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

A fiber platform breaks the quarter its upgrade bill swallows its operating cash, and that can happen while traffic and EBITDA both grow.

The old assumption was that fiber is a build-once asset: lay the glass, then collect rent for decades.

Investors who still price fiber that way will pay a growth multiple for a business that spends almost everything it earns to stay current.

Lumen shows the gap in a single quarter. In Q2 2026, Lumen reported $971 million in operating cash flow and $902 million in capital expenditures, according to its earnings release filed with the SEC on 4 August 2026.

Free cash flow was $69 million. Those are company-wide numbers. They do not benchmark data center fiber on their own. But the point stands.

A large fiber network can generate nearly a billion dollars in operating cash in a quarter and keep very little of it.

The pressure is building because compute is moving.

TeleGeography’s 16 June 2026 analysis found that power and compute availability now shape where long-haul networks need capacity.

The same analysis says AI traffic cannot be separated from the rest of the traffic mix. So, the demand is real, but nobody can measure it cleanly.

Capital is moving toward platforms that can add capacity fast, and away from routes that don't end at the new campuses.

Every Fiber Upgrade Restarts the Capex Clock

Fiber modernization is rarely a rip-and-replace job. It is a stack of separate spending decisions, and each one has its own test.

The cheapest layer is equipment. New transponders and optical line systems push more capacity through glass already in the ground.

Nokia’s 19 April 2023 announcement described C+L-band optical systems for Lightstorm in India, which widen the band of light a single fiber can carry.

The next layer is more cable in ducts the operator already owns.

Lumen’s Private Connectivity Fabric material describes a cable design that fits more fiber into existing conduit. The most expensive layer is new routes: new ducts, road and rail crossings, access rights, and building entries.

Here is the catch. Equipment upgrades are fast and cheap per bit, but they do not last.

Optics get refreshed. Line systems age. Each refresh creates a new bill for the same route, often for the same customer. So, a better cost per bit can sit right next to a higher total capital requirement.

The capacity gain makes the headline. The refresh schedule sets the cost.

Procurement adds a second clock. Zayo’s 20 August 2026 announcement expanded a long-term Corning agreement to secure a substantial share of the cable for its planned build.

The disclosed benefit is supply assurance. A locked supply contract protects a schedule. If the campus or the permit slips, that same contract becomes a commitment with nothing to install.

Capacity Gains Arrive Before Fiber Platform Returns

The industry has long assumed that traffic growth pays for capacity. That assumption held when a route, once lit, carried growing volumes at steady prices for years.

It no longer holds because prices on some products keep falling while the network must refresh to keep its customers.

TeleGeography’s 16 September 2026 pricing analysis documents continuing IP transit price declines, at different levels and speeds by market.

Read the scope closely. It covers IP transit only. It excludes dark fiber, dedicated wavelengths, local access, and installation fees.

So, the lesson is narrow, and it is sharp. Volume and price move separately, and each product carries its own price risk.

A fixed-price dark fiber contract and a usage-based transit service do not share one price curve. A renewing wavelength deal sits somewhere between them.

The facts here are public. What gets lost is the consequence. Growth in bits and growth in cash are two different lines. A model that blends them will overpay.

Share

Financing Headlines Overstate the Capital Left for Upgrades

The second cost is time, and financing pays for it. Zayo’s own announcement names planning, permitting, materials, and execution as parts of long-haul delivery. Its build targets are company forecasts.

A delayed project does not take one hit. It takes several at once. Customer payments arrive late. Old and new systems run side by side for longer. Interest keeps accruing, and contract remedies may apply. Run those as one linked downside case, because they arrive together.

That is why financing totals mislead. On 16 October 2025, EXA Infrastructure announced more than €1.3 billion of refinanced and new debt facilities over seven years, to support expansion and acquisitions. Part of that total replaces debt that already existed.

On 27 August 2024, euNetworks closed a €2.1 billion equity recapitalization. It involved a Stonepeak-managed vehicle anchored by Mercer and Aware Super, with IMCO and APG investing alongside.

The announcement does not split that total between shareholder liquidity and new investment.

Neither headline is a modernization budget. The key number is the investment capacity left after refinancing.

Customer prepayments carry the same trap. S&P Global Ratings, in a 5 March 2026 analysis, identified upfront hyperscaler payments to Lumen as a source of liquidity.

Liquidity and profit are different things. The same dollar cannot fund construction and also count as distributable earnings.

Lumen’s own release shows why the definitions matter. Free cash flow excluding special items was $327 million, with $258 million in cash special items removed.

That is a different measure from the $69 million, and neither one is a fiber modernization figure.

Private Capital Owns the Upgrade Bill at Exit

This problem lands hardest on private capital: infrastructure funds, private equity, and private credit that buy or lend against fiber platforms.

The binding variable is cash flow after upgrades. Fiber platforms are often valued on EBITDA multiples, and EBITDA does not see capex.

If the refresh cycle is real and recurring, a low normalized capex number in your model inflates the price you pay and the debt you size. You find out at exit, when the buyer rebuilds the capex schedule from scratch.

Scale makes the stakes larger. DigitalBridge put an approximate $4.25 billion value on Zayo’s acquisition of Crown Castle Fiber Solutions in its 13 March 2025 announcement, and Zayo completed the deal on 1 May 2026.

Deals of that size buy metro density and route reach. Any acquired network also brings its own overlap with the buyer’s routes and its own future upgrade needs.

So, start every fiber deal with the bottleneck. Find the constraint that limits the route and then find out who pays to remove it and what contracted cash follows.

Leave a comment

Ask Which Bottleneck the Upgrade Capex Removes

Five checks would catch the failure before the money goes out.

First, name the bottleneck. Ask the engineers and the sales team to point at the same constraint on each route: tired optics, a full duct, an expiring access right, or a campus with no connection. Each is a different investment case. Run this at screening, before valuation work starts.

Second, survey the ducts. An asset register doesn't prove you can pull a cable through every segment. Require surveys of occupancy, condition, and continuity before the model counts spare conduit as cheap expansion. This belongs in confirmatory diligence.

Third, match the rights to the model. Record access rights segment by segment, with renewal and expansion terms. Do not value a route past the shortest enforceable right beneath it, a problem GDCH covered in its analysis of dark fiber residual value at IRU expiry. Do this in legal diligence, before any asset life enters the model.

Fourth, separate capex into maintenance, competitive refresh, contracted expansion, and speculative expansion schedules. Treat customer-required refreshes as mandatory, even when classified as growth. Include them in the base case before investment committee review.

Fifth, net the financing. For every headline financing or prepayment, compute what is left after refinancing, restricted cash, and completion obligations. Use only capacity the customer has accepted as the revenue milestone. Do this before you size debt or set distributions.

Policy can lower the cost of some of this work. The EU’s Gigabit Infrastructure Act, Regulation (EU) 2024/1309, became fully applicable on 12 May 2026, with rules on infrastructure sharing and coordinated civil works.

India’s Telecommunications (Right of Way) Rules, 2024 took effect on 1 January 2025 with a standard approval process and set timelines.

Neither guarantees that a given crossing gets built on schedule. Test reuse rights before you accept a greenfield cost estimate, then confirm with the local authority.

Route-Level Returns Remain the Missing Number

The open question is the one no public filing answers. No operator in this piece discloses realized returns by route after a full refresh cycle.

Lumen, Zayo, and the European platforms that raised billions all report at the company level.

Until that number exists, you are underwriting the refresh cycle on your own data or on faith.

The funds that build their own route-level cash record will price the next fiber platform correctly.

Everyone else will learn the refresh schedule from the buyer’s diligence team at exit.

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

Each month, I track the data center transactions shaping AI infrastructure across North America, Europe, Asia-Pacific, Latin America, and the Middle East and Africa. September produced 122 verified deals. Ten patterns stand out.

1. Anthropic was the month’s largest buyer of compute, on three continents.

Akamai signed seven-year project plans committing Anthropic to about $11.6 billion in payments. In Australia, Anthropic signed a long-term lease at Zerra DC’s Western Downs Digital Park, subject to foreign investment review, and in Texas it anchors the month’s top deal.

2. The longest commitments now run a decade or more.

Google signed a 22-year nuclear power purchase agreement with Fortum for the Loviisa plant in Finland. Blockfusion signed a 15-year anchor lease with CoreWeave at Niagara Falls, and Bitdeer AI a ten-year agreement for 65.1MW in Malaysia.

3. Gas is becoming the AI campus’s private power plant.

Aligned broke ground on its 2GW Project Phoenix in Pennsylvania with onsite gas power planned. INNIO booked a firm 450MW order for gas-engine prime power, and Woodway signed for a 22-mile pipeline feeding behind-the-meter generation at an undisclosed hyperscale site.

4. Debt came in four shapes.

VIRTUS closed £2.45 billion of bank debt from 13 lenders. Vantage closed a $2 billion revolving credit facility, NEXTDC priced A$1.1 billion of convertible notes, and ASP Data Center raised NOK 1.6 billion in Norwegian bonds.

5. Nscale played three roles in one month.

It became preferred compute provider to Figure, the physical AI developer, under an initial $3.5 billion commitment in Barstow, Texas. It filed for a New York Stock Exchange listing, and it closed a $2.36 billion first tranche of a $3.36 billion convertible facility, with $1 billion committed by NVIDIA.

6. Equity investors paid up for AI platforms.

Mistral closed a €3 billion Series D at a valuation above €21 billion and Crusoe raised $3 billion at a $30 billion valuation.

7. Governments rewrote the rules while deals were being signed.

Scotland now requires a full environmental impact assessment for new data centers above 50MW. Virginia restricted nondisclosure agreements with state agencies, California signed seven data center bills on electricity, water and land use, and Thailand paused sector development across 166 projects while it drafts national standards.

8. One conference week produced a cluster of Saudi commitments.

During LEAP 2026, MIS announced a $1.2 billion expansion to 192MW, and NHC Innovation set out an $880 million Khuzam Digital Valley program anchored by BytePlus and NAVER Innovation. AWS and HUMAIN committed up to 50MW at Saudi Arabia’s first AI Zone by 2028.

9. Big buyers are entering Asia through existing buildings and partners.

Amazon signed conditional agreements to buy two Grand Ming data center towers in Hong Kong for at least $278 million. OpenAI became an anchor customer at two Firmus sites in Malaysia, and Microsoft Taiwan is preparing to double its data centers to four.

10. The Nordics drew the largest corporate pledge and a completed platform sale.

Google committed at least €13 billion to Finnish digital infrastructure across four locations for 2027 and 2028. CPP Investments, Equinix and Partners Group completed their acquisition of atNorth at a $4.0 billion enterprise value, alongside a separate $4.1 billion financing package.

Leave a comment

The Top 10 Deals Globally

The ten highest-ranked deals split four in Europe, three in North America, two in Asia-Pacific and one in the Middle East and Africa. Latin America’s September activity sat earlier in the build sequence, in a new tax-incentive law, land reservations and grid applications, covered in the regional companion piece.

1. Anthropic–Lambda Texas AI Cloud Agreement (North America)

Anthropic, the AI lab behind Claude, agreed a $35 billion cloud computing deal with Lambda, the AI cloud provider, for roughly 350MW of capacity at Hut 8’s Beacon Point campus in Nueces County, Texas. NVIDIA, the chipmaker whose processors power most AI training, holds the underlying lease, and Lambda supplies the compute.

The $35 billion is contract value, which is revenue, not construction cost. The 350MW sits on capacity already leased at Beacon Point, so it counts once in any total. The largest AI labs now buy compute through layers of landlord, chip supplier and cloud provider.

2. PPC–AWS Agios Dimitrios Data Center (Europe)

PPC, Greece’s largest power utility, signed a binding memorandum with Amazon Web Services, Amazon’s cloud business, for a data center at Agios Dimitrios in Western Macedonia. AWS will lease the finished facility for 15 years. PPC provides the land, buildings and energy infrastructure, and AWS installs and runs the IT equipment.

The first phase is planned at 240MW of IT load within 300MW of total power demand, at an estimated cost of about €1.2 billion ($1.38 billion). A power company now earns 15 years of rent on the electricity and the building together.

3. Cerebras–Compute Nordic Mikkeli AI Capacity (Europe)

Cerebras Systems, the AI chipmaker, signed a seven-year capacity agreement with Compute Nordic Finland for 165MW of contracted IT capacity in Mikkeli. Construction is under way on the first 50MW, built for high-density Cerebras compute.

An analysis by Ramboll cited by Cerebras puts the full build at €1.0 billion to €1.7 billion; the lower figure, about $1.16 billion, is the disclosed lower bound. The deal shows a chipmaker locking in purpose-built capacity for its own hardware.

Share

4. Goodman Project Apollo, Macquarie Park (Asia-Pacific)

Goodman, the Australian industrial property group, received New South Wales planning approval for Project Apollo in Macquarie Park, Sydney. The development has about 135MW of total power demand and 185MW of diesel backup generation, with a development cost of A$1.365 billion (about $979 million). The site will run on grid power. No tenant has been named, and IT load is not disclosed.

5. Nabiax ADC3 Alcalá (Europe)

Nabiax, which operates the Alcalá de Henares data center campus near Madrid, broke ground on ADC3. The project adds about 100MW of planned IT capacity at an estimated €800 million (about $918 million), which is project cost, not closed financing. Scaling through 2029, ADC3 adds more than four times the campus’s existing 22.3MW of IT load.

6. Oracle Project Jupiter Renewable Power Tender (North America)

Oracle, the database and cloud computing company, issued a request for proposals for 2GW of new renewable generation in New Mexico to support its Project Jupiter AI infrastructure. Delivery is targeted for 2027-2031. No contracts have been awarded, and no value is disclosed. The tender is separate from the project’s earlier onsite fuel-cell design, and the two figures are not combined. Oracle is lining up clean power at gigawatt scale before it's needed.

7. Nostrum Evergreen Badajoz Power Authorization (Europe)

The Extremadura regional government granted a preliminary authorization for the electrical infrastructure of Data Riocaya’s Nostrum Evergreen AI campus at the Badajoz logistics platform in western Spain. Environmental filings show a total power demand of 170MW, including 110MW of critical load, and an estimated investment of €1.913 billion (about $2.22 billion). Full campus approval is still to come. A separate 300MW figure is grid connection capacity, not IT load. Securing the grid link first makes the rest of the campus possible.

8. ZanKore $3.1 Billion Term Loan (Asia-Pacific)

ZanKore, a Southeast Asian AI cloud platform, signed a senior term loan of up to $3.1 billion with regional and international lenders. The loan funds an initial 100MW of NVIDIA AI infrastructure in Indonesia. ZanKore’s longer-term target is 1GW. It is one of the month’s largest debt packages, and it shows banks willing to lend billions against AI compute in an emerging market.

