the guaranty terminates, they have underwritten OpenAI.

If they demand replacement credit support at termination, they have priced the risk correctly and NVIDIA’s exit is narrower than the filing implies.

What Each Desk Prices Before The First Lease Commences

Private Capital. Underwriters of SB Energy debt or equity should get three items in writing: the minimum value schedule by lease, the definition of satisfactory credit, and whether OpenAI’s reimbursement obligation is secured.

Firms that price off the cap instead will inherit the template the first executed financing sets, written to clear that deal.

Public Markets. NVIDIA’s contingent exposure sits in a footnote rather than on the balance sheet, and it grows as each lease commences.

Five-year credit default swaps stood at 80.77 basis points on August 18, per a secondary report citing Mott Capital, up from 72.11 basis points on August 11 after Jensen Huang clarified the earlier support framework.

Holders waiting for a quantified liability hold a position whose contingent exposures compound before they are measured.

Operators. The guarantee came attached to exclusivity. NVIDIA is the exclusive AI compute infrastructure provider at the campus and holds equity in the landlord.

Developers competing for frontier lab tenants now compete against landlords carrying vendor credit support, and the price of that support is control of the silicon.

Operators who want those terms should decide what they will give up before a competitor sets the rate.

The Unexercised Option Is The Disclosure To Watch

NVIDIA guaranteed 4.25 IT-gigawatts and took an option on the other 3.75.

That option is the cleanest statement NVIDIA has made about how it values this risk, because it costs nothing to hold and everything to exercise.

A company certain of the residual value of purpose-built AI capacity would have taken the full campus at signing.

The first leases commence in 2028. Between now and then the guaranteed minimum values get set and the project debt gets structured.

The question for the next four quarters is whether NVIDIA exercises before the first building is delivered or after.

Exercising early says the residual value case is proven.

Waiting says NVIDIA wants to see a building relet before doubling the exposure, which is what every lender here should want to see.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

In this week’s issue:

First, Nvidia guaranteed OpenAI’s Ohio leases. The cap is $105 billion across 4.25 IT-GW at SB Energy’s PORTS-Pike campus, and each guaranty switches on only when its lease commences. So what is SB Energy actually holding until 2028?

Then, the equity floor moved. SEC staff confirmed on 29 July that data center securitizations sit outside the Exchange Act definition of an asset-backed security. QTS launched a reported ~$3.9 billion five-year bond the same week at price talk near 7.63%.

Finally, Pennsylvania shut the door. Executive Order 2026-05 makes the GRID standards binding and bans nondisclosure agreements on data center projects.

Let us get into it.

The Guarantee That Pays Nothing Yet

The guarantee behind the largest AI campus ever announced is currently worth nothing. Nvidia entered residual value guaranties with SB Energy on 17 August covering leases for roughly 4.25 IT-GW at the PORTS-Pike campus in Pike County, Ohio. The aggregate cap is $105 billion.

Each guaranty becomes effective when its lease commences. SB Energy has to hit ready-for-service first. The first 800MW is expected in 2028.

Then look at what triggers a payment. OpenAI has to default or become insolvent. SB Energy has to work the asset. Only then does Nvidia cover the shortfall between a guaranteed minimum value and what the relet or the sale recovers, and Nvidia picks the remedy.

So the number everyone is arguing about is a ceiling on a contingent obligation that does not exist yet. The number that decides the trade is the guaranteed minimum value schedule inside those leases. It is not in the filing.

The credit behind the $105 billion, and the two deals that repriced the same risk this week, are below.

THIS WEEK BY REGION

The week’s biggest moves — what happened and what it signals.

North America

The rules for holding data center risk loosened, reducing retention costs. SEC staff confirmed data center securitizations fall outside the asset-backed security definition, removing the 5% retention floor.

Pricing did not follow. QTS launched a reported ~$3.9 billion five-year bond near 7.63%, with Baa3/BBB- ratings expected. Pennsylvania then removed AI data centers from Fast Track permitting and banned project NDAs.

If you allocate to infrastructure funds, the sponsor equity you thought was mandatory just became optional, and the market repriced anyway.

Europe

Capital is going into European sites and stepping back from European compute. SWI Group confirmed on 13 August that over 80% of its capital now sits in a transatlantic platform above 4GW, with the intent to reach 90%.

SWI Group replaced its planned Polarise acquisition with high-double-digit-million-euro debt financing. Polarise was valued at €500 million, with €1 billion planned for buildout. CVC DIF also took majority control of Firstcolo.

Real estate investors should note the split: shells and power attract equity, while the compute inside increasingly attracts debt.

Asia-Pacific

Long tenor is arriving before the tenants are named. A hyperscaler signed a 20-year lease on 50MW at Goodman’s Tsukuba Tech Central, without disclosing who. The regional pipeline reportedly stands at a record 26.5GW.

Equinix signed its fourth Singapore solar agreement, an 11.5MW virtual PPA with Flo Energy, in a market where new capacity depends on securing clean supply first.

Regional data center operators are being asked to commit 20 years of asset life to contracts the market cannot yet price. Watch whether the tenant is ever named.

South America

Chile’s queue is filling with power developers rather than data center companies. Ignis Energy Chile and Lader Energy Chile each filed for grid connection on 500MW projects, targeting the Valparaíso and Santiago regions.

Neither has named an offtaker. Neither has disclosed committed capital.

For tech and venture investors, the scarce asset here is the interconnection slot, and the parties holding it are the ones who own the electrons.

PUBLISHED THIS WEEK

• Nvidia’s $500B AI Infrastructure Financing Platform — Six credit committees disperse borrower risk, but every project in the program is secured on hardware from a single manufacturer.

• Eskom’s Number Is Not 6 Gigawatts. It Is 40 Percent. — The asset Eskom is carving out is the transmission network hyperscalers actually need, and creditor consent gates when it moves.

• What a Data Center Announcement Actually Means (And Why Most of Them Never Get Built) — The number that settles the AI buildout argument is energized capacity, and in Texas it sits under four gigawatts against 445 requested.

• How Cooling, Power, and Network Design Actually Fit Together — Rack density, liquid cooling, electrical redundancy, rail-optimized fabrics, copper reach, stranded capacity, and water and grid exposure.

Sponsored By: Global Data Center Hub

Your trusted source for global AI infrastructure analysis. Every week, we break down the trends shaping the future of data centers, cloud, and AI, from emerging markets to hyperscale battlegrounds.

📰 Subscribe to get expert insights delivered to your inbox.

The Credit Behind the $105 Billion

Nvidia now holds three positions in the same asset. It is the exclusive compute supplier at PORTS-Pike.

It owns equity in the developer through a separate $1.5 billion investment in SB Energy. It guarantees the tenant’s leases.

Each position pays off if the campus fills. Each is impaired by the same event.

The instrument matters more than the cap. A residual value guaranty is neither a loan nor rent backstop.

Nvidia backs lease value after SB Energy attempts to relet or sell; upon a trigger, Nvidia can assume, relet, sell, terminate, or defer for up to a year while covering project costs.

That makes $105 billion a floor on a recovery, not an exposure. The figure that sets the floor is the guaranteed minimum value schedule, and it is not in the 8-K.

The repricing moved in opposite directions this week.

SEC staff removed the 5% risk-retention requirement for data center securitizations, while QTS launched five-year investment-grade debt near 7.63%, closer to high-yield pricing. Sponsors were told to hold less, yet buyers still demanded more.

Europe ran the third version. SWI Group had agreed in February to buy majority control of Polarise at a EUR 500 million valuation with EUR 1.0 billion committed to buildout.

This month it converted that to a loan of up to a high double-digit million-euro amount and kept the exposure senior.

Applying the discipline across all three: Nvidia took the residual risk and disclosed a cap without the schedule that sets it.

QTS took the pricing risk and got paid to. SWI declined the equity risk and moved up the stack.

Only one of those three is a real transfer of risk to a balance sheet that can absorb it, and I cannot tell which from disclosure alone. The missing schedule is the finding.

The document to watch is Nvidia’s 10-Q for the quarter ended 26 July 2026. The full form of the guaranty agreements will be filed as an exhibit.

NOTABLE TRANSACTIONS

Key structures and capital moves from this week’s deal tape.

QTS: reported ~$3.9 billion five-year bond for a Microsoft-linked Georgia campus

Bloomberg reports the Blackstone-owned operator priced five-year notes near 7.63%, upsized by $1 billion, against a Microsoft-linked Georgia campus. At Baa3/BBB-, investment grade no longer drives the sector’s coupon.

If you buy investment-grade credit, the mechanism is a sector premium stacked on the rating, and that premium is now your whole question.

Wingspire Equipment Finance: $140 million against high-density GPU servers

Wingspire, a Blue Owl portfolio company, closed $140 million secured by servers for a private equity-backed GPU cloud operator. The deal underwrites GPU resale value over just a few years.

Lend against compute and the mechanism is hardware residual value, which nobody has priced past one GPU generation.

SWI Group and Polarise: agreed majority stake converted into debt

SWI Group replaced its planned majority purchase of AI cloud firm Polarise with a high-double-digit-million-euro loan. A €500 million valuation plus €1 billion buildout effectively shifted equity into debt.

For a neocloud founder, the mechanism is a backer who takes the site and leaves you the service business.

If you are new to Global Data Center Hub, visit our Start Here page for more information about the newsletter.

If you found this newsletter useful, share it with a colleague.

Have a great week.

— Obinna

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

This article is the 10th article in the series: Ground to Grid: A Free 21-Lesson Guide to Mastering Data Center Development

Join Global Data Center Hub

Power, cooling, and network are not three systems designed in sequence.

They are one system designed at once, because a decision in any one of them reshapes the other two.

This is the engineering integration step. It sits after the conceptual and detailed design work and after capacity planning, and it is where those earlier decisions either hold together or come apart.

A facility can pass every individual discipline review and still fail as a building, because the three systems were reviewed separately and never reconciled against each other.

For most of the industry’s history that separation worked. Racks drew 3 to 15 kW. Air moved the heat.

Cable runs were short enough that the network team could take the room as given.

Each discipline handed a finished specification to the next, and the interfaces between them were standard.

AI workloads ended that arrangement. The three systems now move together, at a scale that punishes poor integration.

The three systems and what each one governs

Power sets the ceiling. It determines how much compute the facility can host at all.

Every added megawatt of IT load cascades directly into cooling load and into structural demand, because the equipment that removes the heat has weight, footprint, and its own power draw.

Cooling determines how dense that compute can be. Density selects the cooling method, not the other way around. Up to roughly 20 to 40 kW per rack, air cooling with containment works at a PUE of about 1.3 to 1.5.

Between roughly 35 and 100 kW, direct-to-chip liquid or rear-door heat exchangers become necessary.

Above 100 kW, immersion or two-phase liquid is the practical option, and PUE falls toward 1.02 to 1.2.

The reason is physical. Water has a volumetric heat capacity of roughly 4,180 kJ per cubic metre-kelvin. Air is about 1.2.

Fan power scales with the cube of airflow volume, so at high density the air-moving equipment consumes an unacceptable share of the power it was installed to support.

Network determines which workloads the facility can serve. AI clusters run two physically separate networks. A front-end network carries storage and management traffic.

A back-end network carries GPU-to-GPU traffic on InfiniBand or RoCEv2 in a non-blocking spine-leaf topology.

Route diversity and interconnect capacity decide whether a building can host distributed training or only inference and enterprise tenancy.

How a decision in one reshapes the others

Raising rack density raises power draw and cooling load at the same moment.

These are not sequential effects. They are the same decision expressed in two disciplines.

Choosing liquid cooling then changes the electrical design. In the Siemens and nVent 100 MW GB200 reference design, each pod uses four 415V, 3,000A switchboards feeding overhead busways.

Coolant distribution units receive four independent 30A feeds; standard network and support racks receive two.

That asymmetry exists for a reason. A rack can briefly survive primary power loss if coolant keeps flowing.

It cannot remain powered without coolant flow. Cooling therefore carries its own redundancy layer, separate from IT load redundancy.

Electrical topology then sets the shape of that layer. N+1 adds one spare unit, 2N provides two fully independent systems, and 2(N+1) combines two independent N+1 systems.

Cooling loops mirror the chosen topology so cooling does not become the weak link in a fault-tolerant design.