9. HUMAIN–DataVolt Oxagon AI Data Center (Middle East and Africa)

HUMAIN and DataVolt, two Saudi AI infrastructure companies, broke ground on a 100MW AI data center at Oxagon in NEOM, the first development inside a planned 360MW first phase of DataVolt’s 1.5GW Red Sea campus. DataVolt’s chief executive put the cost at around $1 billion. Sources do not say whether the 100MW is IT load or total power, and financing was nearly secured, with no amount disclosed. The facility is designed to run on renewable energy and export computing capacity abroad.

10. DDC 01 Propco CoreWeave-Linked Secured Notes (North America)

DDC 01 Propco, backed by a Blue Owl, Cedarwood and PowerHouse joint venture, priced about $1.085 billion of senior secured notes for the Digital Drive data center in Chesterfield County, Virginia. The 76MW IT facility already has a 15-year lease with CoreWeave, the AI cloud provider. The notes are priced at 98.5% of face value with an 8.875% coupon, a yield of about 9.25%. The deal shows a long lease with an AI cloud tenant financed directly in the bond market.

Leave a comment

September reads as two markets running on different fuel.

In developed markets, the fuel was financial engineering. DDC 01 Propco turned a CoreWeave lease into secured notes, VIRTUS pulled 13 banks into one package, and Nscale used convertible notes that become shares when it lists.

In emerging markets, the state did the heavy lifting. Brazil enacted REDATA, a nationwide tax-incentive law for data centers. Vodafone Business and a partner agreed to develop Egypt’s first sovereign AI data center, and Saudi Arabia’s HUMAIN put its name on both a NEOM construction site and an AWS capacity target.

Indonesia breaks the pattern and leads rather than waits. ZanKore secured $3.1 billion from regional and international lenders, and BDx is building the first 120MW building of its CGK4 campus in West Java, with the first phase due in early 2027. Private builders are moving on private money.

The split may not hold. The UAE’s €40 billion state-backed program for Germany, which includes data centers, sends state money into a developed market, while Indonesian AI compute draws commercial bank lending. If state capital keeps flowing into developed markets and private debt into emerging ones, which model will build the next gigawatt faster?

Share 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

The Funding Question Is Back, and the Answer Changed

In Q1, this series argued that capital structure was the first deciding variable in AI infrastructure and cost structure the second.

Meta’s Q3 shows what happens when the second variable can’t carry the load. The problem goes back to the first, and this time the answer is equity.

Demand is not the constraint. Meta says it is building four superclusters of more than one gigawatt each and expects its infrastructure capacity to nearly double in 2026.

What binds is the physical delivery of that capacity: power, sites, equipment, and the cash to pay for all of it before it earns anything.

Meta’s net property and equipment rose 43.5% in nine months, to $253.2B.

That asset base now starts to depreciate, and the depreciation lands on income statements long before the new products built on it report revenue.

The Ad Engine Is Not the Problem

The stock sale is not the rescue of a weak business. Advertising revenue rose 25.9% to $63.1B, 97.6% of the total.

Ad impressions grew 10% and the average price per ad grew 14%, so Meta sold more ads and got paid more for each one.

Revenue beat the $63.34B consensus and cleared the top of Meta’s own $61B to $64B guidance.

AI is paying off inside ads, measurably. Meta reports a 9.6% lift in app-ad conversions from expanding its GEM ranking model, and partnership ads now run above a $15B annual rate after growing more than 97%.

These gains show the infrastructure already earns a return where Meta has a business to put it in.

They don’t show a return on the whole program, because Meta doesn’t disclose one.

The Payroll Swap Did Not Close the Gap

In Q1, Meta framed compute and people as its two main cost centers and cut 8,000 roles to shift spending from one to the other.

The cut happened. Headcount ended Q3 at 66,775, down 15% from a year ago.

Costs grew anyway. Excluding a $9.89B legal charge tied to an agreement with 52 state attorneys general, expenses rose 35.9% to $41.7B, well ahead of 26.1% revenue growth. Cost of revenue rose 45.2%, depreciation 44.1%, and research and development 32.4%.

Operating margin, again excluding the charge, fell to 35.4% from 40.1%.

The mechanism matters for anyone building or financing capacity.

Meta named third-party cloud spending and third-party AI token costs as expense drivers alongside its own depreciation and data center costs.

Owning more capacity hasn’t ended Meta’s need to rent it while new workloads scale.

The labor swap is real, but it is small next to the infrastructure bill.

Capex Crossed Operating Cash

Q3 capex, including finance-lease principal, reached $34.9B against $32.8B of operating cash flow. That is 106.5% of the cash the business produced, and 54% of revenue. Free cash flow came in at negative $2.1B.

The year is not yet negative. Free cash flow through nine months was $11.0B, but that is down 62.6% from $29.5B in the same period of 2025. The bigger pressure is ahead.

Meta’s capex floor has risen in every quarter of 2026, from $115B to $125B, then $130B, now $135B, while the $145B ceiling has held since Q1. Hitting the new range means Q4 capex of $49.2B to $59.2B, which is 41% to 70% above Q3.

Free cash flow also overstates what is left. Meta paid $4.1B in dividends through September and $12.8B in taxes on employee share settlements, and both sit outside its free cash flow measure.

On the balance sheet, cash and securities of $82.4B now exceed long-term debt by only about $1.5B, after debt rose $22.2B since December.

Meta had already borrowed heavily this year. The equity raise is what came next.

Why Equity, and What It Costs

The financing has three parts. A $12.5B mandatory convertible preferred pays a 6.50% dividend, about $812.5M a year, and converts around August 2029.

A $25B common stock offering, upsized from $22.5B, sold 40 million shares at $625 through forward contracts.

Meta expects about $24.9B when those forwards settle, on or before November 19, 2027, so the cash is not on hand today. A further $60B Class C at-the-market program is planned but not yet set up.

Together that is $97.5B of potential funding, not money raised.

Class C stock is the structural piece. Meta plans a one-for-one Class C dividend to existing holders around January 19, 2027, which leaves voting power unchanged.

After that, Meta can issue non-voting shares to raise money, pay employees, or buy companies without diluting founder control.

The economic dilution still falls on holders. Equity avoids more leverage on a balance sheet where debt now nearly matches cash, but it shifts the cost onto per-share returns.

One detail shows how tight the loop has become. NVIDIA, AMD, and Broadcom indicated interest in buying up to $5B of the common offering.

Those indications are nonbinding. Even so, the companies Meta pays for chips and networking offered to help fund the purchases.

Meta Funds AI From a Single Business

This is where Meta’s position differs from its peers. Microsoft, Alphabet, and Amazon each rent capacity to outside customers, so part of their buildout earns cloud revenue as it comes online.

Meta’s buildout earns through advertising, which is 97.6% of revenue, and through new products that do not yet report revenue.

Those products are growing in usage. Muse onboarded more than 10 million users in the US and Canada, more than two million businesses use Business Agent weekly, and Business Agents handle over 30 million weekly conversations.

Management expects incremental revenue to begin in 2027 but gave no target. It also expects Muse compute costs per unit to improve more than tenfold over the next year while total serving costs rise as usage grows. Cheaper units don’t mean a smaller bill.

The obligations stretch far past one year’s capex. Reuters reported about $825B in future spending commitments, largely data centers, AI infrastructure, and cloud capacity.

That figure comes from secondary filing analyses rather than a reconciled filing schedule, and it spans many years.

The same analyses show about $121.4B in contractual payments due in 2027.

What the Buildout Means Outside Meta’s Stock

Demand visibility and shareholder visibility are now separate questions.

Developers, equipment makers, and power providers can benefit from committed construction even while Meta’s equity holders face dilution and uncertain monetization.

For suppliers, the question is contract quality and payment timing, not Meta’s share price.

The location data Meta disclosed is US-only: owned data centers in 23 states and the supercluster program. Emerging markets show up on the other side of the ledger.

Meta’s ad pricing growth was partly offset by impression growth in lower-monetizing regions and surfaces, and management named compute intensity and localization as limits on rolling out agents internationally.

Emerging markets supply users and impressions. They are not where Meta disclosed capacity, and global reach is not a global deployment map.

What Resolves Next

Three signals will show whether the equity raise is a bridge or a permanent cost.

First, Q4 capex against the implied $49.2B to $59.2B, and Q4 cash conversion.

A quarter at the top of that range with flat operating cash flow would make the $60B Class C program look necessary rather than optional.

Second, the first revenue line for Muse or Business Agent in 2027. Management has said incremental contributions begin then.

A disclosed revenue figure confirms the thesis. Usage metrics alone, through another year of rising capex, would not.

Third, the pace of issuance. Watch when Meta settles the common forwards and whether it establishes the Class C program and draws on it.

The speed of that drawdown is the clearest read on how far operating cash falls short of the buildout.

The Bottom Line

The ad engine is strong, and AI is visibly improving it. That wasn’t enough.

Meta cut 15% of its workforce, and costs still grew faster than revenue. Capex passed operating cash, and Meta turned to equity five days after the print.

The cycle has now moved from capital structure to cost structure and back again, this time with shareholders carrying the bridge.

If Meta’s new AI businesses do not show revenue in 2027, that bridge becomes a lasting cost to shareholders.

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

Nebius signed a $3.5 billion loan with Citi that matures in 364 days. It is built to be replaced, and that is the point of it.

The facility gives Nebius cash to keep buying GPUs and building data centers into 2027 without locking in long-term debt terms today.

The company says the facility gives it flexibility over the timing and structure of further capital raising. What must go right is just as clear.

New capacity has to reach paying service fast enough to support longer, cheaper loans before the bridge comes due on October 16, 2027.

So, the headline number matters less than the order in which the money arrives.

What Nebius Signed

Nebius signed the agreement on October 17 and announced it on October 22. The SEC filing calls it a Bridge Facility Agreement. Citibank, N.A. is the mandated lead arranger and lender, and Citi affiliates act as agent and security agent.

There are three borrowers: Nebius Group N.V., the Amsterdam parent, plus Nebius Inc. in Delaware and Nebius B.V. in the Netherlands.

The loan is secured by substantially all of each borrower’s assets. That includes the parent’s shares in the two subsidiaries, subject to excluded assets and permitted liens.

Pricing starts at Term SOFR plus 2.75%. The margin rises by 0.25 percentage points for each 90-day period after signing, up to 3.50%. On the filing’s summary terms, the steps fall around January 15, April 15 and July 14, 2027.

Nebius can prepay without premium or penalty. A minimum liquidity covenant applies. Its threshold is not disclosed.

Proceeds go to data center development, GPU procurement and general corporate purposes, according to the company announcement.

The announcement gives the maturity, the starting rate and the cap. The filing adds the bridge label, the 90-day steps, the parent as a borrower and the reach of the security.

Read together, those terms describe timing capital. A margin that climbs each quarter, paired with free prepayment, rewards early refinancing.

The cost of waiting is easy to size. If the full $3.5 billion were drawn, the margin alone would run about $96 million a year at 2.75% and about $123 million at 3.50%, before SOFR, fees or hedging.

That is a GDCH calculation on an assumption the record does not confirm. It shows why the facility rewards a quick takeout.

Two facts are not public: how much Nebius drew at closing and how much remains available. The $3.5 billion is the size of the facility, not a confirmed cash receipt.

How the Money Stacks Up

Nebius funds its buildout in layers, with customers supplying the first.

In the first half of 2026, operating cash flow was $4.5 billion. Deferred revenue, mostly customer prepayments, contributed $4.4 billion of that, about 98%.

Cash spent on property, equipment and intangible assets was $8.1 billion. Operating cash flow minus that spending was-$3.6 billion, according to the company’s financial statements.

Management says about 70% of Q2 deals included prepayments covering 50% to 60% of the related capex. It expects more than $9 billion of customer prepayments in 2026. Both are management figures, and the second is a forecast.

The capital markets supplied the next layer. Nebius sold about $2.8 billion of shares through its at-the-market program in Q2.

It closed $5.7 billion of convertible notes on August 24: $3.45 billion of 0.50% notes due 2030 and $2.30 billion of 4.50% notes due 2034. Alongside that offering, holders exchanged $800 million of older notes for about 15.8 million shares.

Then came the first asset-level loan. In July, Nebius announced a $775 million senior secured facility that matures October 31, 2030, priced at SOFR plus 2.50%.

Deployed GPU infrastructure and contracted cash flow from an investment-grade customer back it. That customer’s contract had already entered its servicing phase.

Nebius said the loan and the customer’s cash flows together cover more than 100% of the capex for that deployment. The figure combines two cash sources. It does not describe a loan larger than the equipment cost.

The Citi bridge sits between these layers. It is about 4.5 times the size of the July loan, shorter by more than three years, and priced higher. It also lends against the company as a whole, whereas the July loan lends against one deployment.

When a Contract Becomes Collateral

The Citi facility shows where the Bankability Gap sits for an AI cloud. The Bankability Gap is the distance between announced or contracted capacity and capacity that is truly bankable.

Bankable capacity has secured its power, financing, equipment and customers on terms that fit the asset.

On demand, Nebius has closed much of that distance. Management cites more than $40 billion of long-term commitments from investment-grade customers.

The June 30 financial statements report $37.5 billion of remaining performance obligations, with 36% expected to become revenue within 24 months. The two are different measures and should not be added.

What remains is a question of financing form. The July loan was written against capacity already running for a customer. An asset-backed lender takes security over hardware and an assigned stream of payments. Both are easier to value once the hardware is installed and the customer is paying.

So, the bridge funds a specific stretch: the months between a signed contract and a deployment a lender will finance for years. The binding layer is capital, and it binds at that conversion step.

Construction-stage capacity has to become service-stage collateral. Equipment delivery, power and commissioning set the pace. Nebius is building at scale on that front.

In August, it raised its year-end contracted power target to 5 GW and plans to deploy more than 1 GW a year starting in 2027. Contracted power means secured land and power commitments, which come before energized, billable capacity.

The calendar makes this concrete. Nebius’s Meta agreement includes $12 billion of dedicated capacity orders over five years, deployed in tranches starting in early 2027.

In August, management said that capacity was on track for early 2027. The first Meta tranches land inside the bridge’s life.

There is evidence that the sequence works. The Microsoft contract made both parties’ obligations conditional on Nebius confirming it had secured the required capex financing.

The Q2 shareholder letter reports that all capacity tranches under that contract have now been delivered. Contract, financing and delivery have happened in that order once. The July loan showed the final step, an operating deployment turned into secured debt that runs to 2030.

What It Means on Each Side of the Table

For lenders and private credit, the bridge points to a pipeline. If Nebius repeats the July structure, each deployment that reaches service becomes a candidate for a multiyear secured loan.

The terms to watch are advance rates, reserves, the reach of any parent guarantee and how assets are released from the corporate security package.

Broad corporate security can sit awkwardly beside asset-level loans unless the documents let assets move into a separate vehicle. The executed agreement, which is not public, decides that.

For operators and developers, delivery dates now carry financing weight. A site that energizes on time supports the next asset-backed loan. A site that slips lengthens the bridge-funded stretch, at a margin that rises every 90 days.