The network then feeds back into power. Back-end AI fabrics are rail-optimized, with the same NIC position on each node connecting to the same leaf switch.

That physical cabling requirement dictates where GPU racks can sit.

Distance then becomes a power cost. Direct attach copper draws zero watts but reaches only 1.5 to 3 metres.

A 1.6T optical module draws 25 to 30 watts, while a fully populated 32-port switch dissipates nearly 1,000 watts before compute.

At cluster scale, that adds megawatts to IT load, UPS sizing, cooling capacity, and interconnection demand.

So the floor plan is now drawn backward from the fabric. Leaf switches and rail-connected racks stay within copper reach to avoid optical power and cooling costs.

Cable pathways, chilled water piping, and busways then compete for the same space, requiring coordinated planning under ANSI/TIA-942.

What AI changed

GPU racks draw many times the power of traditional racks.

H100-class accelerators drew roughly 700 watts each. B200 and GB300-class parts draw more than 1,000 watts each.

That density forces liquid, and liquid reshapes the building.

Then the network intensity of distributed training raises interconnect capacity and interconnect power together.

The three systems arrive at the design table simultaneously rather than in order.

A worked example

NVIDIA’s GB200 NVL72 holds 18 compute trays, 36 Grace CPUs, 72 Blackwell GPUs, and 9 NVLink switch trays in a single cabinet.

It draws roughly 120 kW, with some configurations reaching 132 kW.

None of it works without all three systems designed together.

Standard AC power cords cannot deliver 120 kW to a cabinet, so central DC busbars and power shelves replace individual server power supplies, freeing space for GPUs.

Air cooling also falls short at this density, requiring cold plates and rear-frame manifolds to cool GPUs, CPUs, and NVLink switches directly with coolant at roughly 45°C.

The network decision closes it. Inside the rack, passive copper twinaxial cables form a blind-mate backplane delivering 57.6 terabits per second of full-duplex bandwidth per switch tray.

Copper eliminates thousands of transceivers and saves kilowatts per rack. Copper also attenuates quickly, which is why all 72 GPUs and 9 switches must sit inside one frame.

That packaging is only possible because the DC busbar freed the space and the liquid loop removed the heat.

Vertiv and NVIDIA describe the objective of their joint reference architecture as reducing stranded power by aligning AI clusters to data center capacity blocks.

In the Siemens design, the pod boundary is set jointly by electrical block size, a 1.63 MW liquid heat load per pod, and the NVLink domain size. No single discipline sets it.

What this means for each party

For developers. Co-design the three systems from the start.

Sequencing them invites the rework that moves schedules, because a cooling change late in design reopens electrical, structural, and layout work that was already approved.

For investors. A design that treats power, cooling, and network as separate work streams is a design that will change during construction.

Two assets with identical grid capacity can differ substantially in the compute density they can host, and that difference is set during design, not at interconnection.

For operators. The integration quality inherited at handover sets the efficiency, the density ceiling, and the failure modes for the life of the asset.

A hall commissioned correctly for 10 kW racks cannot take GPU tenancy without a cooling retrofit, and that retrofit brings a redundancy layer never included in the original electrical budget.

For policymakers. These choices drive grid load and water use together.

Open cooling towers consume roughly 2.0 to 4.0 litres per kWh of IT energy. Closed-loop direct liquid cooling consumes roughly 0.0 to 0.1.

Grid interconnection studies commonly take 12 to 24 months or longer, and in constrained markets the total delay runs to several years.

Key takeaway

You do not design a data center’s power, then its cooling, then its network. You design one integrated system, because the three are bound together, and the facilities that ignore that binding pay for it in construction.

The cost rarely appears as a single failure.

It appears as stranded capacity, as rework when a cooling change reopens approved electrical work, and as a density ceiling the operator inherits and cannot move.

None of that surface in a discipline-by-discipline review. Each review passes. The building still falls short.

Which of the three systems do you think design teams most often treat in isolation, and where does that isolation show up later?

Part 11 examines the contracts that set your timeline: Procurement Strategy — Equipment Vendors, EPC, and the Contracts That Set Your Timeline

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• A data center announcement is a request for electricity, not a commitment to build. In most American power markets, a developer can join the waiting line to connect a large facility to the grid at little or no cost, without owning the land, financing the project, or signing a customer. The announced figure and the built figure are two different numbers.

• The gap between the two is enormous and measurable. ERCOT, the organization that runs the Texas power grid, reported roughly 445 gigawatts of large-load connection applications through 2033 in its April 2026 update. About 9 gigawatts of that has clearance to switch on, and the most electricity these facilities have ever actually drawn at one time is under 4 gigawatts.

• The sorting is happening now, and it is being paid for by ordinary electricity customers. PJM’s independent market monitor calculated that planning for forecast data center demand rather than operating data center demand added about $6.2 billion to a single capacity auction. Twenty-three states have now approved rules making large power users pay for the capacity they reserve.

A company announces a new data center campus.

The number attached to it is enormous.

Two gigawatts. Ten billion dollars. A county most readers have never heard of.

A week later, another one. Most of those announcements will never become a building.

That claim comes straight from the public filings.

The organizations that run America’s power grids publish, every month, exactly how much electricity has been requested and exactly how much is flowing.

The two numbers sit about a hundred times apart, and almost nobody outside the sector reads the second one.

Learning to read it separates the buildings that are going up from the ones that exist only on a spreadsheet.

It also explains a fight underway right now in utility hearings across most American states, over who pays for the buildings that never arrive.

Texas has requests for 445 gigawatts and a grid that has never delivered 86

A gigawatt is a unit of electrical power.

One gigawatt is roughly the output of a large nuclear reactor, and it is about enough to supply 800,000 American homes.

In its April 2026 monthly report, ERCOT listed large-load connection applications totalling 445.8 gigawatts through 2033.

Large load is the grid operator’s term for any single facility drawing enough electricity to matter to the system, and in Texas close to 90 percent of these applications are data centers.

Set that against what the Texas grid has ever actually done. Its all-time record, on the hottest afternoon in its history, is about 85.5 gigawatts.

The requests are more than five times everything the state has ever consumed at a single moment, including every home, hospital, refinery, and air conditioner at once.

At the end of 2024 the same queue held 63 gigawatts. It has grown roughly sevenfold in eighteen months.

Nothing about that growth curve describes construction. It describes paperwork.

If you read the 445 gigawatts as a forecast of what Texas is about to build, you have misread a waiting list as a schedule.

Most of the requests have never been studied

Here is the detail that does not travel into the headlines.

Of those applications, ERCOT reported that 321 gigawatts had no studies submitted.

An interconnection study is the engineering work that establishes whether the local grid can physically carry the load and what equipment would have to be added.

It costs money and takes months, and it is the first point at which a project has to prove it is serious.

Roughly seven out of every ten gigawatts in the Texas queue had not reached that first checkpoint.

This is possible because joining the line has historically been close to free. A developer can file requests at five sites in three states while intending to build at one.

Every one of those filings enters a public database, and the database gets reported as demand.

Texas regulators approved a new review process in June 2026 called Batch Zero. Its stated purpose is to sort the serious projects from the speculative and duplicate ones.

ERCOT expected roughly 100 gigawatts of the 450 to survive that sort.

When a headline reports a national or state pipeline figure, it is almost always counting requests at this stage. You are reading a number that no engineer has yet checked.

Nine gigawatts cleared, under four actually drawing power

The other end of the pipeline is published too, and it is much smaller.

As of ERCOT’s March 2026 update, 9,042 megawatts (or 9.0 gigawatts) of large load had received Approval to Energize.

That is the clearance a facility needs before it can legally start pulling electricity from the grid.

Among the projects that hold it, the most ERCOT has ever seen them draw at once is 3,883 megawatts (or 3.9 gigawatts).

Roughly 9 gigawatts cleared. Under 4 gigawatts running. Against 445 gigawatts requested.

That final ratio is about one percent, and the intuitive explanation for it is wrong. The obvious reading is that 441 gigawatts of real projects are stuck in a slow queue.

The filings tell a different story. Most gigawatts were never tied to projects with land, financing, or customers, making the queue more of a list of options than a backlog of buildings.

Read it as a delayed pipeline and you will spend the next three years waiting for capacity that was never coming.

Ohio asked for a deposit and two thirds of the demand walked away

The clearest evidence for that reading comes from a single American utility, and it is the closest thing this sector has to a controlled experiment.

American Electric Power’s Ohio utility supplies electricity to central Ohio, including Columbus.

In March 2023 it stopped accepting new data center connection requests entirely.

It had received more than 30 gigawatts of applications. Demand across its whole Ohio customer base has never risen above about 10.5 gigawatts. The requests were roughly three times everything the utility had ever supplied to everyone.

The freeze lasted 28 months. It ended in July 2025, when Ohio’s utility regulator approved a new billing arrangement for data centers drawing more than 25 megawatts. The terms were straightforward.

Sign a twelve-year contract and pay for at least 85 percent of the capacity you reserved whether or not you use it.

The pipeline then fell from about 30 gigawatts to about 13 gigawatts. In a later progress report to the regulator, the utility put its signed data center contracts at roughly 5.7 gigawatts.

AEP’s own explanation was that the arrangement removed the most speculative and uncertain projects, the ones that had never submitted a load study.

Two-thirds of the demand disappeared the moment somebody attached a price to reserving it.

That story has a contested half, and it belongs here.

The Ohio Manufacturers’ Association represents large industrial electricity customers in the state.

In February 2026 it asked regulators to investigate the utility’s forecasting. Its argument runs in three parts.

The 30-gigawatt figure was never a real planning forecast. The utility has not reduced what it reports to the regional grid operator about future demand.

And Ohio businesses and households may be paying today for demand that may never appear.

Both sides are on the public record. Neither has won yet.

What the episode establishes, regardless of who wins, is the mechanism. A queue that is free to join fills with requests that were never going to be built, and the announced number tells you more about the entry price than about demand.

The buildings that do exist are 99 percent full

At this point, the natural conclusion is that the whole thing is inflated.

The finished buildings say otherwise, and holding both facts at once is the sector's actual state.

CBRE, the commercial property firm that tracks the data center market, reported that vacancy across primary North American markets fell to 1.4 percent at the end of 2025.

That figure came in a year when supply in those markets grew 36 percent to 9,432 megawatts.

In Northern Virginia, the largest data center market on earth, vacancy reached 0.3 percent in the first quarter of 2026.

Preleasing tells the same story. Somewhere in the mid-70 percent range of capacity currently under construction has already been committed to a customer, against a historical norm of 40 to 50 percent.

CBRE’s data center lead has described tenants signing leases for buildings that will not be delivered for three or four years.

So the announcements are mostly fiction, and the finished space is scarcer than it has ever been.

Both hold at once, for the same reason. Real capacity is genuinely hard to produce.

That makes it valuable. And that makes it worth announcing far more of it than anyone can deliver.

If you take the phantom queue as proof that AI demand is fake, you have drawn the wrong conclusion from the right number.

Malaysia signed contracts for 5.9 gigawatts and used 405 megawatts

The same gap opens wherever the data center boom lands, well outside the reach of American grid rules.

Malaysia’s southern state of Johor became Southeast Asia’s fastest-growing data center market, on the strength of cheap land and power next door to Singapore.

As of December 2024, 38 Malaysian data center projects had signed electricity supply contracts with the country’s main utility, Tenaga Nasional Berhad, for 5.9 gigawatts of maximum demand.

Their actual draw at the time was 405 megawatts.

The contracted figure came to about 43 percent of the utility’s total supply. The real draw was about 3 percent.

Malaysian officials have since said publicly that they want data centers to use 85 percent of the demand they declare, which is a polite way of naming the same problem Ohio priced.

Take any country’s pipeline figure as a picture of what has been built, and you will be off by roughly a factor of ten.

Why this could show up on your electricity bill

The sorting has a price, and it lands on ordinary electricity customers.

Power grids have to be planned years ahead. Generators, substations, and high-voltage lines take three to five years to build, so grid operators commit to buying capacity long before the demand shows up.

When the forecast is inflated by requests that will never become buildings, that extra capacity still gets bought. The cost goes into the rates everybody pays.

PJM manages the power grid across 13 states, from Illinois to Virginia, and runs an annual auction to secure capacity.

Its independent market monitor ran the comparison directly.

Planning for forecast data center demand, rather than only the demand from data centers already operating, raised the cost of one auction by about $6.2 billion. That is an increase of 61.4 percent.

That is the price of planning for the announcements.

The response has been fast, and it is happening in public. The Smart Electric Power Alliance counts 77 large-load tariffs pending or in place across 36 states.

The Edison Electric Institute counted 23 states with at least one approved as of May 2026.

These rules all do what Ohio’s did. Sign a long contract. Post collateral. Pay for most of the capacity you reserve whether you use it or not.

Virginia’s version covers customers drawing 25 megawatts or more and takes effect in January 2027.

Every one of those rules was decided in a proceeding that anyone could read, comment on, and attend.

The number that settles the argument

There is a loud public argument about whether the AI buildout is a real industrial expansion or a bubble.

Most of it runs on announcement figures, because those are the numbers that make news.

The filings offer a better instrument. Two numbers, both published monthly and free: how much electricity has been requested and how much is flowing.

The first measures enthusiasm and the price of a place in line. The second measures existing buildings, with machines inside them drawing power.

Watch the second one. If it climbs steadily over the next three years, the buildout was real, and the queue was noise around it. If it stays flat while the requests keep growing, the story was mostly paperwork.

The answer will surface first in a grid operator’s monthly report, and in a utility commission docket in your own state, long before it reaches a headline.

For an argument this big, that is an unusually open place for the evidence to sit.

When you see a data center announced near you, what would you want to know before believing it?

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• Chip allocation, not construction, now sets when a data center starts earning. A powered shell can be delivered in eighteen to twenty-four months, while a fully configured accelerator cluster is gated by advanced packaging and high-bandwidth memory capacity no developer controls. The interval between substantial completion and rent commencement is where levered development returns are decided.

• A signed pre-lease is not de-risked revenue, and the market still prices it as though it were. Roughly three-quarters of capacity under construction in primary markets is committed before delivery, but committed capacity is not energized capacity, and energized capacity is not paid capacity. Pre-leasing percentages disclose demand, not the timing of cash flow.

• Disciplined capital underwrites to rent commencement and requires proof of chip allocation before releasing construction draws. Firm commencement terms move silicon delivery risk to the tenant, and buildings designed for a liquid-cooling transition survive a hardware generation change that strands fixed-specification assets. Both are contract terms, not forecasts.

Absorption in data centers is now set by chip allocation, not by construction schedules.

The old underwriting paradigm treated substantial completion as the moment risk ended and revenue began.

Apply that paradigm today and you will fund a finished, powered, cooled building that sits dark for two or three quarters while your tenant waits in a queue it does not control.

Two Clocks Have Come Apart

The gap is widening because two clocks have come apart. A powered shell can be delivered in eighteen to twenty-four months.

A fully configured accelerator cluster, from order to racked and running, often takes longer.

The tenant’s place in that line is set by a supplier the developer never meets.

So the question worth asking is no longer where capacity is being built.

It is where capacity is actually being energized, and who carries the asset until it is.

The Bottleneck Sits Above the Landlord

The constraint sits upstream of anything a developer controls.

Advanced accelerators are gated at three chokepoints: lithography equipment, advanced packaging, and high-bandwidth memory.

Packaging is the tightest of the three.

Memory is close behind, because each gigabyte of high-bandwidth memory consumes several times the wafer capacity of standard DRAM.

New capacity at those nodes tends to get absorbed by the next chip generation.

Supply of the generation already in the field eases very little.

Allocation Decides Before Manufacturing Does

On top of the manufacturing layer sits an allocation layer. The dominant chip designer distributes output through a tiered system.

The largest cloud buyers take the majority of leading-edge supply under multi-year advance commitments.

Sovereign programs and specialist compute platforms split most of what remains.

Everyone else waits and waits at the back.

Consider a developer that delivered a large campus on schedule in a secondary US market. Shell, power, cooling, all complete and commissioned.

The tenant, a specialist compute platform, had signed a five-year take-or-pay contract eighteen months earlier.

Its upstream allocation slipped two quarters. Rent commencement slipped with it.

The developer carried construction debt on a finished asset earning nothing, and the equity return fell by several hundred basis points on a schedule variance nobody had modeled.

A Generation Change Is a Re-Engineering Event

The generation problem compounds this. A previous-generation deployment sits comfortably at forty to fifty kilowatts a rack on air cooling.

The current generation demands more than one hundred kilowatts a rack and mandates direct liquid cooling, with coolant distribution units and manifold networks the older design never anticipated.

When a tenant’s allocation shifts generation mid-build, the developer does not get a schedule delay.

The developer gets a re-engineering event, priced at construction-stage rates.

Committed Is Not Energized

The industry has long assumed that a signed pre-lease is de-risked revenue.

That assumption held when tenant fit out was measured in weeks and the equipment was commodity hardware available on demand.

It does not hold now, because the tenant’s hardware sits in a queue the tenant does not control and the landlord cannot see.

Roughly three-quarters of capacity under construction in primary markets is committed before delivery. Committed is not energized and energized is not paid.

Capital Cannot Buy Queue Position

Time is the real cost here, more than price. Grid interconnection is the longest pole.

Queues at PJM and ERCOT run four to five years, longer than the useful life of the chips that will eventually fill the building.

Substation equipment and transformers carry multi-year lead times of their own.

Where chips arrive and power does not, operators power-cap or rotate fleets, which strands a meaningful share of installed capacity even with the silicon physically on site.

Friction also runs through labor and permits. Skilled electrical and mechanical trades for high-density fit-out are scarce in every hub that matters.

Environmental review and permitting add quarters. None of these compresses with capital.

You cannot pay your way to the front of an interconnection queue, and you cannot pay your way to the front of an allocation queue either. A named-market version of that constraint: Eskom’s Number Is Not 6 Gigawatts. It Is 40 Percent.

Three Lenses, One Interval

For independent operators, the binding constraint is specification under uncertainty.

You commit to floor loading, piping runs, and cooling topology before the tenant’s silicon generation is confirmed.

Specify too low and the asset is obsolete within one hardware cycle. Specify too high and you fund cooling the first tenant will neither use nor pay for.

The recurring variable to underwrite is the transition between hardware generations. Design for the transition and the tenant becomes replaceable.

For private equity and infrastructure investors, the binding constraint is the interval between substantial completion and rent commencement.

That interval is unfunded carry, and it does not appear as a line item in most development models. A two-quarter slip on a levered development is not a rounding error.

It is the difference between a top-quartile outcome and a forgettable one. Model the interval explicitly, size a reserve against it, and settle who pays for it before the shovel goes into the ground.

For public equity, the binding constraint is the credit quality sitting behind the absorption. Platforms with investment-grade offtake and firm commencement terms are being valued on contracted cash flow.

Platforms leased to leveraged compute resellers are being valued closer to a credit spread, whatever the reported pre-leasing percentage says.

Rating agencies have started to draw that line explicitly in how they treat hardware-collateralized facilities.

The dispersion is widening, and a headline occupancy number will not tell you which side of it a platform sits on.

What Disciplined Capital Does Differently

Five disciplines follow from this.

Underwrite to rent commencement, never to substantial completion. Make firm commencement the default term and treat any tenant-controlled trigger as a repricing event, not a concession.

Require evidence of allocation before authorizing construction draws. Confirmed allocation rights from the chip supplier, not a purchase order and not stated intent.

A tenant who cannot produce it is asking you to underwrite its queue position.

Design the building to survive a generation change. Structural floor loading and piping for liquid cooling, even where the first tenant runs air-cooled racks.

The incremental cost at design stage is a fraction of the cost of a mid-build retrofit.

Price the counterparty behind the compute contract, not just the lease. If the tenant resells compute, require credit support and underwrite the megawatts’ re-leasing value independently of the tenant.

Separate the power entity from the real estate entity where grid timing binds. Behind-the-meter generation financed as project debt, the shell financed as real estate debt.

Two structures, two cost-of-capital profiles, one delivery date you actually control.

Run these and the discipline stops being defense. You buy assets other bidders cannot price, on terms they will not accept, to a timeline they cannot meet.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• Anthropic, Macquarie Asset Management and GIC established Theseus Infrastructure on August 10, 2026, a platform that will develop, own and lease purpose-built data centers to Anthropic under long-term agreements. Macquarie-managed funds and GIC own the platform and fund the majority of the equity in each project. The parties disclosed no capital figure, no capacity target, no sites and no timeline.

• The disclosed equity floor in each Theseus project is above 50 percent, and the ceiling is undisclosed. Coverage has treated Macquarie and GIC as the owners without asking who funds the remainder. Whether Anthropic holds project-level equity determines whether construction risk and asset ownership actually sit outside the company.

• Anthropic committed to cover electricity price increases that consumers in host communities would otherwise face from Theseus sites. New York Executive Order 62, signed July 14, 2026, paused discretionary state environmental permits for data centers of 50 megawatts or more on the stated ground that grid upgrade costs should not fall on ordinary ratepayers. Anthropic is already developing New York capacity with Fluidstack.

Part One: The Record

Anthropic, Macquarie Asset Management and GIC announced a strategic partnership on August 10, 2026, to establish Theseus Infrastructure.

Theseus will develop, operate and lease data center infrastructure to Anthropic under long-term agreements.

Each facility will be purpose-built for Anthropic’s capacity needs.

The initial focus is the United States. The parties disclosed no capital figure, no capacity target, no site locations and no delivery timeline.

Ownership And Equity Structure

Funds managed by Macquarie Asset Management, together with GIC, will own the platform.

Those funds will also supply the majority of the equity for each individual project.

Anthropic will act as anchor tenant on each site rather than as developer or landlord.

The three parties will jointly identify and develop new sites.

Macquarie Asset Management manages US$497.6 billion in assets. GIC was established in 1981 and manages Singapore’s foreign reserves.

The parties framed the arrangement around each firm’s prior activity.

Macquarie was cited for developing, financing and operating large-scale digital infrastructure. GIC was cited for infrastructure investing globally.

Data Center Dynamics reports that Macquarie previously invested in AirTrunk and Aligned, exiting both in recent years, while GIC holds stakes in Vantage EMEA, Equinix xScale, EdgeCore, and a Brazilian joint venture.

The Ratepayer Commitment And Undisclosed Terms

Anthropic committed to cover electricity price increases that consumers may otherwise face from Theseus sites.

The company described this as consistent with commitments it announced earlier in 2026.

The parties stated that the developments will require significant capital investment and will create construction jobs and permanent operational roles in host communities.

No figure was attached to the investment; the job counts or the electricity commitment.

Theseus joins an existing set of Anthropic capacity arrangements.

Anthropic said in 2025 that it would invest $50 billion in US data center capacity, developing custom sites with Fluidstack in Texas and New York. Anthropic leases capacity from TeraWulf, Hut 8 and SpaceX.

In July 2026, Anthropic signed a 20-year lease with TeraWulf at its Hawesville, Kentucky campus covering approximately 401 megawatts of critical IT load, with initial capacity expected in the second half of 2027.

The parties said they will work together to identify and develop new sites. They gave no date for a first site announcement.

Part Two: The Read

The Majority Is Disclosed. The Rest Decides The Structure.

Macquarie and GIC will fund the majority of the equity in each Theseus project, which fixes a disclosed floor above 50 percent and leaves the ceiling open.

Coverage has read that sentence as settling the ownership question, and it does the opposite: it names one side of a split and leaves the other side unnamed.

If Anthropic holds project-level equity in the vehicles that own its buildings, Theseus is a co-investment carrying lease documentation, and the assets have not left the company in the way the announcement implies.

No Figure Exists Because No Project Exists Yet

Every other arrangement in Anthropic’s compute stack carries a number.

The Fluidstack program carries $50 billion.

The TeraWulf lease carries roughly $19 billions of contracted revenue across 20 years.

The Riot Platforms agreement carries $9.1 billion with extensions to $16.1 billion.

Theseus carries none, and the reason is structural rather than evasive.

Capital has been committed to a process for identifying sites, not to assets that exist.

That distinction matters because platform announcements and funded facilities behave differently.

NVIDIA announced its own financing platform on August 10, 2026, targeting over $500 billion. Read the inverse disclosure: Nvidia's $500B AI Infrastructure Financing Platform

A platform can be announced, staffed, and quietly deprioritized without a single dollar moving.

The pattern is now common enough to have a shape: large institutional coalitions announce mobilization capacity, and the first executed project agreement is what converts the announcement into a market.