For policymakers and utilities, a facility of this size shows access to bank capital. It does not, by itself, show that any single campus is funded through energization. Incentives and grid commitments work best when tied to project milestones.

For investors outside the US, the model has a second path. Under the partnership model Nebius introduced in July, partners finance and own the infrastructure and hardware while Nebius supplies software, system design and customers.

The Q2 letter lists sites in Finland, Israel, Iceland, the UK, France, Estonia and Spain. Partner-owned capacity moves the financing question to whoever owns the building and the chips.

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

Edition No. 74

THE WEEK’S THESIS: THE POWER DATE DECIDES

The power date decided what a data center contract was worth this week. The deals that came with their own power moved. The deals waiting on someone else’s grid got cut.

Microsoft will fund new generation for its Indiana campus through NiSource. Google is paying homes and businesses across the PJM grid to free up capacity through Voltus. In Norway, Crusoe keeps adding halls on Polar’s hydro power.

The ones that stalled were waiting on a power line. Nscale’s Loughton site lost its anchor customer when the grid date moved out by years. NEXTDC cut its outlook after work stopped on the line feeding its Melbourne campus. A contract with no power date is a promise with no start date.

The structure behind Nebius’s deal is in The Mechanism below.

MARKET SIGNALS

Amazon commits to Germany. Amazon plans to invest more than €16 billion (~$18 billion) in Germany in 2027. AWS takes €4.8 billion of that to expand its Frankfurt region. The release doesn’t separate new money from projects already announced. Either way, Frankfurt’s grid now has to carry a bigger cloud on a 2027 timetable.

Nvidia stands behind its customers’ buildings. Forbes tallied ~$164.5 billion of data center commitments Nvidia has reportedly guaranteed or taken on, based on The Information’s reporting. The largest piece is reportedly a ~$105 billion lease guarantee for an OpenAI campus in Ohio. The chipmaker’s own quarterly filing confirms a much smaller slice: $3.53 billion of land, power and shell guarantees. Most of the headline number sits outside the filing, so the full size of the backstop is still a reported figure.

OVHcloud has room to sell. OVHcloud has 250MW connected and 140MW in use, CEO Octave Klaba told Reuters. That leaves 110MW, about half, ready for new customers. Its first AI model launches in early 2027. Powered, empty halls are the scarcest thing in Europe right now, and the French provider has them.

Thailand writes the rules first. Thailand is finalizing data center rules while projects worth more than 1 trillion baht (~$30 billion) wait on them, Bangkokbiznews reported. Proposed setbacks run 100 to 200 meters from homes, schools, hospitals and temples, scaled to how much fuel a site keeps. A draft framework is due in the first quarter of 2027. Until then, the pipeline has a price and no permit.

CAPITAL WATCH

Blackstone and its energy fund. Blackstone closed its fifth energy transition fund at $8.7 billion, its hard cap and more than 50% larger than the last one. The fund has already backed three power generation and transmission platforms. It also bought Flow Control Holdings, which makes parts for data center liquid cooling. Private equity is paying for the plant behind the building, the part that sets the power date.

Elice and the Kosdaq. Elice, a Korean GPU cloud provider, fell 32% on its Kosdaq debut on Oct. 20. It raised 201.1 billion won (~$151 million) at 90,500 won a share, the top of its range, then closed at 61,500 won. The retail tranche was reportedly ~400 times oversubscribed. Every GPU cloud heading to market now gets priced against that first day.

DayOne and its bankers. DayOne is weighing whether to delay its US listing or shrink it to $2.5 billion to $3 billion, Bloomberg reported. The original plan reportedly targeted as much as $5 billion at a $20 billion valuation. No decision has been made. A smaller deal keeps the listing alive but lets this month’s market set the price.

Nvidia and Firmus. Nvidia is in talks to invest hundreds of millions of dollars in Firmus, The Information reported. The GPU cloud provider had shelved its ~$5 billion IPO. Bloomberg reported earlier this month that Firmus was exploring a $2 billion to $3 billion private round. The private round’s price will show whether that check sets the value or props it up.

GLOBAL LENS: THIS WEEK BY REGION

Europe: hydro and cooling get booked first

VINCI’s Cobra IS won a €165 million contract for the cooling plant at a large data center campus in the Iberian Peninsula. The job covers 126 dry coolers sized for about 200MW, plus mechanical and electrical work. The client and the country aren’t named. Cobra is bidding for later phases it puts at several billion euros. Cooling now gets bought as a plant of its own, and the phase-one contractor stands first in line for the rest.

Polar plans a second building at its Tørdal site in Norway, with 40MW of liquid-cooled IT for Crusoe. The first building, at 12MW, is already live on hydro power. Crusoe plans to deploy Vera Rubin NVL72 systems in the new hall, due to go live in late 2027. The release doesn’t say that capacity is under contract yet. Norway gives the tenant a power date it can count on, so the open item is the lease.

North America: capacity found outside the queue

Google expanded its Voltus deal to a 1GW virtual power plant across PJM, up from 100MW signed in June. Voltus pays homes and businesses to cut their use at peak times through thermostats, batteries and industrial equipment. Google funds it, and the agreement runs about three years. That’s headroom on the grid that already exists, with no new plant to build. The test is whether PJM keeps counting that flexibility as capacity.

Applied Digital and Soluna signed a definitive agreement for a joint venture to build a 50MW data center in Briscoe County, Texas. Applied takes 51% and brings the capital. Soluna takes 49% and brings land and power assets next to the 150MW Briscoe wind farm. ChronoScale, majority-owned by Applied, signed only a non-binding letter of intent to lease it. Closing waits on a county tax abatement, and the tenant is still a sister company.

Asia-Pacific: old buildings get the GPUs

GreenSquare brought SYD1 online in Western Sydney, a 1995 data center rebuilt for AI in eight months. Sharon AI is the tenant and is deploying ~8,200 Blackwell Ultra GPUs. IT capacity rose 275% inside the original footprint. Programs managed by Partners Group back the developer. The proposed 80MW second stage will show whether reuse works at scale.

Aolani says it signed non-binding letters of intent for up to ~500MW across Tier III sites in Malaysia. It names neither the sites nor the counterparties, and the letters still need definitive agreements. The startup expects up to 580MW committed by 2027, up from ~80MW this year, with capacity arriving from the second half of 2027. It also plans 22,000 Blackwell Ultra GPUs across Malaysia and the Philippines. The letters hold the megawatts, but a signed lease and the GPU financing decide whether anything gets built.

Share

CONSTRAINT OF THE WEEK: THE GRID DATE

The grid date decided more contracts this week than price did. Microsoft confirmed on Oct. 19 that it pulled out of Nscale’s Loughton site in Essex. Power was due around the end of 2026, and the grid now points to the 2030s. Darren Hardman, who runs the company in the UK and Ireland, told ITV its £22 billion UK plan won’t change. So the money stays in Britain and moves to sites that can switch on sooner.

NEXTDC cut its FY27 guidance on Oct. 21 after Jemena suspended parts of a power line upgrade feeding its M3 campus in Melbourne. Net revenue guidance fell to A475–500millionfromA615–640 million. Work won’t restart before February 2027, so the next power increase at M3 slips to FY28. The shares fell about 40% that day, Rask Media reported. One suspended line repriced a listed operator’s whole year.

Developer Curtis Hall wants a gas plant of up to 49.9MW to start its Cardiff data center, the BBC reported. The plant would run from July 2028 to July 2033, when the full 150MW grid connection is expected. Critics say the size sits just under the 50MW line that sends a project to the Welsh government. Cardiff councillors have objected. Bridge power buys five years, and it moves the fight to the planning committee.

The next bottleneck is the courtroom. Two Lowndes County, Alabama residents sued the county commission over its approval of Project Red Clay under the state’s Open Meetings Act. The project carries a $1.15 billion tax break and has asked Alabama Power for 1,500MW. The judge has recused, and the case is pending. A campus can have its power lined up and still sit idle while a court reviews the vote.

THE MECHANISM

One-year money for a multi-year build

Nebius is an AI cloud company listed on Nasdaq and based in Amsterdam. It builds data centers, fills them with GPUs and rents the computing power to developers and companies. On Oct. 22 it signed a $3.5 billion senior secured facility with Citi. The money goes to data center development, GPU purchases and general corporate purposes.

The facility matures in 364 days. It charges Term SOFR, a benchmark rate, plus 2.75% at first, and that margin can rise to 3.50%. It’s secured, and Citi is lead arranger, agent and security agent. Customer prepayments and operating cash flow make up the rest of the funding base, the company says.

So this is bridge money. Nebius says longer-term asset-backed debt can follow, based on more than $40 billion of long-term contracts with investment-grade customers. That debt rests on the contracts. The contracts rest on halls that are powered and running. Citi gets repaid from that next financing, which puts every power date in the build inside a one-year clock.

The release doesn’t say what secures the loan or what moves the margin up. It doesn’t name the customers behind the $40 billion or say when their capacity goes live. The risk sits in the gap between a 364-day maturity and a build that runs longer.

I’m watching whether the longer-term debt closes well before the 364 days run out. I’d press on what happens to the margin if a site slips. The document to watch is Nebius’s third-quarter results filing, where the security package and the step-up terms should appear.

Leave a comment

NOTABLE TRANSACTIONS

NiSource: 20-year Microsoft power agreement for ~1,100MW of new generation

NiSource signed a long-term energy agreement with Microsoft on Oct. 16 for its planned campus in northwest Indiana. Service starts in 2028 and runs for 20 years. Its generation arm will add about 1,100MW to its pool. No contract value was disclosed.

What makes it notable is that the customer pays for the power plant before it draws a megawatt. The tech company fully funds the generation and transmission needed to serve its load. The utility says that keeps the cost off its existing NIPSCO customers.

The utility now expects about $1.75 billion of savings for those customers over the contracts’ lives, up from $1.4 billion. That works out to as much as $140 per customer a year. The generation arm also contracted 600MW of battery storage under a long-term tolling agreement. An affiliate of AGP Group will build it.

What to watch is the regulatory approval. The agreement still needs it, and the conditions attached will show how much risk stays with ratepayers. A clean approval gives other utilities a template to copy.

Alcoa: Massena East smelter site sale for up to $315 million

Alcoa signed definitive agreements on Oct. 15 to sell parts of its idle Massena East smelter site in New York to a data center developer. Total proceeds could reach $315 million. The company doesn’t name the buyer. Datacenter Dynamics reported it is North Country Colocation Services, an NYDIG affiliate.

What makes it notable is that most of the price depends on megawatts that don’t exist yet. The seller gets about $50 million to $65 million at closing, set by how much data center capacity is approved.

Another $60 million to $250 million comes over seven years after closing, tied to approved expansions and commissioned capacity. Most of it is expected in the first five years. Alcoa’s CFO said on the earnings call that 180MW is initially permitted, expandable to 635MW, DCD reported. So the seller shares the permitting risk and gets paid as the buyer’s power comes through.

What to watch is the permitting. Closing needs regulatory approvals and the conditions to start construction, and it’s expected within six months. The approved megawatts decide how much of the $315 million Alcoa actually sees.

CBRE: Onnec acquisition for £460 million plus up to £300 million earnout

CBRE signed a definitive agreement on Oct. 20 to buy Onnec Group, a London-based data center infrastructure services firm. The initial price is £460 million (~$610 million) in cash. Closing is expected around the end of the year, subject to regulatory approvals.

What makes it notable is that a real estate services firm is paying a buildout price for cabling and fiber crews. Onnec designs, deploys and maintains network, fiber and data center infrastructure for hyperscalers in the UK, Europe and Asia-Pacific.

Onnec is expected to bring in about £370 million of revenue and more than £50 million of core EBITDA in 2027. That puts the upfront price at no more than about nine times 2027 core EBITDA. The earnout adds up to £300 million if Onnec hits growth thresholds. CBRE expects its critical infrastructure services line to bring in nearly $4.5 billion of revenue in 2027.

What to watch is the earnout. Its terms haven’t been published. They’ll show how much of the price rides on hyperscaler build spending staying high for several more years.

THE FRAMEWORK: WHEN THE POWER ARRIVES

Every deal in this edition turns on the date its power shows up. The table sets each one against its power plan and where it stands.

THE QUESTION FOR NEXT WEEK

What does NiSource disclose on Oct. 28? The utility promised more operational and financial detail on the deal at that day’s third-quarter call. If it puts a cost on the new generation, every utility courting a hyperscaler gets a price to copy.

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

Join Global Data Center Hub

Share 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.

This article is the 19th article in the series: Ground to Grid: A Free 21-Lesson Guide to Mastering Data Center Development

Join Global Data Center Hub

Why Operations Is Where the Return Lives

A data center earns its return over years of reliable operation.

Completion only starts the clock. Development gets the asset built. Operations is what makes it pay.

Operations is the long phase that follows handover. It begins at acceptance, when the operating team takes responsibility from the builder.

ISO/IEC TS 22237-7:2018 covers that transition and the management processes that follow it. From that day, the asset has to earn for as long as it stays in service.

The tenant pays for one thing above all: power, cooling and connectivity that stay on, within the limits the contract promises.

Reliability is the product. The building is how it gets delivered.

Maintenance and lifecycle discipline decide how much of the return survives over time.

The owner keeps the cash left after service promises are met, maintenance is done and worn equipment is replaced. A low operating bill can hide a backlog that shows up years later as a failure or a bill.

The Pillars of Operation

1. Reliability and Uptime Management

An SLA (service level agreement) sets the performance the tenant pays for and the remedy when the operator misses it. Meeting it is the core operating product, and it is monitored every minute of every day.

Power is where most of the risk sits. Uptime Institute’s 2026 outage analysis names power as the leading cause of impactful outages, with UPS systems, transfer switches and generators prominent among the failures.

Uptime can also hide risk. A facility can stay online while part of its protection is gone. A UPS (uninterruptible power supply, the battery system that carries load through a power break) may be bypassed for repair.

The tenant sees no outage, yet the site is one fault away from one.

So, the strongest operating reports track degraded states next to uptime: how long each lasted, which loads were exposed and what protection remained. They answer a simple question. What happens if the next fault arrives now?

2. Preventive and Predictive Maintenance

Maintenance keeps critical systems working before they fail. It is the discipline that prevents the downtime that breaches SLAs.

Preventive maintenance runs on a schedule, by time, use or a set requirement. Condition-based work starts when measurements show wear. Predictive maintenance uses data to forecast a failure before it happens. The right mix depends on what a failure would cost and how fast the asset can be restored.

ISO 55001:2024 sets requirements for an asset management system. It leaves the financial and technical rules for each asset type to the owner. Each asset still needs its own plan.

Predictive tools are arriving, but trust is uneven. In Uptime Institute’s 2026 global survey, 65% of respondents would trust AI for predictive maintenance. Only 16% would trust it to change mechanical and electrical settings on its own. Those figures measure trust. Savings still need proof from a real deployment.