GDCH covered the same structure in NVIDIA’s $500 billion financing platform, where six memoranda produced a headline figure and no binding agreement.

For Theseus, the event to watch over the next twelve months is the first named site with a project-level equity check attached.

That is the moment the platform prices. Until then, the announcement establishes intent and a partnership, and it establishes nothing about scale.

Singapore Is Buying The Tenant And The Building

Data Center Dynamics reports that GIC has previously invested in Anthropic. GIC now also co-owns the entity that will own Anthropic’s buildings and collect its rent.

Sovereign capital is taking a position in the operating company and in the real estate that operating company occupies, inside the same compute stack.

This is a different exposure from sector allocation, and it will not surface in the diligence on either position.

A fund underwriting the Theseus equity examines lease terms, counterparty credit and construction risk. A fund holding the pre-IPO stake examines revenue growth and model competitiveness.

Neither review shows the correlation, because the correlation lives at the portfolio level and both positions clear on their own merits.

The single variable underneath both is Anthropic’s ability to pay rent through a compute cycle. If that holds, the landlord position and the equity position compound together.

If it does not, they impair together, and the lease that looked like a senior claim on a diversified sovereign book turns out to be a claim on the same credit the equity was underwriting.

Expect more of this shape over the next four to eight quarters as sovereign and pension capital moves from data center platforms into tenant-specific vehicles and expect the concentration to be visible only to whoever is looking across both books.

The Ratepayer Guarantee Is A Permitting Instrument

New York Executive Order 62, signed July 14, 2026, paused discretionary state environmental permits for data centers drawing 50 megawatts or more.

The order’s stated rationale is that the cost of electric system upgrades to serve large loads should not fall on ordinary New Yorkers.

Nearly 12 gigawatts of data center load sat in the NYISO interconnection queue as of May.

Anthropic is already developing New York capacity with Fluidstack.

Read the electricity commitment against that order and it stops looking like a community gesture.

It is a pre-negotiated answer to the precise objection that has frozen permitting in one of Anthropic’s named markets, offered before a single Theseus site has been proposed.

Sites that arrive with the ratepayer question already settled move through discretionary review faster than sites that arrive arguing about it.

The commitment is also an unpriced liability, and nobody has said where it sits.

If Theseus carries it, it is an operating expense that reduces the landlord’s yield and falls within Macquarie and GIC’s lease economics. If Anthropic carries it, it is an uncapped, undefined contingent obligation tied to a company still in registration.

Over the next twelve to twenty-four months, expect host-community negotiations across the US to converge on this term as standard, and expect the first structure that discloses where the reserve sits to become the reference point.

What Each Segment Prices Before A Site Is Named

Private Capital. Ask for the equity waterfall before the platform economics.

Two questions decide the structure: what percentage above 50 the sponsors actually fund, and whether the anchor tenant holds any project-level equity.

Firms that wait for the first Theseus project to close will underwrite against terms Macquarie and GIC negotiated on a template written to clear that deal.

Public Markets. Anthropic filed a confidential draft S-1 on June 1, 2026, after raising $65 billion at a $965 billion post-money valuation.

Long-term lease obligations, any residual commitments to Theseus, and the ratepayer guarantee appear in that document, not in the August 10 release.

Holders pricing off run-rate revenue are pricing a company whose compute cost structure sits in agreements they have not read.

Operators. The dedicated-landlord platform competes for the same tenant you do, with site co-development and a permanent lease attached.

Operators bidding Anthropic capacity in the next two quarters are bidding against a partner, not a customer.

The First Named Site Sets The Template

Theseus establishes a coalition, a structure and a role for Anthropic that sits closer to development partner than tenant.

It establishes no capital, no capacity and no schedule, and the distance between those two states is measured in one document nobody outside the three parties has seen.

Two things arrive in the next two to three quarters.

The S-1 becomes public and shows what Anthropic actually owes across leases.

The first Theseus site gets named and shows what the equity split actually is.

Ask which of those lands first, because whichever one it will set the price of every question about the other.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• NVIDIA signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on August 10, 2026, targeting over $500 billion of third-party capital. NVIDIA states the partnerships remain subject to execution of the final agreements. No partner has disclosed a dollar commitment, and no first project has been named.

• NVIDIA may provide residual-value support for up to 25 percent of an opportunity, assessed project by project. NVIDIA has not disclosed what an opportunity measures, whether the support sits in a first-loss position, who values the collateral, or whether an aggregate cap applies. The percentage remains a term-sheet concept until the first agreement is executed and read.

• Goldman Sachs is the only traditional bank among the six partners and holds investment and distribution roles. Goldman is in talks with banks, insurers, asset managers and private-credit firms, and asset managers plan to retain a sizable share of the financing. NVIDIA five-year credit default swaps traded at 72.11 basis points on August 11 after Jensen Huang clarified the limited scope of the support.

Part One: The Record

Six Memoranda And A $500 Billion Target

NVIDIA announced on August 10, 2026, that it had signed memoranda of understanding with six financial institutions to establish AI compute infrastructure financing platforms.

The six are Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. The stated objective is to mobilize more than $500 billions of third-party capital over time.

The capital is intended to create dedicated pools at scale, at attractive rates, for NVIDIA customers including frontier AI labs, enterprises and AI clouds.

NVIDIA’s release states that the partnerships remain subject to execution of the final agreements.

NVIDIA published a follow-up post on August 11 defining the figure.

The more than $500 billion represents aggregate third-party capital that the platforms are designed to mobilize over time.

NVIDIA states the capital is not NVIDIA revenue, a single fund, or a commitment to a single customer.

The Division Of Roles And The Residual Support Mechanism

Under the announced structure, NVIDIA supplies the technology platform and the financial institutions supply and underwrite the capital.

NVIDIA states that the financial institutions will independently assess each opportunity, covering the customer, demand, utilization, cash flow and residual value.

NVIDIA disclosed one form of support. In some cases, NVIDIA may provide a residual-value support mechanism for up to 25 percent of an opportunity, assessed carefully on a project-by-project basis.

NVIDIA describes the arrangement as substantially lower than other compute-financing arrangements, and grounds it in four claimed properties of its compute: fungibility, universal adoption, software upgradability and redeployability.

Jensen Huang framed the commercial premise in one line: in AI, compute is revenue.

Apollo president Jim Zelter called modern compute a scarce, mission-critical asset class.

Brookfield chief executive Bruce Flatt described compute as fast becoming the essential layer of infrastructure.

KKR co-chief executives Joe Bae and Scott Nuttall put their emphasis on execution, saying delivery, not ambition, is the hard part.

Goldman’s Mandate And The State Of Play

Goldman Sachs is the only traditional bank among the six and holds both investment and distribution roles.

Reuters reported on August 14 that Goldman can provide junior capital and private credit through its asset management arm while placing debt with private-credit funds and, in time, public markets.

Goldman is in discussions with banks, insurers, asset managers and private-credit firms, and asset managers plan to retain a sizable share of the financing.

Goldman Sachs Research estimates the top four hyperscalers will spend more than $5 trillion on technology and data centers by 2030.

Credit markets moved twice. NVIDIA five-year credit default swaps narrowed as much as five basis points to 72.11 basis points on August 11 after Huang clarified the limited scope of the residual support.

The yield spread on NVIDIA’s 5.625 percent 2056 bonds narrowed two basis points to 113 basis points over comparable Treasuries.

Four days after the announcement, the Wall Street Journal reported that NVIDIA had cut its planned guarantee of the OpenAI campus in southern Ohio from roughly $250 billion to under $120 billion, limited to the first five gigawatts of a ten-gigawatt program.

No memorandum has converted into a binding final agreement. No partner allocation has been disclosed. No first project has been named.

Part Two: The Read

The Whole Deal Turns On 25 Percent

NVIDIA disclosed one number that determines whether any of this gets financed, and that number is 25 percent.

Coverage since August 11 has converted the percentage into something NVIDIA never described: a quantified first-loss tranche sitting beneath 75 percent of senior debt, multipliable to $125 billion across the program.

NVIDIA disclosed a ceiling on a support mechanism whose base, position in the waterfall and valuation method are all undefined, and a lender who prices that ceiling before reading the agreement is pricing a term that does not yet exist.

Twenty-Five Percent Of An Undefined Base

Start with the words. Up to 25 percent of an opportunity. An opportunity could mean total project cost.

It could mean eligible NVIDIA equipment cost, which would be a much smaller number. It could mean debt principal, expected residual value at a stated date, or a defined loss tranche.

Each reading produces a different credit enhancement, and the spread between the widest and narrowest reading is large enough to change whether a project clears a credit committee.

The instrument is equally open. NVIDIA has not said whether the mechanism is a put, a repurchase undertaking, a deficiency guarantee, a funded reserve account, a lease residual guarantee, or a commitment to take and redeploy equipment.

Also undisclosed: the valuation agent, the appraisal method, which hardware generations qualify, the support tenor, the draw conditions, the fees, and whether an aggregate cap applies across NVIDIA’s whole book.

The consequence over the next twelve to twenty-four months is a documentation race.

The first executed residual-support agreement becomes the market template, because every subsequent lender will benchmark to it and no lender wants to negotiate a bespoke version of a term the market has already standardized.

Whoever closes first writes the definitions everyone else inherits.

Five Asset Managers And One Bank Is A Distribution Design

Count the partners by type. Five are alternative asset managers. One is a bank. That composition tells you the intended holding pattern.

The capital is designed to be originated and held privately first, with public distribution arriving later, once portfolios season and ratings exist.

Goldman’s dual role sits precisely on that seam, supplying junior capital and private credit through asset management while structuring and placing senior paper through the bank.

Two developments make the second stage cheaper.

Larry Fink’s comparison to the early mortgage-backed securities market signals ambition around standardization and distribution.

On July 29, the SEC’s Division of Corporation Finance agreed, based on facts presented by a law firm, that certain securities issued by entities directly owning data-center facilities fall outside the Exchange Act’s Section 3(a)(79) definition of asset-backed securities.

The staff was explicit that the position is not a Commission rule and may change on different facts.

For qualifying structures it removes credit-risk-retention, Regulation AB reporting and third-party diligence reporting requirements, and it does so at the exact moment six institutions want a tradable market.

The binding constraint on this program is not the supply of capital. It is the absence of a rated benchmark for compute-backed credit.

Watch for the first rated tranche, because that is the event that converts a private origination business into a market.

Six Managers Diversify The Borrower And Not The Collateral

The strongest structural argument for the platform is that risk genuinely leaves NVIDIA’s balance sheet at origination.

Six independent credit committees underwriting separate projects across separate customers do disperse idiosyncratic borrower risk, and that is a real improvement over a single vendor lending a single customer the purchase price.

The Ohio revision is evidence of the direction of travel: NVIDIA cut a concentrated corporate guarantee by more than half within four days of announcing the distributed alternative.

Now hold that against what does not disperse. Every project in the program is collateralized by hardware from one manufacturer.

That manufacturer sets the product cadence that determines the depreciation curve.

That manufacturer supplies the software layer that underwrites the claim of extended economic life.

That manufacturer may provide residual support on the same collateral it produces, and it earns revenue when financed proceeds buy more of it.

The recovery model of every lender in the program and the value of NVIDIA’s support both depend on one variable: the future market-clearing price of NVIDIA compute.

Diversifying across six managers reduces the chance that one borrower’s failure impairs the pool.

It does nothing about a common shock to GPU-hour pricing, because a common shock arrives at every project at once.

What Each Segment Prices Before The First Close

Private Capital. Underwrite as though the residual support is worth zero until the executed agreement is in front of the credit committee.

Put the five definitions in the term sheet before the indicative pricing: the base the 25 percent measures, the position in the loss waterfall, the valuation agent and method, the support tenor, and the aggregate cap.

Firms that skip this step and wait for the market template will inherit definitions negotiated by whoever closed first on terms written to clear that deal and to protect nobody after it.

Public Markets. NVIDIA’s economic-at-risk amount is undisclosed and accretes transaction by transaction.

Five-year credit default swaps at 72.11 basis points price the August 11 clarification of scope, and they do not price an aggregate that no one has published.

The disclosure threshold that would make this quantifiable may not be reached for several quarters, and holders who wait for it are holding a credit whose contingent exposures compound in the interim without appearing anywhere.

Operators. Contracted offtake is the currency in this market.

Projects carrying investment-grade offtake, secured power and explicit equipment-redeployment rights will price toward infrastructure spreads.

Merchant GPU fleets and single-customer frontier-lab exposure will keep a technology premium.

Operators that lock power and offtake within the next four quarters compete for the first pools, when documentation is still negotiable, and everyone after them competes for allocations under templates already hardened.

The Decisive Document Is The One Nobody Has Read

The August 10 announcement created origination capacity and a coalition with the balance sheet to use it.

It did not create a funded facility, and the distance between those two things is measured in documents.

The first residual-support agreement will define the base, fix the loss position, name the valuation agent and set the cap, and every deal that follows will be priced against those four terms whether or not its lenders negotiated them.

That leaves one question for the next two quarters.

If the first project’s residual-support agreement measures the 25 percent against eligible equipment cost, the enhancement is a fraction of what the market currently assumes, and the pricing on every deal behind it moves.

Ask which reading the first document takes before the market decides it already knew.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

In this week’s issue:

First, the tenant that owns nothing. Riot Platforms signed a 20-year, 191MW lease at Rockdale, Texas worth a reported ~$9.1 billion, and Bloomberg identified the tenant as Anthropic through people familiar with the matter. Riot is funding early construction on a $573 million interim facility. What is a purpose-built 191MW campus worth to anyone other than the lab it was designed for?

Then, the owners changed but the users did not. Anthropic launched Theseus Infrastructure with Macquarie Asset Management and GIC, who own the platform and fund most of the equity per project. They disclosed no capital figure, no site count, and no timeline.

Finally, the exit window is opening. Vantage is reportedly exploring a listing or sale at ~$100 billion and DayOne filed confidentially for a reported ~$5 billion, which will set the price public markets pay for contracted AI capacity.

Let us get into it.

The Landlord as the Investment

The party using the compute is now the only party without an asset. Anthropic proved it twice in 48 hours. Riot Platforms disclosed a 20-year lease for 191MW at Rockdale, Texas, worth a reported ~$9.1B through June 2048.

Riot named the tenant only as a leading frontier AI lab. Bloomberg identified Anthropic through people familiar with the matter. Riot declined to comment and Anthropic did not respond.

Anthropic launched Theseus Infrastructure with Macquarie and GIC, which own the platform and fund most project equity. Anthropic is the anchor tenant, with no capital, site count, or timeline disclosed.

So look at what Anthropic holds in both. It holds a long-dated payment obligation. It does not hold a building, a grid connection, or a share of the equity.

So here are the two tests: Who owns the asset when the tenant owns nothing? And what is it worth to the next tenant if the first one walks?

What happens to Rockdale if Anthropic walks, and how the same question runs through two more deals this week, is below.

THIS WEEK BY REGION

The week’s biggest moves — what happened and what it signals.

North America

Ownership separated from use across the biggest US deals this week. Riot Platforms signed a 20-year, 191MW Rockdale lease worth a reported ~$9.1 billion, with early construction funded by a $573 million interim facility. Anthropic, Macquarie Asset Management and GIC launched Theseus Infrastructure, where the investors own the platform and the lab is the tenant.

Invitium Energy, the PPL and Blackstone venture, reserved more than 5GW of turbines for Pennsylvania with no customer signed. Vantage Data Centers is reportedly exploring a listing or sale at ~$100 billion.

If you run an infrastructure fund, the equity is now the scarce position in these deals, and the tenants are handing it to you.

Europe

European capacity is being paid for by parties that will never run a server. Nebius agreed to lease high-density capacity at Vantage’s CWL1 campus in Newport, Wales, the first announced commercial commitment in the South Wales AI Growth Zone.

Uniper has identified more than 10 European sites that could be sold or leased for data centers, three of them at an advanced stage across the UK and Germany. Berlin still owns 99.12% of Uniper and is preparing to sell down.

Real estate investors should watch what Uniper is really selling: an existing grid connection.

Asia-Pacific

Asian capital moved toward the exit and toward home assets in the same week. DayOne Data Centers confidentially filed for a US IPO, reportedly seeking around $5 billion, eight weeks after a $4.5 billion Series C.

Digital Core REIT agreed to sell interests in three North American data centers to its own sponsor for about $316 million, and its Asia-Pacific exposure doubles to 22% of assets.

Public markets investors get a clean read here. A listed vehicle is selling American assets into the strongest American demand on record.

South America

Connectivity capital arrived ahead of the compute. Google announced three subsea systems under Americas Connect: Alisios, Canoa and OlaLuz.

Alisios links the Dominican Republic, Panama, and Chile; Canoa links Bermuda and the Dominican Republic. Google disclosed no capacity or timelines.

For finance institutions, the route map is the tell. The Dominican Republic just became a hub, and capital that follows fibre usually arrives within three years.

PUBLISHED THIS WEEK

• Microsoft Opens India South Central In Hyderabad On $20.5 Billion — Data residency is the product, and the four-region footprint is what Microsoft is selling to regulated Indian buyers.

• The EU Gigafactory Tender Selects Balance Sheets. Polarise Is Selling Operating Skill. — The capital sponsors take the equity, and the operating specialists take a contract, because the tender gates on money and secured power.

• The Data Center Glossary: What the AI headlines have been saying all along — Every term that appears in a data center headline or term sheet, defined once for readers without a sector background.

• Two Carriers, One Trench: The Diversity You Paid For and Never Verified — Two carrier contracts no longer produce two physical paths, and a physical trace before close is now a condition of underwriting.

Sponsored By: Global Data Center Hub

Your trusted source for global AI infrastructure analysis. Every week, we break down the trends shaping the future of data centers, cloud, and AI, from emerging markets to hyperscale battlegrounds.

📰 Subscribe to get expert insights delivered to your inbox.

What Happens to Rockdale If Anthropic Walks

Riot is building a single-tenant campus for a tenant it will not name. The lease runs 191MW at Rockdale, Texas, and generates a reported ~$9.1 billion through June 2048.

Two five-year extensions would take it to ~$16.1 billion. Delivery is staged, with 96MW targeted by December 2027 and the full 191MW by June 2028.

The funding behind that is not settled. Riot secured a $573M interim facility through Morgan Stanley for early construction while pursuing permanent financing. The 22-year revenue stream currently rests on bridge funding.

Riot’s second tenant is AMD. Together the two leases give Rockdale 241MW under contract and roughly $9.8 billion of long-term contracted revenue, all signed inside six months.

Theseus discloses limited details: Macquarie and GIC will fund most project equity, with Anthropic as anchor tenant. The US-focused platform has no disclosed fund size, site count, or first project.

Now run the two tests across the week.

At Rockdale, Riot owns the asset, but residual risk is less clear. The 191MW lab-specific campus has fewer potential buyers than a colocation facility, and lease termination and reletting terms remain undisclosed.

At Theseus, Macquarie and GIC own the asset and Anthropic pays rent. That is the cleanest answer to the first test all week.

The second test cannot be run at all, because no site, no cost, and no lease term has been disclosed. A missing disclosure is a finding.

At Invitium, PPL and Blackstone own turbines and interconnection queue positions with no customer attached.

Those assets have a buyer whoever the tenant turns out to be. It is the only structure this week that passes the second test on its face.

So the marker I am watching is Riot’s next filing. The permanent backstop replaces the $573 million bridge, or it does not.

Until it does, the largest contracted revenue stream a bitcoin miner has ever signed sits on interim credit.

NOTABLE TRANSACTIONS

Key structures and capital moves from this week’s deal tape.

Invitium Energy: over 5GW of combined-cycle turbines reserved for Pennsylvania data centers

The PPL-Blackstone venture has secured 8–14GW of land, reserved 5GW+ of turbines, and queued ~5GW in PJM, with $12.5B–$15B in potential investment through 2032. Construction awaits acceptable-risk energy agreements.

If you develop generation, the mechanism is optionality bought cheap, and the turbine slot is now worth more than the site it will stand on.

Digital Core REIT: about $316 million of North American interests sold to its own sponsor

The Singapore-listed REIT is selling Toronto, Los Angeles, and Northern Virginia stakes to Digital Realty while increasing Singapore and Osaka holdings. ~$176M of 3% SGD/yen debt replaces ~4.4% USD/euro debt.

For real estate investors, the mechanism is currency and rate arbitrage dressed as a growth story, and unitholders vote on it before year end.

DayOne Data Centers: a confidential US IPO filing eight weeks after a $4.5 billion private round

DayOne raised $4.5B in a Series C led by Coatue and Hillhouse, then confidentially filed for a reported ~$5B US listing. Private capital funded the build; public capital is now being tapped for the next one.

Buy the listing and you are accepting a private round as the valuation marker, and the IPO tests whether others agree.

If you are new to Global Data Center Hub, visit our Start Here page for more information about the newsletter.

If you found this newsletter useful, share it with a colleague.

Have a great week.

— Obinna

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

This article is the 9th article in the series: Ground to Grid: A Free 21-Lesson Guide to Mastering Data Center Development

Join Global Data Center Hub

TL;DR

• Rack density has separated into two curves, with fleet-wide averages between 8 kW and 27 kW per rack and AI and HPC cabinets between 40 kW and 142 kW. Air cooling reaches a physical limit near 41 kW per rack, above which direct-to-chip liquid becomes the only viable path. A capacity plan calibrated on a single blended density figure is calibrated against a facility that does not exist.

• Redundancy in a data center is specified as N, N+1, or 2N, and the cost of 2N scales with every megawatt rather than sitting as a fixed premium. N provides no spare capacity, N+1 adds one backup component, while 2N duplicates critical components across two independent distribution paths. Redundancy topology and Uptime Institute Tier classification are distinct: Tier evaluates overall facility capability, not component counts alone.

• High-voltage substation transformers carry lead times of 36 to 48 months and medium-voltage switchgear runs 20 to 30 months, against a lease cycle that runs roughly 18 months from signature to required energization. Mechanical and IT infrastructure can be deployed in 5 MW to 10 MW blocks as leases are signed. The electrical backbone has to be committed before that demand is visible, which is the boundary the phasing decision runs into.

Capacity planning is the discipline of building for demand you cannot yet see.

Build too little and you cap your growth. Build too much and you strand capital in empty halls. The whole skill is calibrating the gap.

This is the quantitative core of the design stage.

Everything before it describes what a market might absorb.

Capacity planning converts that description into engineered megawatts, cooling capacity, and floor area.

Three decisions carry the work. They are usually taught as three separate exercises.

They resolve as one, because each decision moves the other two.

Get the calibration wrong and the cost appears in one of two forms: tenants you cannot serve, or capital sitting in space that earns nothing.

The Three Variables

IT load. This is the compute the facility must power and cool, measured in megawatts.

It is set by the tenant and the workload type, and it anchors every other system in the build.

Two related numbers sit above it. Facility load adds cooling plant, UPS conversion losses, distribution losses, lighting, and security.

Utility load is the gross capacity reserved at the substation, which includes facility load plus transformer losses and utility engineering margin.

Power Usage Effectiveness is facility load divided by IT load.

PUE changes with utilization because fixed overhead remains even when compute load falls.

One source estimate day-one PUE at 1.85–2.10 for a 40 MW plant carrying 10 MW initially, versus 1.18–1.25 for a modular plant at the same load; these figures are document-sourced and independently unverified.

Density is where IT load becomes a design constraint, not just a number. Average rack density ranges from 8–27 kW, while AI/HPC racks can reach 40–142 kW.

Air cooling approaches its practical limit around 41 kW per rack, making direct-to-chip liquid cooling increasingly necessary at higher densities.

Redundancy. This is how much backup the design carries.

N carries no spare capacity, so any failure or maintenance event sheds load.

N+1 adds one component beyond base capacity. 2N duplicates every critical component across two independent distribution paths, either of which can carry the full design load.

More redundancy buys more reliability and costs more per megawatt.

The level chosen is a commercial decision as much as a technical one, because it is set by what the tenant will pay for and contractually require.

Scalability. This is the path to add capacity in phases as demand arrives.

The design objective is that expansion does not require rebuilding what already exists.

The standard approach separates the building from its equipment: developers build a powered shell for long-term capacity, then deploy 5–10 MW blocks as leases are signed.

This works only if the shell is designed for future loads, with floor slabs above 350 psf and clear heights over 24 feet to support liquid-cooled racks without structural retrofits.