3. Monitoring and Response

Monitoring watches power, cooling and network systems around the clock. Response turns an alarm into action before a fault becomes a failure.

Both depend on the control systems. NIST’s SP 800-82 Rev. 4, released as an initial public draft on September 21, 2026, includes building automation, access control and physical monitoring in its scope.

It is still a draft. Facility controls are operating infrastructure, with their own security and recovery needs.

Fast response also needs parts and people. A spare in a supplier’s network is of little use if it cannot reach the site inside the restoration window.

And in the same survey, 53% of 607 respondents reported trouble finding qualified candidates for open roles.

4. Lifecycle and Capital Planning

Every major asset has four lives: accounting, physical, supported and economic. The supported life ends when parts and vendor help run out. The economic life ends when running it stops making commercial sense. A chiller can work perfectly and still be commercially out of date.

Lifecycle planning replaces and upgrades equipment across those lives, so the asset keeps performing for decades. The plan is built asset by asset, with a clear split between funded and unfunded needs. An approved future budget is different from cash on hand when the switchgear fails.

Liquid cooling brings these decisions forward. On August 27, 2026, Singapore launched SS 726:2026, Liquid Cooling in Tropical Data Centres, which covers operation and integration with existing facilities.

In Latin America, ODATA said on September 17 that installation and first-phase testing of its DeltaFlow system were underway at SP04 in Brazil and QR03 in Mexico. Savings results are not yet public.

Why Operations Determines Asset Value

Reliable uptime keeps tenants. A tenant whose service has stayed within its SLA has a reason to renew and to expand in the same building. Each renewal avoids the empty months of finding a replacement.

Downtime works in reverse. A breach triggers the SLA remedy, usually service credits. It also strains the relationship, and the tenant remembers it at renewal.

Efficient operation protects margins, but the contract decides whose margin. PUE (power usage effectiveness) divides a facility’s total energy by the energy its IT equipment uses. Uptime’s survey puts the 2026 average at 1.52, across 644 responses.

Google reports 1.09 for its fleet in 2025, a company figure that does not transfer to another building. When energy passes through to the tenant at actual cost, a lower PUE cuts the tenant’s bill and can leave the owner’s margin flat.

When energy sits inside a fixed price, the operator keeps some of the gain. Under a contractual PUE multiplier, the billing formula decides.

PUE also says nothing about useful computing output, which is why Singapore’s Green Data Centre Roadmap tracks facility and IT efficiency separately.

Lifecycle discipline extends the earning life, and it shows up in the cash. Digital Realty’s 2026 outlook, published July 23, includes $400 million to $425 million for recurring capital spending plus capitalized leasing costs.

That is forward guidance. Its definition includes second-generation tenant improvements and leasing commissions, so the line covers the cost of keeping existing revenue.

The same disclosure reports a $94 million insurance settlement, net of income tax, with about $27 million recognized in Core FFO (funds from operations, a REIT earnings measure). A careful reader separates recoveries like that from recurring performance.

No source reviewed for this piece sets a standard maintenance-cost ratio, retrofit return or valuation premium for operating quality. Each asset’s own records have to carry that case.

Where Operations Fails

Deferred maintenance is the most common failure, because it looks like savings.

A documented deferral with a reason, an approval and a date is a managed decision. A task deferred again and again becomes unplanned downtime.

Monitoring and configuration gaps let small faults become large ones. On March 26, 2024, the PDX01 data center Cloudflare uses in Oregon lost power. Cloudflare’s April 8 postmortem says the facility’s initial assessment pointed to incorrectly set breaker coordination settings.

Breaker coordination makes the breaker closest to a fault trip first, so the rest of the system stays up.

Cloudflare calls this a preliminary finding and one contributing factor. Its own APIs and dashboards recovered about seven minutes after the power loss, without human intervention.

The equipment was installed. The settings decided how it behaved.

A missing capital plan leaves aging equipment to fail in service. The worst timing is a major replacement that falls due before the tenant renews. The owner then pays ahead of the revenue meant to fund it.

The last failure starts before operations does. An operating team that inherits a facility without current drawings, approved settings, an asset register and trained staff is running a building it was never prepared to run.

Open construction items with no owner drift into the operating budget.

Strategic Implications

For developers: Involve operators early in design reviews, when changes are still affordable. Before completion, document all open items with assigned owners and fix dates, and hand over current drawings, approved protection settings, and configuration backups.

For investors: The operating record is where the return is proven. Start with uptime and tenant retention. Then ask what sits behind them: the degraded-state log, the maintenance backlog and the capital plan.

Before you credit an efficiency project to the owner, read the energy clause. Under a pass-through, the tenant keeps the saving.

For operators: Reliability is the product. Report degraded states next to uptime every month. Give every deferral a reason, an approval and a completion date. Then test the response itself.

Check that critical spares can reach the site inside the restoration window, and that the site can handle a fault when its most experienced engineer is away.

For policymakers: Reliable operation is what makes the sector dependable infrastructure. Say clearly which rules bind, to whom and from when.

The European Commission’s September 21, 2026 proposal for a data centre rating scheme still faces a two-month scrutiny period. Until it clears, no one can price the obligation with confidence.

Key Takeaway

The build is a cost. The operation is the return.

A data center that is developed brilliantly and operated poorly is a bad investment.

The tenant pays for the years of service that follow handover. Uptime, sustained over those years, is what actually pays.

So, when you evaluate a data center, ask how it has run, and how it will keep running, before you ask how it was built.

The maintenance backlog, the degraded-state log and the capital plan will tell you whether the return in the model is still there.

Do you think operational excellence or development skill is more often the deciding factor in a data center’s returns, and why?

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

When a storm knocks out the power at home, the lights go dark and the fridge goes quiet. When the same storm hits a data center, the email still loads and the AI chatbot still answers.

That steady service rests on something most people never see rows of diesel engines parked beside the building, waiting for the grid to fail.

In 2024, Amazon bought about 350 acres near Xcel Energy’s Sherco coal plant outside Becker, Minnesota, according to MPR News.

The plan called for a large data center and 250 diesel backup generators to keep it running through any outage.

Amazon asked the Minnesota Public Utilities Commission, the state board that oversees power plants, to waive a “certificate of need.”

State law requires that ruling before anyone builds a plant of 50 megawatts or more.

Amazon said its generators would never connect to the grid and would run fewer than 15 hours a year, Data Center Knowledge reported on March 3, 2025.

The commission disagreed. On February 28, 2025, it voted 4-0 to deny the exemption.

E&E News reported on March 4, 2025, that Amazon would have to prove the need for up to 600 megawatts of diesel generators. Data Center Frontier noted on March 7, 2025, that the total came close to the 647 megawatts of Xcel’s Monticello nuclear plant.

Then the project stopped. On May 23, 2025, MPR News reported that Amazon had suspended development in Becker while keeping the land. Amazon gave its own reasons.

A spokesperson cited uncertainty and said the company would move resources to projects that could add capacity faster. Whatever the cause, a campus the town had expected sat paused, with no new date.

Becker is one example of a pattern that grows with every new campus. A data center has to stay on when the grid fails, so each building gets its own backup. Bigger buildings need more engines. Counted one at a time, each engine looks small.

Counted together, they reach the size of a power plant, and the rules written for power plants start to apply.

Thirty Seconds of Batteries, Then the Engines Take Over

When the grid fails, the switch to backup happens in two steps. First, large banks of batteries take the load at once.

Microsoft explained in a June 1, 2021, post that these batteries cover roughly the first 30 seconds, the time a diesel generator needs to start.

Then the generators take over and run for as long as the outage lasts, burning fuel stored on site.

Most days, the engines sit idle. They still have to run on a schedule, because an engine that never starts in practice may fail on the day it is needed. So, crews start them for testing and upkeep.

The US Environmental Protection Agency’s rule for emergency engines, as published in October 2026, allows up to 100 hours a year of testing and maintenance, with no time limit during a real emergency.

Each start sends out diesel exhaust and engine noise. That is why the engines need air permits, the state approvals that set how much pollution a machine may release and for how many hours.

The counting question shows up outside the United States too. In England, a site whose engines together burn fuel at a rate of 50 megawatts or more needs an environmental permit from the Environment Agency, the national pollution regulator.

Amazon’s December 2024 permit application for its data center in Hayes, west London, covered 28 generators, according to the agency’s public consultation file.

A Backup Generator Sounds Small Until You Count Them

Most people picture a backup generator as one machine, like the one behind a hospital or in a neighbor’s garage. That picture is reasonable.

For most buildings, it is accurate. An office or a clinic might keep a single engine for emergencies. The word “backup” also suggests something that rarely runs, and data center engines do sit quiet most of the year.

Amazon estimated its North Carolina engines would run about 5 to 15 hours a year for testing and upkeep, WRAL reported on July 30, 2026.

The count changes the picture. Amazon’s planned campus in Richmond County, North Carolina, will have 588 emergency backup engines. Duke Energy, the local utility, will add 57 temporary engines to power the site until its grid connection is ready.

North Carolina’s Department of Environmental Quality approved both sets of air permits on August 28, 2026. That is 645 engines for one campus.

In Virginia, the state’s environmental agency listed 177 air permits for data centers as of November 12, 2024, the Virginia Mercury reported.

Backup power at a modern data center is a fleet of engines. The fleet can be large enough that a state treats it as a power plant.

A Cleaner Backup Is a Ten-Year Project

Replacing diesel takes years, and the companies that want to do it have said so. In July 2020, Microsoft, the company behind Windows and the Azure cloud, set a goal to end its dependence on diesel fuel for backup power by 2030. That is a ten-year plan.

In 2020, Microsoft ran a row of servers for 48 hours straight on hydrogen fuel cells, machines that make electricity from hydrogen without burning it. The company called that test a proof of concept.

The grid adds its own delay. Duke’s 57 temporary engines in North Carolina exist because the permanent grid connection will arrive later, and the state requires Duke to retire them within one year of starting them up.

When a new campus waits on the grid, the engines carry more of the load for longer.

GDCH has covered why new data centers wait years for power.

A strained grid can also pull the engines into a bigger role. In 2023, Virginia’s Department of Environmental Quality proposed letting data centers in Loudoun County run their generators during grid emergencies without the usual hour limits, to free up power for homes.

It dropped the plan on April 12, 2023, after public comment, saying the issues were being worked out among data centers, utilities, and PJM, the grid operator for the region, according to Data Center Frontier.

Every one of these paths runs through public review. In North Carolina, the Amazon and Duke permits went through a public hearing on July 30, 2026, before approval four weeks later.

In Minnesota, the denied exemption meant a full certificate of need case, in which Amazon would have to show the engines were needed and weigh cheaper alternatives.

The Engines Sit Next Door to Someone

If you live near a planned data center, the generators may be the part you hear about last. WRAL reported on July 30, 2026, that some Richmond County residents had expected a distribution center and learned the scale of the project late.

A local group, the Richmond County Coalition for Justice and Black Empowerment, told WRAL it wanted residents included in decisions.

If you follow the AI story, the engines explain a trade built into every outage plan.

Amazon told WRAL that each building needs backup to keep cloud services running for hospitals, first responders, utilities, and national security agencies. Reliability for users far away rests on engines in one town.

If you pay a power bill in a crowded grid region, the same engines can enter the picture during a shortage.

The 2023 Virginia proposal was written to free grid power for homes by moving data centers onto their own diesel.

Four Places the Generator Count Shows Up

Start with the air permit notice. State environmental agencies post draft permits for public comment, and those drafts list the number of engines.

North Carolina’s notice for Amazon’s Richmond County campus listed 588. A large engine count signals a large campus.

Next, read the hours in the permit. The draft North Carolina permits allowed each backup engine up to 500 hours a year, while Amazon expected about 5 to 15.

The permit sets what the engines are allowed to do during a long outage. The estimate describes a normal year. Both numbers sit in the same public file.

Then look for temporary engines. When a utility brings in “bridge” generators, as Duke Energy did in Richmond County, the site is ready before its grid connection is.

That shows up in utility filings and air permits, and it tells you the grid side of the build is running behind the building.

Finally, watch for a power plant review. When a state utility board takes up a data center’s generators, as Minnesota’s commission did in Becker, the state has decided to count the engines together.

England reaches the same result through its 50-megawatt permit rule. Those cases sit on public dockets; with filings anyone can read.

The Diesel Under the Cloud

The cloud feels weightless. It lives in buildings that can keep a power plant’s worth of engines on standby, and every AI answer that survives a storm survives it on stored fuel.

As campuses grow, the engine counts grow with them.

Who decides how many engines one town should host, and what replaces them once diesel goes?

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

Jensen Huang has a habit of calling the next phase of AI infrastructure before the market prices it in.

In a recent long-form interview on an investor podcast, he laid out where the build-out stands heading into 2027: what is scarce, what is getting more expensive, and who is actually paying for it.

The conversation spent little time on chatbots or model launches.

It went to the physical and financial layers: the sites, the power, the cost per gigawatt, and the money that turns chips into running capacity.

The message was clear: the chips are no longer the hard part. Power, sites, and financing are.

One caveat governs everything that follows: Huang runs the company that sells most of the world’s AI chips, and the hosts are Nvidia shareholders. Every claim about demand, market share, and durability comes from people with a direct stake in the answer.

Most of his argument holds up. Three parts do not, or at least not yet.

Here are the ten takeaways where he is right, and the three where investors should push back.

The Bottleneck Has Moved

1. Chips are no longer the binding constraint.

Huang named land, power, and shell (buildings ready to receive equipment) as one of the most important obstacles to next year’s build-out. The host relayed a street estimate of roughly 50 GW of announced chips against 35 to 40 GW of sites that may be ready. The exact gap is an estimate, but the direction matches what the market is showing: chips can be made faster than sites can be energized. The scarce asset is now the ready-to-power site, not the GPU.

2. Power scarcity rewards the most efficient chip.

The common view is that a power crunch hurts chipmakers by capping how many chips can be installed. Huang argues the opposite. When every watt is limited, buyers who still need to hit revenue targets choose the hardware that produces the most output per watt. He says this is why Nvidia’s share is rising even at its current scale. The logic is sound: when one input is scarce, whatever uses it best becomes more valuable.

3. Supply will follow ready sites, not borders.

Because demand exceeds supply, Huang said Nvidia can redirect chips to whichever customers and regions have sites ready. He cited early build-outs in Africa, Central Asia, Southeast Asia, Latin America, and the Middle East, and said he does not like that the U.S. may not keep pace. For markets that can deliver power and permits quickly, this is a direct opening: the world’s largest supplier is telling buyers it will go wherever the megawatts are.

The Cost Curve Has Turned

4. The end of Moore’s law makes spending rise on its own.

For decades, twice the computing power cost the same a year later. Huang says that era is over: twice the transistors now cost twice as much, however long you wait. That means capex climbs even if demand for compute stays flat. He put pre-AI computing spend at about $0.5 trillion a year and sees compute alone reaching about $1.5 trillion by 2030, inside $3 to 4 trillion of total annual infrastructure spending. Most AI spending debates leave this effect out.