How the Three Interact

Higher IT density raises cooling and power requirements across the whole build, not only in the hall where the dense racks sit.

The distribution topology, the plant sizing, and the structural loading all move with it.

Higher redundancy multiplies the cost of every megawatt. A 2N design does not add a fixed premium.

It scales with capacity, so the redundancy decision compounds as the facility grows.

A scalable design front-loads some cost to avoid a larger cost later.

Oversized fluid headers installed during shell construction cost under 1 percent of total project capital.

Retrofitting the same pipework into an occupied hall costs $2.0 million to $3.0 million per MW.

The tenant type sets the acceptable trade-off among all three.

An AI training tenant will accept lower redundancy in exchange for maximum power and density, because the software layer handles fault recovery.

A multi-tenant colocation operator cannot make that trade, because the reliability commitment is contractual and uniform across the hall.

Where Capacity Planning Fails

Designing for peak demand that never materializes.

The plant is sized for an end state that absorption never reaches, and the capital sits idle while the equipment depreciates.

Under-provisioning redundancy for a tenant that requires 2N. The facility is complete and cannot serve the lease it was built to win.

Building a facility that cannot phase forces full upfront commitment, and long equipment lead times make late additions difficult.

High-voltage transformers can take 36–48 months, while medium-voltage switchgear may require 20–30 months. If the electrical backbone is ordered only after a tenant signs, energization may miss the planned timeline.

Locking density assumptions that the workload later exceeds. A hall engineered for 8 kW per rack exhausts its power and cooling capacity while a large share of the floor sits empty, and the retrofit to liquid arrives at brownfield cost.

One distinction matters when reading these levels against a certification claim.

Redundancy describes component counts and pathway layouts.

Uptime Institute Tier classification measures operational capability across the whole facility, including distribution path count and concurrent maintainability, so the two are separate measures rather than one scale.

Strategic Implications

For developers. Phase your capacity to demand. The empty hall you built too early is capital earning nothing.

Separate what phases from what does not, because the components with multi-year lead times have to be committed on a different schedule than the mechanical plant.

For investors. Understand what the capacity plan assumes about absorption.

Overbuilt redundancy and stranded phases both erode returns, and they erode them through different mechanisms.

Overbuilt redundancy raises cost per megawatt permanently. Stranded phases delay revenue against capital already spent.

For operators. The redundancy level and density ceiling set at this stage define your reliability commitments and your headroom.

Both are difficult to change once the hall is occupied.

For policymakers. Capacity decisions drive the power load your grid must serve.

The density trend is raising it fast, and the fleet-wide average understates the pace because high-density deployments sit far above it.

Key Takeaway

Capacity planning is not about building the biggest facility.

It is about building exactly enough, phased to arrive as the demand does. The waste is on both sides of that line.

The calibration question is where each subsequent stage of the development process begins.

The financing structure, the construction sequence, and the operating cost base all inherit the numbers set here.

When you plan capacity, do you err toward overbuilding for growth or phasing tightly to demand, and what has that choice cost or saved you?

Part 10 examines how cooling, power, and network design fit together: How Cooling, Power, and Network Design Actually Fit Together

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• The AI buildout is described in three separate professional vocabularies, and almost no article tells the reader which one it is using. Data center coverage borrows from commercial real estate, from electrical engineering, and from project finance. A reader fluent in one of the three is still lost in the other two.

• The terms that cause the most confusion are the ones that sound ordinary. Words including load, tier, capacity, stranded, and platform each carry two or three unrelated meanings inside the same sector, and headlines move between those meanings without warning.

• The vocabulary decides public outcomes, not only private ones. Whether a data center gets built, how long it waits for electricity, who pays for the power lines it needs, and whether a town approves it are all settled in these words, in filings and hearings that are open to the public.

A sentence like this one appears in data center coverage most weeks:

“The developer secured 300 MW behind the meter, signed a fifteen-year triple-net lease with a hyperscaler, and is still sitting in the interconnection queue.”

Nine words in that sentence are doing real work, and a reader who has followed the AI story closely for two years may still be unable to define a single one of them.

The sector never wrote anyone a glossary. This vocabulary was built by people who assumed that everybody in the room already had it.

The reason the vocabulary is so hard is worth stating plainly, because it explains why other explainers have not fixed it.

Data centers sit at the intersection of three industries that were unrelated until recently.

Commercial real estate contributed the language of leases, tenants, and yields.

Electrical engineering contributed megawatts, substations, and redundancy. Project finance contributed the debt structures that pay for all of it.

Each of those fields had a century to build its own jargon. Then they collided in a single building, and nobody harmonized the dictionaries.

What follows is that dictionary. It is organized the way the sector actually divides: the building, the power, and the money.

Each term gets a plain definition, what it decides in practice, and where a reader is likely to meet it.

At the end are five ordinary-looking words that carry more than one meaning, which are the words that cause the most trouble.

The building

Start here, because the physical object is the one thing in this sector a reader can picture without help.

A data center is a warehouse full of computers. The computers are servers, which are desktop computers with the screen and keyboard removed and the working parts flattened into a tray.

The trays slide into a rack, which is a tall metal shelf about the height of a refrigerator, holding dozens of them. A large building holds thousands of racks.

Everything else in this section describes how those racks are housed, cooled, powered, and sold.

Hyperscale and hyperscaler. A hyperscaler is one of the handful of companies that operate computing at planetary scale and build data centers for their own use: Amazon, Microsoft, Google, and Meta are the names a reader already knows.

Hyperscale describes a facility built to their standard, which generally means a campus rather than a single building.

The word appears in almost every article about the buildout, and it is a description of the customer as much as the structure.

Colocation. A business that owns the building and rents space, power, and cooling to companies that own the computers.

The tenant brings its own servers and pays for the room and the electricity to run them.

Retail colocation rents by the rack or the cage to smaller customers.

Wholesale colocation rents in blocks of megawatts to large ones. A reader will see the word shortened to colo.

Rack. The metal frame that holds servers. Capacity is measured by the electricity a rack can draw, which is why data centers are sized in megawatts rather than square feet.

Density, power density, rack density. How much electricity a single rack consumes.

This is the number that changed when AI arrived. A rack of ordinary business servers draws a modest amount and can be cooled with moving air.

A rack built for AI training packs far more computing into the same shelf and produces heat that air cannot carry away fast enough.

NVIDIA’s GB200 NVL72, which holds 72 of its Blackwell chips in one rack, is described by NVIDIA itself as a liquid-cooled design for exactly this reason.

Density is why the buildings had to be redesigned rather than simply multiplied.

Liquid cooling. Running coolant through pipes directly to the hot components of the server instead of blowing cold air at them.

Direct-to-chip cooling brings the liquid to a plate mounted on the processor.

Immersion cooling submerges the whole server in a fluid that does not conduct electricity.

Both exist because high-density racks generate more heat per square foot than air can remove, and both add plumbing to buildings that previously needed only fans.

Latency. The delay between a request and a response, measured in thousandths of a second. Distance creates latency, because signals take time to travel.

This is why a video call routes through a nearby facility while a large AI model can be trained in a remote one: the training run does not care about the delay, and the video call does.

Training and inference. Training is the one-time process of building an AI model (ChatGPT, Gemini, Claude, etc) by running enormous amounts of data through it, which takes months and vast amounts of electricity in one place.

Inference is what happens every time someone asks the finished model a question, which takes a fraction of a second and happens billions of times.

Training built the industry’s early demand.

Inference is the demand that keeps growing after the model is finished.

Redundancy, N+1, 2N. Spare equipment held ready so that a failure does not stop the building. N is the amount of equipment needed to run at full load.

N+1 means one spare unit beyond that.

2N means a complete duplicate of everything. Redundancy is most of what separates a data center from a warehouse, and most of what it costs.

Tier III and Tier IV. A certification system run by the Uptime Institute, an independent body, that rates how much of a building can be repaired or replaced while it keeps running.

Tier III means maintenance can happen without shutting anything down.

Tier IV adds tolerance for an unplanned equipment failure on top of that.

Operators cite the ratings in marketing, and buyers of space treat them as a floor rather than a distinction.

PUE, or power usage effectiveness. A ratio comparing all the electricity a facility draws against the portion that reaches the computers.

A score of 1.0 would mean every unit of power went to computing and none to cooling or lighting. Real buildings score above that, and the gap is the overhead.

PUE is the industry’s headline efficiency number and the one most often quoted without explanation.

WUE, and water. Water usage effectiveness measures liters of water consumed per unit of computing.

It exists because one common way to cool a building cheaply is to evaporate water, which trades electricity for a local water supply.

This is the term at the center of most community objections to new sites, particularly in dry regions, and operators are now asked about it in public hearings.

Greenfield, brownfield, retrofit. Greenfield means building on empty land.

Brownfield means building on a site that was previously industrial, often chosen because it already has a heavy power connection.

Retrofit means rebuilding the inside of an existing data center, which is now common as older buildings designed for low-density racks are converted for AI.

Speculative build. Starting construction before any customer has signed for the space.

The developer is betting demand will arrive. In a slow market this is how companies fail.

In the current market it is how they win, because a building already under construction can be delivered years sooner than one that starts after the contract is signed.

Build-to-suit. The opposite arrangement, where a customer commits first and the developer builds to that customer’s specification.

Lower risk for the developer, and a longer wait for the tenant.

Most large campuses are one of these two things, and knowing which tells a reader how much risk somebody took.

Powered shell. A finished building with electricity brought to it and very little else inside.

The buyer installs its own cooling and electrical fit-out.

It is sold this way because the shell and the power are the slow parts to obtain, and the equipment inside can be added quickly by whoever will operate it.

Commissioning. The testing period after construction and before customers move in, when every system is deliberately failed to confirm the backups work.

It takes months and it is the last thing to be compressed when a schedule slips.

Transformer, switchgear, UPS, generator. The chain of equipment between the public power lines and the chips.

A transformer changes electricity from the high voltage used to move it long distances to the lower voltage a building can use.

Switchgear routes and isolates it. A UPS, or uninterruptible power supply, is a bank of batteries that carries the load for the seconds between a grid failure and the generators starting.

The generators, usually diesel, run the site until the grid returns.

Transformers in particular now have delivery waits measured in years, which has made a piece of electrical equipment into a scheduling constraint on the AI buildout.

Subsea cable, dark fiber, cross-connect, peering. How a building talks to the world.

Subsea cables are the fibre-optic lines on the ocean floor that carry almost all intercontinental internet traffic.

Dark fiber is cable already laid in the ground but not yet lit with equipment, which a buyer can lease and light itself.

A cross-connect is a physical cable joining two customers inside the same facility.

Peering is two networks agreeing to exchange traffic directly rather than paying a third party to carry it.

AI factory. A term promoted by chip makers and adopted by operators for a facility built for AI training rather than general computing.

It reflects a real design shift, mainly much higher density and widespread liquid cooling. But it is also a marketing term, so readers should treat it as a description of intent, not a technical standard.

If a reader takes one thing from this section, it is that the building is measured in electricity rather than in space.

Every argument in the next section follows from that, and missing it is why the phrase “a 500 MW data center” reads as a fact about size when it is a fact about consumption.

The power

This vocabulary comes from the electricity industry, which is regulated, slow-moving, and public. It is also where the AI story now actually happens.

Megawatt and gigawatt. A watt measures the rate at which electricity is used. A megawatt is a million watts and a gigawatt is a billion. Scale it with a household.

The US Energy Information Administration puts average annual consumption for a residential utility customer at 10,791 kilowatt-hours.

A facility drawing 100 megawatts continuously for a year consumes roughly what 80,000 American homes consume in that year.

When a company announces a gigawatt campus, it is describing something on the order of a mid-sized city’s electricity demand.

The grid. The network of power stations, high-voltage lines, and local wires that delivers electricity to everyone connected to it.

Its defining property is that supply and demand must match continuously, second by second, because electricity in bulk cannot be stored easily.

Every constraint in this section comes back to that property.

Interconnection, and the interconnection queue. Interconnection is the permission and the physical work required to connect a new power plant or a new large customer to the grid.

The queue is the waiting line for it, and it is long. The Lawrence Berkeley National Laboratory publishes the count in its Queued Up report.

The 2026 edition, released in June 2026 with data through the end of 2025, found over 2,060 gigawatts of generation and storage capacity actively waiting to connect.

It also found that for projects finished in 2025, the median wait from application to operation ran past five years.

Hold that against a construction schedule of roughly two years, and the shape of the whole sector becomes clear: the building is the fast part.

Utility, ratepayer, tariff, rate case. The utility is the company that delivers electricity in a given territory, usually a regulated monopoly.

Ratepayers are its customers, meaning households and businesses.

A tariff is the published schedule of what each class of customer pays. A rate case is the public proceeding where a utility asks its regulator to change those prices.

Data centers have made rate cases newsworthy, because the question of who pays for new power lines and new generation is settled in them, in public, with transcripts.

Transmission and substation. Transmission is the high-voltage long-distance network that moves bulk electricity between regions.

A substation is the facility where voltage is stepped from transmission levels down to something a customer can use.

A site with a substation nearby is worth substantially more than one without, which is why land near existing electrical infrastructure has been bought aggressively.

Energization and time-to-power. Energization is the moment a site is actually connected and can draw electricity.

Time-to-power is the industry’s phrase for how long that takes from the decision to build.

It has replaced construction cost as the number that decides where projects go, and it is why regions with spare grid capacity now attract investment that would once have followed cheap land.

Stranded power. Electricity that exists but cannot reach a customer, usually because the wires to carry it were never built or the plant sits far from demand.

Developers hunt for it, because a site next to stranded power can skip years of waiting.

This word does a second job elsewhere in finance, which is covered at the end.

Offtake and offtaker. An offtake agreement is a long-term contract in which a buyer commits in advance to purchase what a project produces. The offtaker is that buyer.

The term comes from mining and energy, and it matters because a signed offtake is usually what allows a project to raise money at all.

Lenders are financing the contract as much as the asset.

PPA, or power purchase agreement. A specific kind of offtake for electricity. A buyer agrees to purchase power from a particular plant, at an agreed price, for ten or twenty years.

The technology companies use these to fund new wind, solar, and nuclear capacity, because a signed twenty-year buyer is what makes a plant financeable.

Baseload and firm power. Baseload is electricity available continuously rather than when the weather permits.

Firm power is a contractual promise that a specified amount will be there when called for.

Data centers need both, because a training run interrupted at hour 600 is a very expensive failure.

This requirement is why nuclear generation returned to the conversation after two decades of decline.

Behind-the-meter. A power arrangement where generation sits on the customer’s side of the utility’s meter, so the electricity never travels on the public grid.

In front of the meter means the ordinary arrangement where power is bought from the grid. Behind-the-meter is attractive because it can bypass the interconnection queue.

It is contentious because a large customer generating its own power may contribute less to the shared cost of the network that everyone else still relies on, and regulators are actively arguing about it.

On-site generation, gas turbines, fuel cells. Power produced at the site itself. A gas turbine is essentially a truck-sized jet engine burning natural gas to spin a generator.

A fuel cell converts fuel to electricity chemically, without combustion.

Both are being installed at data centers as bridge power, meaning a temporary supply while the grid connection is waited out, and both attract local objection over air quality.

Nuclear, and SMR. Nuclear plants supply large amounts of continuous, carbon-free electricity, which matches what a data center needs almost exactly.

SMR stands for small modular reactor, a smaller design intended to be factory-built and delivered rather than constructed on site.

Several are under development and none is yet delivering power at commercial scale in the United States, so announcements in this area describe intentions on a timescale of the 2030s.

Curtailment. Deliberately reducing output or consumption because the grid cannot handle it at that moment.

Wind farms are curtailed when they generate more than the wires can carry.

Data centers are now being asked to accept curtailment of their own demand in exchange for faster connection, which turns a computing facility into a flexible participant in the grid rather than a fixed drain on it.

Battery storage and microgrids. Battery storage holds electricity for minutes or hours, smoothing the gap between when power is generated and when it is needed.

A microgrid is a local system that can disconnect from the main grid and run independently.

Both appear in data center plans as ways to reduce dependence on a connection that may not arrive on schedule.

Carbon-free energy and RECs. A renewable energy certificate is a tradeable document representing one unit of electricity generated from a renewable source.

Because electrons cannot be traced through wires, certificates are how a company claims renewable power.

Buying certificates from a distant wind farm and matching them against consumption in a different region is called annual matching.

Matching consumption to carbon-free generation in the same place and the same hour is called 24/7 carbon-free energy, and it is a much harder claim to make.

Load growth. The rate at which total electricity demand rises. It was close to flat in the United States for roughly two decades, which shaped how utilities planned, hired, and ordered equipment.

It is now rising again, and much of the current strain follows from an industry that spent twenty years optimizing for a world that stopped existing.

ERCOT, PJM, and what an RTO is. A regional transmission organization operates the grid across a multi-state area and runs the market that sets wholesale electricity prices there.

PJM covers thirteen states across the mid-Atlantic and Midwest.

ERCOT covers most of Texas and is notable for being largely separate from the rest of the country’s grid.

A reader meets these names constantly in data center coverage because the region determines both the wait and the price.

Moratorium. A pause or ban on new data center connections or approvals, imposed by a local government or a utility. Several jurisdictions in the United States and in Europe have adopted one.

Moratoriums are the clearest evidence that the constraint on this buildout is no longer capital or chips, and the reasons cited are usually electricity prices, water, and land use.

The thing to carry out of this section is that the electricity system was built to serve demand that grew slowly and predictably, and it is being asked to serve demand that does neither.

When you see a story about an AI project delayed, the delay is far more often a power problem than a construction or funding problem.

The money

This vocabulary comes from commercial real estate and project finance. It is essential for understanding why buildings get built and appears throughout industry earnings reports and press releases.

Capex and opex. Capital expenditure is money spent to build or buy something lasting. Operating expenditure is money spent running it.

The enormous figures announced by technology companies for data centers are capex, which is why they appear as investment commitments rather than as costs against current profit.

Lease, tenant, anchor tenant. A data center operator is a landlord, and the arrangement works much like renting an apartment, with a contract, a term of years, and a monthly payment.

The tenant is the company renting. An anchor tenant is a large one whose commitment makes the rest of the building financeable, in the same way a department store anchors a shopping mall.

Triple net. A lease in which the tenant pays the taxes, the insurance, and the maintenance on top of the rent.

The landlord’s income is therefore closer to predictable, because most of the costs that could vary belong to somebody else.

Long triple-net leases with large technology companies are the reason data centers can be financed with debt on terms usually reserved for far duller assets.

Take-or-pay. A contract in which the customer pays whether or not it uses what it reserved.

The phrase is literal. Take-or-pay commitments underpin most large power and capacity deals in this sector, and they are what a lender is really lending against.

Backlog, contracted revenue, RPO. All three describe money a company has signed for but not yet earned.

RPO stands for remaining performance obligations, a formal accounting measure of contracted revenue not yet delivered.

These figures appear prominently in cloud company results, and they matter because they indicate demand that is committed rather than forecast.

Pre-leasing and lease-up. Pre-leasing is signing tenants before a building is finished.

Lease-up is the period after completion during which remaining space fills. A campus that is fully pre-leased before it opens tells a reader something about demand that no forecast can.

Vacancy and absorption. Vacancy is finished capacity with nobody in it. Absorption is the rate at which new capacity gets taken up. In most property markets these move independently.

In this one, vacancy in the major markets has been unusually low, which is why so much construction begins without a signed customer.

Capital stack. The layered set of money funding a project, ordered by who gets repaid first.

Senior debt sits at the bottom and is repaid first at the lowest return. Equity sits at the top, is repaid last, and takes the loss if the project disappoints.

Every announcement about a project being financed is a statement about who occupies which layer.

Project finance and non-recourse. Project finance lends against a single asset’s own future income rather than against the borrower’s whole balance sheet.

Non-recourse means that if the project fails, the lender can take the project and nothing else.

This structure allows very large facilities to be built by companies that could not borrow those sums against themselves.

REIT. A real estate investment trust, a company that owns income-producing property and is required to distribute most of its earnings to shareholders in exchange for favorable tax treatment.

Equinix and Digital Realty are the two large listed data center REITs, and they are the reason a reader can find data center ownership inside an ordinary retirement fund.

Private credit. Lending by investment funds rather than by banks.

It has grown substantially since the 2008 financial crisis, and it is now a major source of data center construction funding, partly because these loans are larger and more bespoke than banks readily provide.

ABS, securitization, asset-backed. Securitization is the practice of bundling the future payments from a set of contracts into a bond that investors can buy.

ABS stands for asset-backed securities, the bonds themselves.

Data center leases have become a recognised category of collateral for this, which means the rent from a specific building is being converted into a tradeable instrument bought by pension funds and insurers.

Joint venture. Two or more parties forming a shared entity to develop something neither would fund alone.

Nearly every very large campus announced in the past two years has been structured this way, typically pairing a developer that knows how to build with an investor that supplies the money.

Cap rate. Short for capitalization rate. Annual income divided by the price paid, expressed as a percentage.

It is the standard shorthand for how expensive a property is relative to what it earns.

A lower cap rate means a higher price for the same income, and the number is quoted constantly in property coverage without ever being defined.

Stabilized, and stabilized yield. A building is stabilized once it is fully leased and running normally rather than filling up.

Stabilized yield is the annual income at that point divided by the total cost to build.

The gap between what a completed building yields and what an equivalent one costs to buy is where development profit comes from.

IRR. Internal rate of return, the annualized return on an investment, accounting for when the money went out and when it came back.

Timing is the point: the same total profit earned in three years produces a much higher IRR than in ten.

A reader will see it quoted alongside EBITDA, which measures operating earnings before interest, tax, and the accounting charges for wear and obsolescence.

The two answer different questions, and they are routinely presented as though they answered the same one.

Depreciation and useful life. Depreciation is the accounting practice of spreading the cost of equipment across the years it is expected to work. Useful life is the assumed number of those years.

This obscure pairing is now central to the AI boom debate: if chips wear out or become obsolete faster than assumed, industry profits may be overstated relative to the underlying economics.

Anyone following the debate about whether the buildout is overstated is following an argument about this number.

Circular deals. Arrangements in which a supplier invests in its own customer, or a company’s revenue is partly funded by its own investment.

Several such arrangements between chip makers, cloud providers, and AI developers have been announced and publicly disputed.

Supporters argue that suppliers funding a new market is ordinary practice in capital-intensive industries and has precedent in telecoms and semiconductors.

Critics argue the structure can make demand appear more independent than it is, and that the same practice preceded losses in earlier technology cycles.

The disagreement is genuine and unresolved, and a reader should know the term because it will keep appearing.

What this section shows is that the buildings get built because somebody signed a contract long enough to borrow against.

When you read that a project is “financed,” the real event is that a technology company promised to pay for fifteen years, and everything else followed from that promise.

Five words that do two jobs

These are the words that cause the most trouble, because none of them looks like jargon. A reader who has understood everything above can still be defeated by a headline built on one of these.

Load. To an electrical engineer, load is demand for electricity. To a computing person, load is the amount of work a machine is doing. When an article says load growth, it means the first.

When it says workload, it means the second. Both appear in the same paragraph regularly.

Tier. Three unrelated meanings. A Tier III building is a reliability rating from the Uptime Institute. A Tier 1 market is a mature metropolitan area with established data center supply, such as Northern Virginia.

A tier of the capital stack is a layer of financing. Nothing connects the three except the word.

Capacity. Grid capacity is how much electricity the wires can carry. IT capacity is how much computing a building can house.

Contracted capacity is how much a customer has already committed to pay for.

Three different numbers, all quoted in megawatts, all called capacity, and frequently compared to each other as though they measured the same thing.

Stranded. Stranded power is electricity that exists and cannot reach a buyer, which is an opportunity for whoever can get to it.

A stranded asset is an investment that loses its value before the end of its expected life, which is a loss.

The word points in opposite directions depending on which industry is speaking.

Platform. In investment language, a platform is the first company an investor buys and then builds around through further acquisitions. In technology, a platform is a service other people build on, such as a cloud.

In development, a platform is the physical campus itself. All three usages are common in the same coverage.

Very little of this vocabulary was written to be understood by the public.

It was written by three professions talking to themselves, and it stayed private because for most of its history nothing about it touched anyone outside the industry.

That has changed. The International Energy Agency estimated that data centers consumed about 415 terawatt-hours of electricity in 2024, or 1.5% of global demand, and projects that consumption will more than double to roughly 945 terawatt-hours by 2030.