5. Each gigawatt now carries more than twice the capital.

Huang said Nvidia’s content per gigawatt rose from about $18 billion to more than $40 billion as the company added networking, CPUs, storage, and security to what it sells. He expects the number to keep rising. He also contrasted roughly $5 billion per gigawatt for general-purpose computing with about $50 billion for the newest systems. For anyone sizing debt or equity per megawatt, older cost assumptions understate the capital each campus now carries.

The Demand Base Is Wider Than It Looks

6. Two AI labs are not carrying the build-out.

The bear case asks whether two frontier labs can earn enough to justify trillions in spending. Huang split Nvidia’s demand into rough thirds: everyday computing such as data processing and simulation, five or six AI labs rather than two, and thousands of startups renting compute through neoclouds. He said about 80 percent of those startups use open models to build new products rather than compete with the labs. The exact split is his, but the broader point is right: demand is more spread out than the headlines suggest.

7. AI labs will build a copy of the internet.

Huang expects labs to build digital twins of the internet, so agents can learn to use websites and tools without acting on the live web. It is the same approach robotics already uses, where machines train in simulated factories before entering real ones. Few people are modeling this, but it would add a new, compute-heavy layer of demand: simulation for agents, on top of training and inference.

8. Breakthroughs are now search problems.

The host cited a frontier lab solving the Navier-Stokes problem with about 10,000 agents and an estimated $4 millions of compute. Huang’s point was that the result rested on decades of human work, and the AI’s role was to search every path from there. He believes many scientific fields sit at a similar edge. If he is right, science becomes a large and recurring buyer of compute, and the cost of each breakthrough falls as compute gets cheaper.

How Compute Gets Bought and Deployed

9. Orders are placed before the product exists.

Huang said buyers now commit $50 to 100 billion to systems that have not been built. A system at that scale cannot be tested in advance, because testing it would take a giant data center and years of work. Trust in the supplier’s track record replaces testing, the same way Nvidia places all its leading-edge chip production with one foundry it can plan with five years out. This changes how delivery and counterparty risk should be judged on large compute contracts.

10. Safety will be engineered, not aligned.

Huang argued that perfect alignment is impossible, because goals like “be successful” can be read countless ways. His answer mirrors how companies treat new employees: start with zero trust and minimum access, monitor from outside the system so the AI cannot adjust its behavior, and widen limits as trust is earned. This matters for infrastructure because enterprise adoption, the next large wave of demand, depends on agents being safe to deploy inside corporate networks.

Where He Is Not

Three claims deserve more scrutiny than the interview gave them.

1. “Flexible” does not mean “durable.”

Huang’s financing argument runs as a chain: general-purpose chips can be re-rented to any customer, for any model, in any region, which makes them durable, easier to finance, and cheaper to own. The first link holds. The second does not follow. He notes Nvidia still updates software on chips sold eight years ago, but software support is not economic life. Each new generation needs more power per rack and better cooling, so older chips and the buildings around them lose value even while they still work. Lenders should price GPUs on their economic life, not their physical one.

2. The chipmaker as lender is a risk, not just an upside.

Huang presented two new business lines as pure growth: software licensing at a few thousand dollars per GPU per year, and revenue share with neoclouds that Nvidia helps finance, supply, and fill with customers. The second deserves a harder look. When one company is supplier, investor, and source of demand, the risks are linked. If demand softens, the neocloud, its lenders, and Nvidia’s revenue share all weaken at once. Operators taking these terms should know exactly what they give up and how dependent they become.

3. The tenfold jump in value per gigawatt is asserted, not proven.

Huang’s own numbers set the test: about $5 billion per gigawatt for general-purpose computing versus about $50 billion for the newest systems. For that spending to pay off, each gigawatt must produce roughly ten times more economic value. He argues it will, and a wider base of buyers helps. But spreading the bill does not shrink it. Someone still has to earn ten times more per gigawatt, and the interview offered conviction rather than evidence.

Strategic Implications for Digital Infrastructure

Strip out the book-talking and three implications survive for anyone investing in or operating digital infrastructure.

First, the energized megawatt is the asset. If chips can be made faster than sites can be powered, value moves to capacity that has cleared every stage: announced, permitted, financed, and energized. Announced capacity should be discounted until it shows the later stages.

Second, cost models need updating. With Nvidia content per gigawatt more than doubling, per-megawatt capex, debt sizing, and equity checks built on older assumptions understate the capital each campus needs. Models should carry both the higher cost and a shorter economic life for the hardware inside.

Third, emerging markets with ready power gain negotiating strength. The largest supplier says it will send chips wherever sites are ready. Markets that can deliver power and permits on a credible timeline can attract capacity that U.S. sites cannot absorb.

The trade-off is dependence: vendor-financed, single-anchor builds should be priced for that concentration.

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

• Capital markets priced the power constraint for the first time in H1 2026. CoreWeave closed an $8.5 billion investment-grade financing backed by GPUs in April. Investors then repriced $14 billion of Oracle-backed data center debt in June secondary trading.

• Asia-Pacific and Europe absorbed the H1 2026 primary-market commitments. North America’s defining developments ran through the credit stack and the courts. AirTrunk committed $21 billion to a project near Mumbai. SoftBank announced plans for 5 gigawatts of French AI data center capacity. Maine enacted a statewide data center moratorium.

• H2 2026 will test whether a second rated-debt repricing event follows the Oracle precedent. Industry estimates put only 50 to 60 percent of US data center capacity scheduled for the next two years on an on-time trajectory. Delivery is now at the center of the next capital-markets conversation.

In April 2026, a rating agency agreed to call GPU-backed debt investment grade for the first time. In June, the same paper traded at a discount in secondary, with PIMCO selling down.

Those two events reset the central question in global data center investing. Capital markets have priced the power constraint. Can they now price its delivery?

Six months ago, the Q1 2026 Global Data Center Hub piece argued that electric power delivery was the gate on US data center growth.

The first half of 2026 extended that argument globally. Industry research now puts only 50 to 60 percent of US capacity scheduled for the next two years on an on-time trajectory, against a 72 percent historical rate.

The Wall Street Journal confirmed in Q2 that the broader US build is behind schedule.

Equipment lead times for critical electrical components continued to exceed 18 months, per the Q1 synthesis, which is driving the slip.

The scale underneath H1 2026 is worth naming. Global hyperscale demand is growing at a 14 percent compound annual rate, and 18 global markets now hold more than 1 gigawatt of operational capacity, up from 13 in 2024.

US data center power demand is projected to more than double from 31 gigawatts in 2025 to 66 gigawatts by 2027, moving data centers from 4.1 percent of US peak summer power to 8.5 percent.

Big Tech capex is up 61 percent in 2026, with Amazon planning $200 billion of AI spend and Google targeting $180 billion. The broader opportunity sits inside a $3 trillion infrastructure investment supercycle.

Private credit still provides 60 to 75 percent of pre-development capital, and neocloud providers control roughly 10 percent of global third-party inventory. These are the base-rate figures the H1 2026 developments sit on top of.

Three H1 2026 developments earn this synthesis. Capital pricing moved from private credit to a rated-debt layer above it.

Social license hardened from county-level opposition to statewide moratorium and central-government intervention.

Geographic allocation reshuffled, with India promoted to primary-market status.

What follows walks each region through the same three questions: what moved on power delivery, capital structure, and social license.

The regions are evaluated, not surveyed.

Where a question did not activate meaningfully in the first half of 2026, the regional section says so.

North America: Capital Priced the Delivery Gap

North America was the H1 2026 region where the capital-markets story landed hardest, with rated debt forming above private credit at the same time the delivery audit went public and Maine moved permitting to statewide policy.

The US power pathway moved from forecast to audit in H1 2026. The 50 to 60 percent on-time rate, measured against a 72 percent historical baseline, is the base case now rather than a tail risk. Utilities continue to estimate that delivering power will take 18 to 24 months longer than hyperscalers and developers expect.

The capital story defined the first half of 2026. CoreWeave closed the first investment-grade financing backed by GPU collateral, at $8.5 billion in April. Hut 8 followed with $3.25 billion of investment-grade senior secured notes for its River Bend project.

Applied Digital surpassed one gigawatt of contracted capacity and raised a proposed $1.59 billion senior secured notes offering plus a $550 million revolver.

Amazon secured a $17.5 billion loan for AI data center buildout, the largest hyperscaler debt raise of H1 2026 and a signal that even the strongest balance sheets now prefer dedicated infrastructure debt over corporate capex.

Bloomberg reported that Meta and Microsoft led an $850 billion data center leasing boom. At that scale, leasing shifts balance-sheet risk from hyperscaler credit to the developers and lenders financing the leased capacity.

The common structure: secure an investment-grade tenant on a long contract, then raise rated debt against the contracted cash flows. Investors repriced more than $14 billion of Oracle-backed data center debt in secondary trading in June, with PIMCO selling down. A secondary market now exists, and it can move against the issuer.

Social license hardened faster in H1 2026 than in any prior six-month period. Maine enacted a statewide data center moratorium. Developers withdrew from projects in New Jersey and Ontario, per Q2 reporting.

The Q1 synthesis counted $64 billion in US data center projects blocked, delayed, or cancelled since 2023. The first half of 2026 moved the response to the state and provincial level, which changes underwriting from a local-permitting question to a jurisdictional one.

Projects advancing through 2026 and 2027 now have to account for policy risk a county council cannot resolve, and that risk is harder to price because the policies are new.

Europe: Capital Routed Around FLAP-D

European H1 2026 saw growth routed around FLAP-D saturation into France, the Nordics, and Spain’s permitting pathway.

European power constraints concentrated in the FLAP-D markets through H1 2026. Vacancy stayed tight and new hyperscale campus starts shifted north and east to reach deliverable capacity.

Zaragoza’s streamlined single-pathway permitting framework, flagged in the Q1 synthesis as the clearest execution path on the continent, drew interest through the first half of 2026 as the counterpoint to FLAP-D saturation.

The Leeds repurposing of former power station sites extended the Q1 theme that recovered industrial land is the fastest route to campus-scale European development.

SoftBank announced plans for 5 gigawatts of AI data center capacity in France in May, naming the capacity in power terms rather than square footage. Mistral AI advanced the sovereign AI theme with an estimated $830 million Paris data center commitment.

Prime Data Centers announced a 550-megawatt campus in Esbjerg, Denmark, and Telis Energie announced a 500-megawatt campus in Lower Saxony, Germany, both in Q1.

The rated-debt mechanics that defined the US capital story did not materially reshape European financing, which stayed closer to the sovereign and strategic capital shape the Q1 synthesis identified.

The United Kingdom government intervened on a $500 million London data center project in H1 2026, moving European social-license risk from local friction to central-government involvement.

EU sustainability mandates, including the Energy Performance of Buildings Directive and transition-plan disclosure, kept the compliance load heavier than in any other region covered.

Across the first half of 2026, Europe routed growth through permitting-cleared geographies rather than fighting the constraint where it bites hardest.

Asia-Pacific: Capital Followed Power to India

Asia-Pacific delivered H1 2026’s largest primary-market commitments and reshuffled the regional league table, with India promoted from diversification destination to tier-one in a single six-month window.

Asia-Pacific’s H1 2026 story is the clearest regional case of the power pathway activating in the market’s favor. The Singapore-Johor-Riau triangle, named in the Q1 synthesis as the strategic node connecting Southeast Asia, kept absorbing Singapore’s overflow.

Thailand and Indonesia moved from secondary mentions to active hyperscaler commitments. Mainland China’s data center demand continued at a mid-teens compound growth rate through the first two quarters, per the Q2 synthesis. Hong Kong and Shanghai remained under government power allocation policies, which kept vacancy elevated.

Microsoft committed over $1 billion to Thailand data center infrastructure and $5.5 billion to Singapore in Q1. AirTrunk, backed by Blackstone and a Canadian pension, committed $21 billion to a data center project near Mumbai in June.

Meta partnered with Reliance on an AI-enabled data center in India in Q2. Amazon and Anthropic expanded their collaboration to include up to 5 gigawatts of new compute alongside an additional $5 billion investment, with capacity landing across multiple regions.

Developers withdrew from a data center project in Australia in H1 2026, per Q2 reporting. South Korea’s regulatory workaround involves developers building sub-10-megawatt facilities in Greater Seoul to bypass the local utility’s mandatory assessment for larger loads.

Southeast Asian regulators continued to use incentives selectively rather than impose moratoriums. Social license stayed closer to a facilitator than a brake across most Asia-Pacific markets evaluated here.

Middle East and North Africa: Gulf Capital Followed Power Abroad

The H1 2026 MENA story is where the region’s capital landed: in Europe and Asia-Pacific more than at home.

Gulf states hold two assets most of the world does not: abundant natural gas and policy infrastructure that can coordinate generation with digital infrastructure buildouts.

The Q2 synthesis flagged gas with carbon capture and storage as a specific opportunity set, with potential US capital expenditures of up to $80 billion attached to the broader technology. Gulf-scale commitments in H1 2026 remained in the announcement phase.

Nexus Core Systems advanced its Morocco data center project through the first half of 2026.

The more notable H1 2026 story is where Middle East sovereign capital landed outside the region: it continued to appear in European and Asia-Pacific stacks, including as a co-investor base in the AirTrunk Mumbai transaction.

Gulf capital priced AI infrastructure globally rather than regionally in the first half of 2026.

H1 2026 produced no material social-license events in MENA. Where power and land coordinate at the state level, the North America and Europe pattern does not translate.

Africa: Capital Arrived, Power Did Not

H1 2026 produced no African capital-markets event at the scale Asia-Pacific absorbed, but it produced one of the most informative delivery datapoints of the first half of 2026.

African activity in H1 2026 was light. South Africa, Kenya, and Nigeria remained the three markets with the most defined trajectories. Kenya’s geothermal advantage remained the specific asset for inference-capable workloads.

A hyperscaler-backed project in Kenya stalled on energy availability in the Q2 window, per the Q2 synthesis. That is the single most informative H1 2026 datapoint on African data center delivery. Capital reached the market and the power did not.

The Q2 synthesis puts Nairobi and Lagos among the highest-loss grid markets globally at up to 24 percent transmission loss, against Singapore’s 0.2 percent, which continues to cap the realistic scale of latency-sensitive inference workloads on the continent.

H1 2026 produced no material social-license events in Africa. The constraint remains structural.

Latin America: Capital Followed Established Grids

Latin American H1 2026 was when the global platform-consolidation trend extended into the region at modest scale.

The Latin American power pathway continued to run through Brazil’s largely renewable grid, Chile’s established solar sector, and Mexico’s progression toward private investment in strategic generation. Through the first half of 2026, Mexico’s opening remained more invitation than installation.

I Squared’s acquisition of Brazil’s Elea in Q2 extended the platform consolidation trend into Latin American infrastructure, matching the global pattern the Q2 synthesis names as capital shifting from single-asset acquisition to platform ownership.

Secondary expansion into Colombia, Peru, and Argentina continued through the first half of 2026, with Bogotá maintaining its first-mover window. The scale of specific H1 2026 capital events in Latin America stayed below the thresholds set in North America, Europe, and Asia-Pacific.

H1 2026 produced no material Latin American social-license events at the scale of Maine or London. Political and energy stability continued to shape where capital was willing to commit, which is a standing Q1 constraint rather than an H1 2026 event.

What the Regional Comparison Reveals

Three cross-regional observations earn mention from the H1 2026 record.

First, the India risk premium compressed in a single six-month window. The AirTrunk $21 billion Mumbai commitment, backed by Blackstone and a Canadian pension, is the clearest signal that the global league table has reshuffled.

Allocators sized the Mumbai transaction at risk-adjusted returns that six months ago would have required a tier-two premium. It priced closer to tier-one.

The Q2 synthesis names the shift directly: India and Malaysia joined Japan, Australia, and mainland China as leading APAC markets. Allocators still sized to the old premium now face either a repricing event on their existing book or a competitive disadvantage on new allocation. The H2 2026 pipeline will reveal which.

Second, social license has become jurisdiction-specific in ways that will matter for underwriting. Maine’s statewide ban and the UK’s central-government intervention on a London project are two different precedents. Neither looks like a one-off.

Permitting and social-license risk now sit alongside interconnection risk in site selection rather than following it. A project that clears capital, power, and site control still has to clear jurisdictional policy, and H1 2026 showed that policy is now set at the state or national level, not the county level.

Third, the regions where capital priced the first half of 2026 most aggressively are the regions where delivery slippage got measured first. North America absorbed the rated-debt events, the US delivery audit, and the Maine moratorium in one six-month window.

Asia-Pacific absorbed the AirTrunk Mumbai commitment in a period of fewer visible frictions.

The constraint shows up together where the pressure is highest. It should be priced as one risk with three indicators, not three separate risks.

H2 2026 Watchlist

Four items earn tracking between July and December. Each carries a decision rule attached.

First, whether a second rated debt repricing event follows the Oracle precedent. If a second secondary trade reprices at a discount in H2 2026, the asset class is systemically liquid rather than one-issuer-liquid, and allocators sized to held-to-maturity exposure face a repricing event on their existing book.

If no second event occurs, the Oracle trade stays a one-off and the rated-debt stack holds its primary-market pricing into 2027 issuance.

Second, whether India’s Mumbai pipeline signs power contracts commensurate with the announced capital. If energization matches commitments, India’s tier-one promotion is confirmed. If power lags capital, H1 2026 India starts to look like the US pipeline: capital racing ahead of grid.

Third, whether a second US state follows Maine. If one does, jurisdictional-policy risk becomes a national underwriting input and site selection has to add a political-sort layer. If Maine remains alone, the precedent stays local.

Fourth, whether the H2 2026 hyperscaler earnings cycle confirms or revises 2026 capex pacing. If the 61 percent year-over-year increase in Big Tech 2026 capex from the Q2 synthesis holds, the baseline for 2027 rated-debt issuance stays intact.

If pacing softens, H1 2026 rated-debt structures underwritten against that trajectory face refinancing risk sooner than currently priced.

Q1 2026 was when power became the constraint. H1 2026 was when capital priced it and started trading it. H2 2026 will test whether that pricing holds.

Join Global Data Center Hub

Between $8 and $12 trillion is expected to go into data centers worldwide by 2030.

The US has about 30 gigawatts of capacity today and plans to add about 30 more, but large US projects can wait four to seven years for power.

Obinna Isiadinso explains what investors check before funding a data center, and why most markets are just getting started: the world averages about 4 megawatts of data center capacity per million people, against about 50 in the US. From his conversation on .

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

Chip makers are now lending their credit to the buildings their chips run in. AMD’s July deal with Core Scientific shows how that works, in filings anyone can read.

Core Scientific signed 15-year leases for 529 MW of data center capacity. AMD rents 377 MW of it directly. An unnamed neocloud rents the other 152 MW, and AMD signed credit support for those leases.

On its earnings call, Core Scientific said AMD will provide “full credit support throughout the full 15-year lease term.”

That matters because a data center is financed on its rent. When a chip maker stands behind the rent, a lender is no longer looking only at the tenant.

And the reported talks at Oracle, Broadcom and SpaceX show the same idea moving through the rest of the chip market.

What Core Scientific Filed

The 8-K sets out two blocks of leases, signed July 27, 2026.

AMD leased 377 MW of critical IT capacity at Pecos and Hunt County, Texas, and Muskogee, Oklahoma.

The neocloud leased 152 MW at Auburn, Alabama, and Dalton Phase 3, Georgia. Every lease runs 15 years, with three five-year options to extend.

AMD also holds a right to lease another 1,925 MW through December 28, 2028, at certain times and under certain conditions. In its prepared remarks, Core Scientific described the full partnership as up to 2.5 GW.

Management said the AMD sites run on triple-net leases and the neocloud sites on a modified-gross lease.

Pecos is the lead site. It is on track for its first megawatts in the first half of 2027.

The Support Is Real. The Fine Print Is Not Public Yet.

The filing lists three things the credit support agreements do.

They protect AMD equipment inside the neocloud’s space. They give AMD the right, but not the obligation, to cure certain neocloud defaults.

And they set AMD’s rights and obligations if certain material defaults occur.

Management’s words go further: full credit support for the full term. Both can be true.

The filing describes the agreement’s parts. The call describes its purpose.

What the record does not show is the trigger, the cap, or how a payment would flow.

The full agreements have not been filed.

So, the careful reading is this. AMD’s credit stands behind the neocloud leases; on terms you cannot yet see in full. It should not be read as AMD paying every dollar of rent the day a payment is missed.

The Warrant Math Matches the Megawatts

AMD also received a warrant to buy up to 30 million Core Scientific shares at $23.47 each. It vests at 12,222 shares for each megawatt of critical IT load in the leases.

About 6.5 million shares vested at signing.

Run the numbers and the design comes into view. 12,222 shares times 529 MW is about 6.47 million shares. That matches what vested on day one.

Thirty million shares at the same rate cover about 2,455 MW. That is almost exactly the 529 MW signed plus the 1,925 MW reserved.

So, AMD’s stake grows as AMD leases more of Core Scientific’s capacity. Every new megawatt it takes adds to the landlord’s rent roll and to AMD’s equity at the same time.

AMD’s own quarterly filing shows this is a pattern, not a one-off. As of June 27, 2026, a month before the Core Scientific deal, AMD had a maximum gross exposure of $4.1 billion from guarantees on commercial partner data center leases, with terms up to 15 years.

The filing says those guarantees “may be issued in exchange for warrants.” AMD records them as a credit derivative and says they were not material to its financial statements.

The Core Scientific filing does not say its warrant was given in exchange for the credit support. But AMD’s filing describes the general structure plainly. A guarantee goes out. A warrant comes back.

A $6 Billion Build Needs Rent a Lender Can Count On

This is where the support earns its weight.

Core Scientific expects the first roughly 530 MW to need about $6 billion of capital, at $11 million to $12 million per megawatt. It plans to fund that with project-level bonds.

Bond buyers on a 15-year building want 15 years of rent they can count on. Core Scientific’s existing $3.3 billion of 7.75% senior secured notes due 2031 show how tightly the two are tied.

Its quarterly report says amortization on those notes is tied to rent starting under the related data center leases. Rent starts, and the debt starts paying down.

I read AMD’s support as a tool for closing the Bankability Gap. That is the distance between capacity that is announced and capacity that can secure power, financing, permits, equipment, customers and returns.

On the 152 MW, AMD’s credit fills much of the financing piece. A lender pricing those bonds can look at AMD’s support as well as the neocloud’s own credit.

What must go right is the build. Rent starts when megawatts are delivered. Pecos has to energize on schedule, and so do the four sites behind it.

In most markets, a finished shell can sit idle for years while it waits in the utility queue for a live connection.

Four Ways to Pay for the Chip

AMD backs the building. The rest of the market is working on the chip itself, in three other ways.

Broadcom backs the rack lease. Its quarterly filing for the period ending August 2, 2026, describes a five-year structure. A financial partner took over agreements to buy AI racks, along with the customer’s leases.

If the customer defaults, Broadcom owes the gap between 85% of the amount still owed and what the racks sell for. Its maximum exposure, once every rack is deployed, is about $29 billion. It had paid nothing as of the filing.

Apollo led the initial $35 billion financing for that platform in June.

The Wall Street Journal reported that Broadcom is now arranging more than $50 billion so OpenAI can buy custom chips. The talks are early. Whether that package uses the same 85% support has not been reported.

Oracle would rent its chips. Per the same report, Oracle is in talks with Apollo and Goldman Sachs on a structure where investors fund a separate company that buys chips and leases them to Oracle. No amount has been reported.

SpaceX would borrow in its own name. The Financial Times reported that SpaceX is seeking about $40 billion, roughly $10 billions of bank loans and $30 billions of investment-grade debt, with Apollo expected to lead.

CNBC reported that the GPUs could serve as collateral. Bloomberg described the talks as early.

None of the three reported deals is signed. Nvidia has a tool of its own. It says it may offer residual value support for up to 25% of a financing opportunity, assessed project by project.

We looked at why the GPU maker backs resale prices rather than rent when it first used this structure.

The Building Outlasts the Chip, and AMD Backed the Building

Here is the structural point.

Broadcom’s support runs five years; the term of its customer’s rack leases. AMD’s support runs 15 years, the term of a building lease, with options that could take the lease to 30.

That difference sets what each vendor is backing. Broadcom’s support is tied to hardware that can be resold, and its exposure shrinks as lease payments come in.

AMD’s support is tied to powered space. Power, cooling and a long lease can host several generations of chips over 15 years.

So, AMD’s support attaches to the most durable asset in the stack. For the landlord and its bond buyers, that is a clean match: long debt against long rent, with a large credit behind it.

The structure also works for AMD. It gets 377 MW of sites for its own accelerators, a path to 1,925 MW more, and a stake in the landlord that grows with every megawatt.

The open question belongs to the deal. Fifteen years is a long time in compute. The public record does not yet show how the leases handle a chip refresh, a change in rack density, or a new tenant if the neocloud’s space ever needs one.

What to Check on the Next Vendor-Backed Lease

If you lend to data centers, read the support agreement before the headline megawatts. Ask what triggers a payment, what caps it, and whether the support steps down over time the way Broadcom’s does.

If you invest in chip makers, count lease guarantees alongside sales. AMD reported $4.1 billion of them before this deal. Watch the next 10-Q for how the Core Scientific support shows up.

If you build data centers, a vendor-backed tenant is a strong tenant. Tie rent start to energization in writing. Agree up front how a chip refresh works inside a 15-year lease.

If you set policy, the useful ask is disclosure. The Federal Reserve Bank of Kansas City found in October that broad sponsor guarantees let some special purpose vehicles in the AI build-out finance about 90 percent of their assets with debt.

Clear reporting of who stands behind which lease is what lets this kind of credit grow with confidence.

In markets where local tenant credit is thin, vendor support could become one-way large AI builds get funded. That is my reading. It is not yet a pattern outside the United States.

The next AMD quarterly filing is the check to run. If the Core Scientific support shows up as a new guarantee, you will see how much data center rent one chip maker now stands behind.

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

The most useful number in Google’s Wyoming power plan is $399 million.

That is the amount Google advanced to its local utility for long-lead equipment, before any of the new generators were built.

That advance tells you how the deal works. The campus pays for the power it needs, and it pays in writing.

The fuel cells AEP announced this week are the newest and most visible piece of that system. The contracts underneath them are the part other markets will copy.

What AEP Announced

On October 14, AEP said it plans to invest approximately $4 billion in a 702 MW fuel cell project near Cheyenne, Wyoming. Google will buy the power under a 20-year power purchase agreement.

The seller is Cheyenne Fuel Cell, LLC. AEP owns and controls about 92% of it. Funds advised by Basalt Infrastructure Partners own nearly 8%. Construction is set to start in 2026, with full operation expected by the end of 2028.

Bloom Energy supplies the fuel cells. They make electricity from natural gas through a chemical reaction rather than by burning it. AEP says the project plans to add carbon capture to make the power near-zero-emissions.

AEP chairman, president and CEO Bill Fehrman framed the project as a way to meet energy needs beyond the company’s traditional service territory. That phrase matters.

This is a utility building a contracted plant for one customer, outside the system its regulated customers pay for.

For scale, the project equals about 5% of AEP’s $78 billion capital plan for 2026 through 2030. By GDCH’s arithmetic, the announced figures work out to roughly $5.7 million per megawatt. AEP has not published a cost breakdown, so that ratio is a simple division, not guidance.

The Fuel Cells Are One Layer of a Larger Plan

The AEP release says the project follows agreements between Google and Black Hills Corp. that took effect on September 30. Those agreements describe the full power setup for the campus.

According to Black Hills’ October 6 filing, Google’s planned data center will be served by a total resource mix of 2.7 GW, including reserve margins. Up to 590 MW comes as grid-connected service from Black Hills.

Black Hills will also build 564 MW of new gas generation at its Cheyenne Prairie site.

And it will manage the output of about 2.1 GW of Wyoming-based, third-party contracted resources through a private microgrid.

Read that 2.7 GW figure carefully. It measures supply with reserves built in. The load the servers will draw has not been disclosed. When a campus is described as “2.7 GW,” the supply plan is doing the talking.

The public record also does not map each plant into the 2.1 GW microgrid portfolio. Google’s site plan application, as reported by Cowboy State Daily, describes Black Hills resources, fuel cells, and a third-party power plant feeding a substation that serves the campus.

So, the 702 MW of fuel cells sits inside a wider system. How much of the 2.1 GW it fills has not been disclosed.

Then there is timing. Black Hills expects the campus to start taking service in late 2027 and to ramp to peak load in 2030.

AEP targets full fuel cell operation by the end of 2028. The power will arrive in stages, matched to a load that grows in stages.

Supply plans also tend to run ahead of real demand, because campuses often reserve far more electricity than their servers ever draw.

The Contracts Put the Cost on the Campus

Here is the part I would study first.

Black Hills CEO Linn Evans says Google bears all costs associated with serving the planned data center throughout the life of the project. The company says the deal is structured so costs do not shift to its existing retail customers. The agreements run through 2048.

The protections are specific. Black Hills says the agreements protect returns from inflation and interest rate swings, provide full recovery of generation capital over the contract term, and guard against early termination.

Collateral and financial assurance back Google’s commitments, and a contracted minimum peak load is expected to begin in 2027. Google’s $399 million advance for long-lead equipment is refundable, with Black Hills expecting to reimburse it by June 30, 2027.

Black Hills also published its own economics. It plans to invest $1.8 billion in company-owned generation between 2027 and 2029.

It expects about $150 million of net income from the project in 2030 and about $2.4 billion of unlevered free cash flow through 2048, net of that investment.