Decisions of that size are made in rate cases, in zoning hearings, and in interconnection queues, and every one of those proceedings is conducted in the words above.

They are open to the public and they are effectively closed to anyone who cannot read them.

The AI that answers a question in a second sits on top of a building somebody had to power, a grid somebody had to expand, and a contract somebody had to sign for fifteen years. Now you can read the sentences that describe it.

Which of these words had you seen a dozen times without ever knowing what it meant?

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• The EuroHPC tender caps the Union contribution at 17 percent of IT capital expenditure against facilities EuroHPC estimates at €4 to €5 billion each. That places the qualifying constraint on capital and power rather than on operating capability, and it is why Polarise cannot bid alone despite already running AI infrastructure at scale.

• Two of the four locations Polarise names, Oslo and the United Kingdom, cannot host an EU AI Gigafactory. The tender requires the hosting entity to be established in a participating state that is an EU member state. Polarise’s eligible offer is its two Bavarian sites and a potential third in North Rhine-Westphalia.

• Polarise built the Munich AI Factory behind Deutsche Telekom’s Industrial AI Cloud, and Deutsche Telekom has partnered with Brookfield on a separate German gigafactory bid. Polarise’s most significant disclosed customer relationship sits inside a competing consortium for the same lot.

Seventeen Percent Is A Capital Gate Wearing A Subsidy Label

EuroHPC estimates the investment needed per AI Gigafactory at a minimum of €4 to €5 billion and caps the Union contribution at 17 percent of IT capital expenditure.

Coverage has read that cap as a measure of how much Europe is contributing.

It is more useful read as a filter. Construction, utilities, site preparation, and operating costs fall outside the base, meaning bidders must cover roughly 90% of a multi-billion euro build before public funding becomes relevant.

Polarise states it already has prepared sites and technology capable of delivering capacity levels the tender framework envisions only for later build-out stages. That claim, if it holds, is an operating credential.

The tender does not gate on operating credentials.

It gates on capital, secured power at 120 MW or 150 MW of IT load in Phase 2, and delivery inside 18 months of signature.

A specialist that can run the asset but cannot fund it enters the process as a component of somebody else’s consortium rather than as a principal.

Over the next 12 to 24 months this determines who captures the value in these seven projects.

The capital sponsors take the equity. The specialists take a contract.

Polarise’s Operating Map And Its Eligible Map Are Different Maps

Polarise operates or is developing five AI Factories in Europe, at locations it names as Oslo, Frankfurt, Munich and the United Kingdom, with colocation capacity scaling to 120 MW at some sites.

That footprint is the company’s strongest evidence of execution.

Most of it cannot count toward hosting.

The tender requires the hosting entity to be established in a participating state that is also an EU member state, and requires the physical facilities to sit there.

Norway and the United Kingdom fail that test regardless of how capable the assets are.

Half the named locations in Polarise’s footprint are therefore evidence of competence rather than eligible capacity.

What remains eligible is what the company has offered: two sites in Bavaria and a potential third in North Rhine-Westphalia, described as prepared rather than as operating gigafactory-scale facilities.

The gap between an operating map and an eligible map is a general reading problem in this tender.

Several credible European operators hold their most mature assets in the Nordics and the United Kingdom, and a track record built there does not carry into a hosting bid.

The Customer Relationship Sits Inside A Rival Bid

Polarise’s clearest commercial validation is the AI server infrastructure it built and operates for the Deutsche Telekom-backed AI Factory in Munich, launched in late 2025 to support Deutsche Telekom’s Industrial AI Cloud.

Deutsche Telekom has separately partnered with Brookfield on a German gigafactory bid, confirmed on the record by Deutsche Telekom’s chief executive.

That places Polarise’s most significant disclosed anchor relationship inside a consortium it is not part of, competing for the same German host position.

Two paths follow. Polarise joins that consortium as the operating partner, in which case its independent announcement is a positioning move ahead of a negotiation.

Or it assembles a rival group, in which case it competes against the customer whose facility it built.

Polarise’s own framing, that the programme is an accelerator rather than a prerequisite and that it remains open to partners, is consistent with either.

The structural point outruns this company.

Germany is one host country with a limited pool of parties holding both capital and secured power, and the same names recur across bids.

Partner scarcity, not applicant scarcity, sets how many credible German tenders reach November 12.

What Each Segment Now Prices

Private Capital. Infrastructure funds looking at this programme are not buying into a subsidised platform.

They are buying merchant AI compute with a five-year public tenancy over part of the revenue, and the operating layer is a procurable input.

Polarise’s position illustrates the trade: a fund that brings the equity can contract the operating capability rather than pay for it in valuation.

Waiting for the early-2027 awards means negotiating a consortium stake after the risk split is already fixed in the bid documents.

Public Markets. Deutsche Telekom is the listed name most exposed here, through both the Brookfield bid and the Munich Industrial AI Cloud.

The unresolved disclosure question is whether a gigafactory commitment would sit on balance sheet or inside a project company.

An award headline tells shareholders the company won something.

It does not tell them what the company now owes, or over what term.

Operators. For data center developers, the tender is a specification.

It requires 120–150 MW of operational IT load in Phase 2, readiness within 18 months, an EU-incorporated and EU-controlled coordinator, and no core technical operations outsourced outside the EU.

Operators holding secured grid capacity at that scale in an eligible member state hold the scarce asset in every one of these consortia.

Operators holding it in Oslo or the United Kingdom hold a strong business and no hosting bid.

The Scarce Input Is Eligible Power, Not Willing Bidders

Bids close November 12, with awards in early 2027 and construction the same year.

The programme will not be short of interest, and the earlier expression-of-interest round drew 76 proposals across 16 member states.

It will be short of parties that combine several billion euros of capital with secured grid capacity inside an eligible member state on an 18-month clock.

Polarise announcing a bid while openly seeking partners is the first public confirmation of how that constraint behaves.

The question the next three months answer is whether Germany produces two credible consortia or one, and what that does to the price the winning group pays for the operating layer it cannot build itself.

Join Global Data Center Hub

Welcome to Global Data Center Hub. Join investors, operators, and innovators reading to stay ahead of the latest trends in the data center sector in developed and emerging markets globally.

Join Global Data Center Hub

TL;DR

• CoreWeave announced 360 megawatts of contracted IT power across three Indonesian facilities on August 4, 2026, and stated it will own and operate the compute environment across all three sites. The release names no landlord, no developer, and no power counterparty. CoreWeave’s ownership claim is scoped to the compute, and the facilities are targeted online in 2028.

• Microsoft opened its India South Central region in Hyderabad on August 6, 2026, on land it assembled itself from 2022 across more than 110 acres. Microsoft owns the land, operates the region, and is its own tenant, backed by a $20.5 billion commitment running through 2029. Amazon Web Services has run a Hyderabad region since 2022 on $2.77 billions of announced investment.

• CoreWeave held more than 3.5GW of contracted power against more than 1GW of active power as of March 31, 2026, across 49 data centers. The Indonesian 360 megawatts adds to the contracted book rather than the active one. Contracted power is an obligation with a delivery date, and the counterparty carrying that obligation in Indonesia is not disclosed.

Where the $20.5 billion went: Microsoft Opens India South Central In Hyderabad On $20.5 Billion

Three Facilities, 360 Megawatts, And One Carefully Written Clause

CoreWeave will own and operate the compute environment across all three Indonesian sites.

That is the company’s own language, and the scope of it is the story.

It does not say CoreWeave owns the facilities or power. It specifies what the company holds and, by omission, what others hold.

Two days later Microsoft opened a region in Hyderabad built on land it bought itself, starting in 2022, across more than 110 acres in Chandanvelly, the Ellikatta and Mekaguda area, and Kottur and Shadnagar.

Microsoft owns the ground, runs the region, and fills it with its own customers. There is no landlord in that structure because Microsoft is the landlord.

Same continent, same demand, two days apart, and the capital sitting underneath could not be more different.

One company put its balance sheet into the dirt. The other contracted for the output and kept its capital in the compute.

Both are legitimate. They are not the same risk, and Asian markets are now absorbing both at once.

The Neocloud Model Needs A Landlord, And Nobody Has Named This One

CoreWeave has done this before and named the partner.

Its Swedish expansion runs with Conapto, named in the announcement.

Indonesia names nobody. Either the shell partner is not signed, or it is signed and undisclosed, and the release does not let a reader choose between those.

That gap matters because the party building three facilities to hold 360 megawatts by 2028 is the one carrying construction risk, power procurement risk, and Indonesian permitting risk.

CoreWeave carries the take-or-pay obligation and the customer relationships.

Whoever builds carries the delivery. In a market where CoreWeave has no operating history, the identity of that counterparty is most of the execution question.

Over the next 12 to 24 months, expect the developer side of neocloud expansion in Southeast Asia to become the more interesting side to underwrite.

The compute operator is listed and its disclosures are public. The people building the buildings are frequently neither.

Contracted Power Is A Forward Book, Not Capacity

As of March 31, 2026, CoreWeave held more than 1GW of active power and more than 3.5GW of contracted power across 49 data centers.

The Indonesian 360 megawatts joins the second number. It is not capacity today and will not be capacity until 2028.

Microsoft’s $20.5 billion is also forward. The $17.5 billion tranche runs 2026 through 2029. But the instruments differ.

Microsoft has committed capital it controls, against land it already owns, in a region already live and already serving HDFC Bank, Bajaj Finserv, Adani Group, Apollo Hospitals, and Air India.

CoreWeave has committed to pay for power that does not yet exist, in facilities not yet built, for customers not yet named.

Both companies are announcing the future. One is announcing spend it can pace.

The other is announcing an obligation with a date on it.

An investor reading both as “Asian expansion” has collapsed a distinction that will matter a great deal in 2028.

The Grid Decides Whether Either Number Is Real

Telangana solved this problem in public.

The state is deploying a new 220kV grid supply, adding substations, and building flood-water canals for the Microsoft buildout.

That is a visible commitment from a counterparty with obligations, and it arrived years before the region opened.

Indonesia has no equivalent in the public record for these three sites.

No power counterparty is named, no interconnection commitment is disclosed, and no locations are given.

The demand case is real: enterprises, AI-native companies, and governments across Asia need compute close to their data and users, driven by latency and data locality requirements.

That is CoreWeave’s stated rationale and it is correct. The constraint is not demand.

For anyone underwriting either market, the grid is where the announcement stops being an announcement.

Hyderabad has shown its work. Indonesia has not yet.

What Each Segment Now Has To Underwrite

Private capital. The Indonesian opportunity in this announcement is not CoreWeave. It is whoever builds three facilities to hold 360 megawatts by 2028.

That party has a listed, contracted anchor tenant, which is the strongest position a Southeast Asian developer can hold. Find out who it is.

Funds waiting for CoreWeave to disclose the partner will be bidding against people who already knew, on assets that are already spoken for.

Public markets. Read CoreWeave’s contracted power line as a liability schedule alongside a growth figure.

More than 3.5GW contracted against more than 1GW active means the gap between the two is the company’s forward obligation, and every new market adds to it before it adds to revenue.

That is not a warning. It is the model working as designed, and it needs to be modeled that way rather than read as a backlog.

Operators. Indonesian and regional developers now have a named, listed anchor tenant looking for 360 megawatts of delivery by 2028 in their market.

That is a specific, dated, financeable demand signal, and it is rare.

Operators who treat this as a competitor announcement rather than a customer announcement will read it exactly backward.

Two Structures Meet In Asia, And Only One Has Been Tested Here

Hyperscaler self-tenancy has a track record in Asian emerging markets.

Microsoft has four Indian regions, and Amazon has run Hyderabad since 2022. The model is slow, capital-heavy, and it works.

The neocloud model has no such record in Southeast Asia.

CoreWeave’s Indonesian entry is the first serious test of whether a contracted-compute operator can site capacity in an emerging Asian market on someone else’s buildings and someone else’s power.

The answer arrives in 2028, and the leading indicator arrives much sooner, the moment a developer and a power counterparty are named.

The open question is which model the region’s own capital chooses to back.

Southeast Asian institutions can finance a hyperscaler’s shell or a neocloud’s shell, and those are different credits with different tenants behind them.

Nothing in either announcement tells them which one Jakarta is getting.

Join Global Data Center Hub