Those figures belong to Black Hills. They say nothing about the returns on AEP’s fuel cells.

This is the structural point. On the grid side, a regulated utility is serving a gigawatt-scale customer while ring-fencing everyone else. On the generation side, an unregulated affiliate owns a plant backed by one long contract. In both, the hyperscaler’s credit is what makes the build possible.

If you invest in power, that is the asset you are really looking at. A 20-year contract with Google turns a fleet of modular fuel cells into a stream of contracted cash flow.

Without that contract, the same fleet would be a merchant bet on a single site.

The Pattern Has Precedent

AEP has used this structure before. In June 2025, Ohio regulators approved AEP Ohio’s plan to install on-site fuel cells at data centers run by Amazon Web Services and Cologix, with the two customers covering the full project costs.

AEP described it as the first project under its agreement with Bloom.

That agreement dates to November 2024, when AEP and Bloom set a framework for up to 1 GW of fuel cells.

Cheyenne is the same idea at a much larger size: the customer funds the power, and the utility builds and runs it.

What Still Has to Go Right

The structure is clear. Several of its terms are not yet public.

The offtake conditions. AEP’s second-quarter 2026 Form 10-Q says that in January 2026 an unregulated AEP subsidiary signed an unconditional purchase agreement for a substantial portion of its 1 GW fuel cell option, for approximately $2.65 billion.

The filing describes a 20-year offtake arrangement with a high-investment-grade third-party customer.

It states that the offtake remained subject to specified conditions, with the deadline extended by mutual agreement to December 2026, and that AEP would be entitled to additional consideration if the revised conditions are not met.

The October release names Google but does not say whether this is the same arrangement, or whether those conditions are now satisfied.

The cost bridge. If the January purchase and the Cheyenne project are the same, roughly $1.35 billion separates the equipment agreement from the $4 billion investment.

Site work, gas supply, electrical systems, and financing costs could all sit in that gap. AEP has not published the bridge, and the financing structure, debt, and target returns are also undisclosed.

The carbon outcome. Bloom’s current datasheet lists 308 to 378 kg of CO2 per MWh at stated efficiency on natural gas, with near-zero local NOx and no water use during normal operation.

AEP’s near-zero-emissions goal depends on carbon capture that has been announced as a plan. The capture rate, vendor, cost, and start date have not been published.

The build itself. Bloom’s current Energy Server is rated at 325 kW. At that rating, 702 MW works out to more than 2,000 units, by simple arithmetic.

The actual equipment count, availability guarantees, and replacement schedule are not public.

Each item is a condition the outcome depends on. Each should show up in future filings.

The Lesson Travels Beyond Wyoming

For this campus, the binding constraint is firm power on a schedule the grid alone could not meet.

In the terms of the AI Infrastructure Scarcity Stack, the eight layers that can hold up a project, Cheyenne binds at the power layer. What is clearing it is a set of contracts.

That lesson matters most where grids are slow or utilities are stretched.

The parts that travel are simple to name modular generation that can arrive in blocks, a creditworthy long-term buyer, a local utility that runs the system, and written terms that keep ordinary customers out of the cost.

The parts that do not travel easily are local. Gas supply, permitting rules, and the legal right to combine private plants with regulated service all vary by market.

A 20-year contract is only as useful as the law that lets the structure exist.

For policymakers, the label tells you little. “Bring your own power” in Cheyenne still runs through a regulated utility.

What protects the public is the cost-allocation contract, and that is the document worth asking for.

Without that safeguard, households can end up paying higher bills for campuses that are forecast but never finished.

What to Watch Next

Three filings will tell you whether this template holds.

First, AEP’s next earnings materials, for any confirmation that the Google contract conditions are met, plus the project’s financing and earnings contribution.

Second, Black Hills’ transmission filings in Wyoming, which will show how the grid side scales toward the 2030 peak.

Third, any carbon capture contract for the fuel cells, with a capture rate and a date.

If you are weighing a similar campus anywhere, start where Cheyenne starts. Ask who pays, ask where it is written, and ask what the contract says if the load ramps slower than planned.

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

Edition No. 73

THE WEEK’S THESIS: BRING YOUR OWN POWER

This week, the data centers that moved forward brought their own power plant. The ones that leaned on someone else’s grid, or on a thin energy plan, stalled at the planning vote.

The evidence runs across three countries. AEP is building a power plant for Google in Wyoming. Fermi’s Texas campus runs on its own private grid. The Taylor campus near Austin will burn gas behind the meter. Solaria’s Uruguay park comes with its own solar, and Echelon’s Irish campus taps offshore wind through a substation it shares with the wind farm.

The projects that stalled lacked that answer. The UK turned down a growth zone bid built around a Drax campus and named its energy plan as a gap. In Waterford, gas turbines on site weren’t enough to win consent. Owning the power is now the entry ticket, and the buyer behind it decides who gets financed.

The structure behind AEP’s deal is in The Mechanism below.

MARKET SIGNALS

Fink puts a recession on the table. BlackRock CEO Larry Fink spoke to CNBC on Oct. 14. “If we have a moratorium of some sort in this country, we will be in a recession,” he said. He blamed most of the opposition on misinformation, three weeks before the Nov. 3 midterms. The data center fight has moved from zoning boards to national politics, and the money is now arguing its side in public.

Microsoft prices chips in tokens per megawatt. Microsoft unveiled Maia 300 on Oct. 12 and says it expects 40% more tokens per megawatt than leading GPU-based systems. Production deployments start in 2027. That is the company’s own projection. When power is the scarce input, output per megawatt decides which chip fills the hall.

Digital Realty pays the neighbors. Digital Realty is offering every household in De Soto, Kansas, up to $1,000 a year for 10 years, KSHB reported. The credits cover property taxes or utility bills from a $20 million community fund. The city council holds a hearing on the project’s tax district on Nov. 19. Local consent now has a price per household, and it sits in the developer’s budget.

Malaysia’s approvals pile up. Malaysia approved 27 data center projects worth RM225.97 billion between February 2025 and July 2026, Bernama reported. Johor holds 16 of them and Selangor 10. Each must meet a 30% local content floor. The government is also reviewing how much power the sector actually needs, and that review sets how many more get through.

Humain goes to tender in Riyadh. Humain has issued a construction tender for a 100MW AI data center in Riyadh, MEED reported, with an option to reach 200MW. MEED puts the contract at ~$1 billion. The design calls for direct-to-chip liquid cooling and racks of 240 to 250kW. Bids are due Nov. 8, and they will show what Gulf contractors charge for halls at that density.

CAPITAL WATCH

Solaria and Uruguay. President Yamandú Orsi signed an agreement with Spain’s Solaria on Oct. 14 for a 500MW data center park and 350MW of solar. The government puts the total at US$5.7 billion, with the first 200MW phase due by 2029. Bloomberg reported before the signing that the deal would take the form of a memorandum of understanding, and no tenant is named. So the figure describes the park’s planned size, and no buyer has put cash behind it yet.

Hillhouse and TeraNova. Hillhouse Investment launched TeraNova DC, an India data center platform, the Economic Times reported. It plans to invest ~$12 to $15 billion and build more than 2GW by 2032. No sites or committed capital have been named. The first land deal will show whether the number is a budget or a headline.

Princeton Digital Group and its lenders. PDG is seeking a syndicated loan of up to $1.2 billion for its 198.8MW Johor campus, Bloomberg reported. The three-year facility also refinances existing debt, with HSBC, Maybank, OCBC, Standard Chartered and UOB as lead arrangers. A three-year loan on a long-lived asset means PDG will be back in the loan market once the campus fills.

Nscale and its Norway site. Nscale’s flagship site at Glomfjord, Norway, runs 6.4MW of a planned 30MW and sits about six months behind schedule, Bloomberg reported. Bloomberg cited cooling leaks, local rules, construction delays and a fire. The company is reportedly preparing an IPO at a valuation of up to ~$35 billion. Public buyers will price that listing on megawatts delivered, and Glomfjord is the first number they’ll ask for.

GLOBAL LENS: THIS WEEK BY REGION

North America: the county signs, the state still decides

Mason County, Kentucky, approved development and bond agreements on Oct. 13 for a ~3,000-acre campus, and Meta was named as the company behind it. The first phase is reportedly ~$10 billion, though the county judge-executive has cited $15 billion, WVXU reported. The deal runs through an industrial revenue bond lease-back, so the county holds the deed and Meta makes fixed payments in place of property tax. State incentive approval is still pending, and that vote now decides when the first phase starts.

Big Watt Digital and PowerHouse Data Centers won approval from the Taylor City Council for a campus in Williamson County, Texas. It covers 664 acres and up to 1GW. It will run on natural gas behind the meter, and the developers asked for no tax incentives. First power is targeted for the first half of 2028, and no tenant is named. Gas on site gets the campus around the grid queue, but it still has to find a buyer for the space.

Europe: own the power, then sign the tenant

Echelon Data Centres broke ground on what it calls Ireland’s first green energy park, at its Arklow campus in Wicklow. The €4.2 billion plan covers two data centers and a 220kV substation shared with SSE’s 800MW Arklow Bank Wind Park 2. On-site generation runs on HVO, a renewable diesel. Completion is set for 2028, but the wind farm is still in development. The campus now runs on a wind farm timeline it doesn’t control.

Nebius signed a letter of intent to be the first customer at Volt’s Port of Rotterdam AI campus. Volt plans 600MW across six buildings, and its first phase of two 100MW buildings costs ~€7 billion including compute. Eneco will supply power from offshore wind. The LOI isn’t a lease, and Nebius hasn’t said how many megawatts it will take. A signed lease is what Volt needs to finance phase one.

Asia-Pacific: India builds, but only one has a buyer

Colt DCS and RMZ began construction on a third data center at their Navi Mumbai campus. The 64MW building is fully pre-let to an unnamed hyperscaler. Contracted power and a 220kV substation are already secured. A full pre-let before the building goes up moves the leasing risk off the developers. The next test is whether the substation energizes on the hyperscaler’s schedule.

CapitaLand unveiled a 53MW data center in Chennai, with a planned investment of ~INR 1,940 crore. It offers 34MW of IT capacity across 13 halls and is due by the end of 2026. Its listed India trust and a dedicated data center fund own it jointly. The site was first planned for a 2025 opening, so it’s running about a year late. No tenant has been named, and the first lease will set what Chennai capacity rents for.

Share

CONSTRAINT OF THE WEEK: THE PLANNING VOTE

The planning vote decided more projects this week than the power supply did. The UK government rejected the York and North Yorkshire AI Growth Zone bid, which centered on a Drax campus of at least 1.2GW. It said the bid fell short on credible energy delivery, environmental impact and local benefit. Drax now plans a 100MW first phase on the national grid, with a planning submission due in the coming months. That first phase is a twelfth of what the bid promised.

Waterford City and County Council refused permission for a data center and energy campus at Kilmeaden. The plan carried three data center buildings, 16 gas turbines, solar panels and a battery. The Irish Examiner reported that Echelon is also behind the project. On-site gas didn’t win consent, and any appeal now goes to An Coimisiún Pleanála.

The 150MW Norwich Apex Data Centre drew objections from local MP Ben Goldsborough and broadcaster Jake Humphrey within a week of its application. The Keswick site sits beside the protected Marston Marsh, and the plans include 72 rooftop chiller units. South Norfolk Council will decide. Once an MP objects in week one, the council vote turns political before the first officer report lands.

The next bottleneck is the ballot box. Amazon Web Services held an open house for an 18-building, 600-acre campus outside Trenton, Ohio. It also said it will stop asking local officials to sign NDAs. On Nov. 3, Trenton votes on Issue 17, a data center ban. A town vote can now stop a campus before any planning board sees it.

THE MECHANISM

A utility builds off its map for one buyer

American Electric Power is one of the largest US utilities, with 5.6 million customers in 11 states and about 33,000MW of generation. On Oct. 14 it committed ~$4 billion to a 702MW fuel cell plant near Cheyenne, Wyoming. Google will buy the output under a 20-year power purchase agreement. CEO Bill Fehrman said the plant serves needs “beyond our traditional service territory.”

The seller is Cheyenne Fuel Cell, LLC. AEP owns about 92%, and funds advised by Basalt Infrastructure Partners hold nearly 8%. Bloom Energy supplies fuel cells that make power from natural gas without burning it, and the plan adds carbon capture. Construction is planned to start this year, with full operation by the end of 2028.

That contract is what makes the plant financeable. No ratepayers stand behind this plant, so AEP’s return rests on one customer’s signature. For Google, a dedicated plant puts its power supply on a contract it controls.

The release gives no contract price. It doesn’t say who carries the gas cost for 20 years, or which permits the plant still needs. So, the risk sits with one buyer, one fuel and one fuel cell vendor, and the disclosures leave the fuel side blank.

Leave a comment

NOTABLE TRANSACTIONS

Fermi: AMD credit support behind a $6.5 billion TensorWave lease

Fermi said on Oct. 13 that AMD signed a credit support agreement for its Project Matador campus in Carson County, Texas. The support backs TensorWave’s lease of 222MW of power and 156MW of critical IT. The lease carries $6.5 billion of contracted revenue over a 15-year initial term. The campus runs on its own private grid behind the meter.

What makes it notable is that a chipmaker now stands behind its customer’s rent. TensorWave runs a GPU cloud built on AMD chips, so the chipmaker is backing demand for its own product. It is rated A by S&P and A1 by Moody’s.

AMD’s obligations begin only after construction is complete and rent starts on the final phase. The support ends if TensorWave or its guarantor holds an investment-grade rating for 12 straight months. The amount isn’t disclosed. Until rent starts, Fermi and its lenders carry the construction risk.

What to watch is the project financing. Fermi calls the agreement a step toward it and lists financing as a risk. The first construction loan will show what a backstop that starts after completion is worth to lenders.

NeoAI Cloud: KRW600 billion financing for a 40MW Pohang AI data center

NeoAI Cloud closed KRW600 billion (~$445 million) of financing for a 40MW AI data center in Pohang, South Korea. SC Capital Partners joined as a key equity investor. The project loan closed on Oct. 7, and service is targeted for the fourth quarter of 2027.

What makes it notable is how much support sits under one small building. A pre-lease, a contractor’s completion commitment and regional public money all back the first phase.

The money splits into a KRW450 billion project loan and KRW150 billion of equity. Woori Bank and KB Kookmin Bank arranged the loan. North Gyeongsang put in KRW4 billion and helped bring KRW131.5 billion from the Regional Revitalization Investment Fund. NHN Cloud has pre-leased 16MW, half the IT load, and Hyundai E&C committed to complete construction.

What to watch is the 260MW second phase. Phase one leased half its IT load before completion. Phase two will test whether lenders fund a build more than six times larger on the same terms.

EdgeConneX and Stockland: 50-50 Australian venture with 450MW of power secured

EdgeConneX entered Australia through a 50-50 joint venture with Stockland, announced Oct. 14. The venture has secured 450MW of power for hyperscale campuses in Sydney and Melbourne. No capital figure, tenant or timeline was disclosed.

What makes it notable is that a listed property company takes half of an operating platform. Stockland brings land, development and project management. EdgeConneX brings hyperscale delivery.

The power came before the tenant. Further power applications and planning work are in progress. Chris Maher was appointed managing director for Australia.

What to watch is the first planning application and the first signed tenant. Either one will show what a half share in land plus power is worth to Stockland’s shareholders.

THE FRAMEWORK: WHO BRINGS THE POWER

Every project that moved this week had a power plan it controlled. The table sets each one against its buyer and where it stands.

THE QUESTION FOR NEXT WEEK

Does Fermi close the TensorWave lease by Oct. 31? That is the deadline the two companies set in September. A close would put the AMD backstop to work and open the door to project debt. Another extension would tell lenders the guaranty still isn’t settled.

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

Share 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.

This article is the 18th article in the series: Ground to Grid: A Free 21-Lesson Guide to Mastering Data Center Development

Join Global Data Center Hub

Why Handover Is Where Development Becomes Operation

A commissioned facility earns nothing until a tenant is live inside it. Handover is the moment development ends and the operating business begins. A clean handover is what makes the asset start performing.

Handover sits at the end of the development lifecycle. Before it, a project team builds and tests the facility. After it, an operations team runs that facility for the life of the lease. Handover is the bridge between the two.

The bridge carries real money. At the end of 2025, Digital Realty reported $817 million of annualized GAAP base rent (yearly rent under standard accounting rules) from leases signed but not yet commenced, at its share.

So, onboarding a tenant is an operational discipline. The lease milestone is only one part of it.

The tenant has to install equipment safely, receive a working service, and start paying on a date both sides can prove.

The quality of that work also shapes the tenant relationship for the life of the lease.

The Steps of Onboarding and Handover

Handover runs in four steps. They overlap, but each has its own owner.

1. Tenant Fit-Out and Deployment

The tenant installs and commissions its own equipment. That means racks, network gear and, in AI halls, liquid cooling connections.

The operator and the tenant have to coordinate closely to bring that load online safely. Every person, vendor and shipment is approved before arrival.

NIST SP 800-53 Rev. 5 sets out the access controls: authorize access, issue and remove credentials, log entry, escort visitors.

All work inside a live facility runs under an approved method of procedure (a written, step-by-step work plan), a change ticket and a back-out plan.

Network service is often the critical path. A cross-connect is complete when optical levels, continuity and speed are recorded and the customer or carrier accepts it.

AI tenants raise the bar. Uptime Institute notes that direct liquid cooling physically links IT and facility systems. Coolant distribution units need factory witness tests and load simulation. Practice is still maturing, so the test scripts belong in the contract.

2. Service-Level Activation

At this step the SLA (service level agreement: the promised performance and the remedy for missing it) moves from a contract to a live operating commitment.

An SLA works only when someone can measure it at an agreed point.

Verizon’s public U.S. colocation SLA defines power unavailability at the customer cabinet. It ties verified downtime to service credits.

So monitoring, reporting and response processes go into effect here. Each SLA line needs a named object, a handoff point, a metric, a data source, start and stop rules, exclusions and a remedy.

How those terms are drafted also decides who absorbs the loss when performance slips.

3. Documentation and Turnover

The operating team needs a full record of what it is about to run. That includes as-built drawings, commissioning records, operating procedures, equipment manuals, warranties, maintenance plans and training records.

ASHRAE/IES Standard 202-2024, the current published minimum commissioning standard for new buildings and systems, sets a consistent approach to documentation and training. ASHRAE’s guideline written for data centers, Guideline 1.6, is still only a proposal.

The site operations leader should formally accept these materials. When operations inherits them by default, nobody has checked that they are complete.

Much of that record is produced earlier, during the trials that prove a facility can carry full load.

4. Operational Handover

The last step moves accountability from the development and construction team to the operations team. From this point, long-term management owns the facility, the tenant and the SLA.

The cleanest way to mark that point is a signed commencement record. It states the effective date, the services, the committed capacity, any accepted open items, the SLA start, the billing start and the opening meter readings.

A short stabilization period follows, with closer monitoring and a final reconciliation of alarms, access lists, circuits, meters and invoices. Then the service moves to normal review.

Why Handover Quality Matters

A poorly documented handover leaves the operations team blind to the facility it runs.

The team cannot maintain equipment it does not know about or follow a control sequence nobody gave it.

The cost of that blindness shows up in outages. In Uptime Institute’s 2026 outage analysis, 57% of respondents to its 2025 survey said their most recent major outage cost more than $100,000. One in five said their most recent impactful outage cost more than $1 million.

Uptime counts human error as a contributing factor in nine out of ten outages.

Uptime also cautions that outage data has limits. Those figures cover all outages, but procedures, training and configuration are set up at handover.

A rushed tenant onboarding spoils the first impression, and that impression sets the relationship.

The SLA becomes real at handover. The operator has to meet it from day one, including on the days the team is still learning the site.

Gaps between development and operations then surface as early operating failures.

What Separates a Clean Handover

The strongest handovers share four features.

First, complete documentation goes to an operations team that is ready for it.

The team was staffed and trained before the tenant arrived, and it signed for what it received.

Second, tenant onboarding is coordinated so load comes online without incident.

Work runs under approved procedures, and capacity is reserved in the company’s systems of record before the tenant’s equipment ships.

Third, SLA processes are tested and active before the tenant depends on them.

Strong operators raise a test alarm, confirm the customer is notified, calculate a sample credit and produce a sample monthly report.

Fourth, the transition is deliberate. Development and operations overlap for a planned period, with one accountable owner from contract signature to stabilization.

The contract, the operating records and the first invoice all show the same start date.

A Real-World Example: How Digital Realty Reports the Wait

Digital Realty shows what this interval looks like in public numbers. It defines backlog as annualized GAAP base rent from signed leases not yet commenced.

At year-end 2025, backlog stood at $817 million at Digital Realty’s share. Its investor materials put total backlog at about $1.4 billion at 100% share.

In the fourth quarter of 2025, Digital Realty booked $400 million of annualized GAAP rental revenue at 100% share, or $175 million at its own share. For those fourth-quarter leases, the average wait from signature to contractual start was eight months.

That lag reflects normal delivery schedules in a growing business. Each of those leases still has to pass through the four steps above before its rent begins.

Strategic Implications

For developers: Your reputation is set at handover. A clean transition is what earns the next deployment from the same tenant.

Plan the documentation package and the operations team’s training from the first design review.

For investors: Handover quality predicts the early operating performance your returns depend on.

Ask what share of signed rent has not commenced, how the company defines “commenced,” and who signs for the handover on the operations side.

For operators: Everything you inherit at handover defines your starting position. That includes the documentation, the SLA commitments and the tenant relationship.

Refuse to start the SLA clock until you have tested the reporting.

For policymakers: Reliable operation from day one is what makes the sector a stable part of the digital economy you are trying to build.

EU Delegated Regulation 2024/1364 applies to data centers with at least 500 kW of installed IT power and lets colocation operators gather required data from customers.

Give operators time to write such duties into tenant contracts.

Key Takeaway

Development ends and the business begin at handover. Everything before that day is cost. From that day on, the asset has to earn.

The tenant does not experience the years of development. It experiences the day it goes live: whether its racks went in safely and whether the first alarm reached the right person. A clean handover makes that day the start of a working relationship.

That day also sets the numbers. Rent starts on the commencement date, and the SLA clock starts with it. Every week of slip moves both further out. Digital Realty’s eight-month lag shows how much rides on that interval.

So, treat handover as the first day of operation. Staff the operations team before the tenant arrives.

Test the SLA before anyone depends on it. If you leave handover to the end, the first months of the lease become a repair job.

In your experience, what is the handover detail most often overlooked, and how does that oversight show up in the first months of operation?

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 late 2022, Meta wrote off $1.3 billion of data center work to redesign for AI heat.

The newest AI racks are cooled with liquid, while most data centers run far cooler machines.

Every ChatGPT answer ends the same way. It becomes heat.

The chips that write the answer run on electricity, and almost all of that electricity leaves the building as warmth. For decades, data centers moved that warmth out with fans and cold air.

The newest AI chips run hot enough that the buildings around them need pipes.

Meta, the company behind Facebook and Instagram, learned this in the middle of a building spree.

Data Center Frontier, a trade publication, reported in April 2023 that Meta had dozens of data centers under construction. Many had been planned around ordinary computer chips.

In December 2022, Meta paused construction in Huntsville, Alabama; Temple, Texas; and Kuna, Idaho. It also dropped two of the three data centers it had planned near Odense, Denmark. On February 1, 2023, Meta told shareholders it had shifted to “a next-generation data center design” that could support AI.

Its results for the last three months of 2022 included $1.3 billion in charges for abandoned data center assets. In April 2023, Reuters reported that the new buildings would need liquid cooling to handle the heat of AI chips.

The redesign cost time and money. Temple and Kuna restarted about ten months after the pause, DatacenterDynamics reported on October 19, 2023.

The problem has kept growing since. A data center is planned around the chips of its day and built to run for decades, but AI chips change much faster.

The International Energy Agency reported in 2026 that AI server power density rose 11 times between 2020 and 2025.

A Shelf of AI Chips Gives Off 30 Boilers of Heat

The heat starts in the rack. A rack is a tall shelf, about the size of a refrigerator, that holds dozens of servers. A server is a desktop computer with the screen removed, flattened into a tray.

For most of the internet’s history, racks were cooled the way a person cools off in front of a fan.

Cold air blows across the front of the servers, picks up their heat, and leaves out the back. Machines outside the building then push that heat into the open air.

That works as long as the heat stays modest. Uptime Institute, a group that surveys data center operators every year, found in July 2025 that the typical rack in its survey drew a little under 9 kilowatts.

AI racks sit on another scale. NVIDIA, the company whose chips power the AI boom, designed its GB200 NVL72 rack to need 120 kilowatts of cooling, according to a design note it published on October 15, 2024.

That is more than 13 times the typical rack in Uptime’s survey.

The IEA put the same scale in plainer terms in its 2026 report: AI data centers must remove the heat of about 30 natural gas boilers from each rack.

At that level, NVIDIA cools the rack with liquid, much as a car radiator cools an engine.

Metal plates sit directly on the chips. Coolant flows through the plates, soaks up the heat, and carries it through pipes to equipment that cools the fluid and sends it back.

Liquid holds far more heat than the same amount of air, so a thin pipe can carry much more of it.

So, a building made for AI needs pipes running to every rack, pumps to push the coolant, and equipment to pull heat out of the fluid. A building made for air has fans, open floors, and none of that plumbing.

Changing one part tends to ripple through the rest, as we showed in our walk-through of the five systems every facility depends on.

Most Operators Still Run Racks Under 30 Kilowatts

A newcomer could fairly assume that a data center is a data center: a secure building full of computers, ready for whatever chips come next.

The assumption has history behind it. For years, upgrading a data center mostly meant sliding newer servers onto the same shelves, cooled by the same air.

News coverage also measures data centers in megawatts, a number that shows how much electricity a building can draw and leaves out how it removes the heat.

Two buildings with the same megawatts can look identical in a headline.

Uptime’s 2025 survey shows how far most buildings sit from the AI frontier. More than 80% of respondents said their facility had no racks above 30 kilowatts.

Next to NVIDIA’s 120-kilowatt rack, most of the buildings in that survey were set up for a different kind of machine.

The clearer picture of the data center world has two parts. A large group of buildings runs ordinary computing on air. A newer group is being built around liquid for AI.

Meta’s 2022 pause shows what happens when a company finds its plans sitting in the first group while its future needs the second.

Old Buildings Stay in Service for Decades

The fix takes years because buildings last much longer than chips. In Uptime’s 2025 survey, 27.5% of respondents worked in data centers 16 years old or older.

Those buildings opened in 2009 or earlier, long before today’s AI chips existed, and they are still running.

Adding liquid cooling to a working building is slow work. Pipes have to run through halls full of live computers.

Pumps and new heat equipment need floor space and power of their own. Every change has to happen while customers’ computers keep running inside.

Even new campuses lose time adjusting. Meta’s redesign delayed Temple and Kuna by about ten months.

Microsoft, the company behind Windows and a major AI cloud, paused early work on the second phase of its Mount Pleasant, Wisconsin, campus in January 2025.

It told The Register on January 3, 2025, that it was evaluating “scope and recent changes in technology” and how they “might impact the design of our facilities.”

On the same Wisconsin campus, Microsoft said in September 2025 that Fairwater would cool over 90% of the facility using a closed-loop liquid system. On June 23, 2026, it announced that its first Mount Pleasant data center was fully operational.

The bigger story is pace, and we tracked how the company is adding about a gigawatt of capacity every quarter.

Why the Heat Problem Reaches the AI on Your Phone

If you follow AI news, the heat problem explains why a new chip takes time to reach you. A faster chip helps only once a building can cool it.

ChatGPT launched on November 30, 2022, and a few weeks later Meta stopped building to redraw its data centers for AI.

The next AI tools will run in buildings that can carry away their heat.

If you live near a planned data center, cooling shapes how much water it uses. Some cooling systems shed heat by evaporating water, like sweat cooling skin.

Closed loops keep the same fluid moving around a sealed circuit.

Microsoft said in September 2025 that Fairwater’s loop is filled once during construction and then reused, and that the building would use roughly the water a typical restaurant uses in a year. That figure is Microsoft’s own.

Many operators publish far less: Uptime found in 2025 that only 47% of operators collected water data at all. At a public hearing for a new data center, it's fair to ask about the cooling design.

Four Clues That Show a Data Center’s Cooling

Start with the cooling words in an announcement. Terms like liquid-cooled, direct-to-chip, and closed-loop mean the building was designed for AI-grade heat.

Microsoft used that language for Fairwater in its September 2025 post, and company blogs and press releases are where these words show up first.

Next, watch for a pause to change the design. Meta used those words for Huntsville in December 2022, and Microsoft pointed to “recent changes in technology” in Wisconsin in January 2025.

Both statements tied a construction pause to the building's design, and in Meta’s case, the new design was built for AI. Local news and company statements carry these pauses first.

Then look for a charge for abandoned data center assets in a company’s quarterly results. Meta’s $1.3 billion line for late 2022 showed in one number how much unfinished building it was leaving behind.

These lines sit in earnings releases and in filings with the US Securities and Exchange Commission.

Finally, compare the rack number in a chip launch with what buildings run today.

NVIDIA’s 120-kilowatt rack sits far above the typical rack, which runs at a little under 9 kilowatts, in Uptime’s 2025 survey. Chipmakers publish their rack designs, and Uptime publishes its survey every year.

The AI people use every day feels weightless. Underneath it sits a shelf giving off the heat of 30 boilers, inside a building that has to carry that heat away every second.

The buildings going up in 2026 are being planned for the chips of 2026. Will they still be able to cool the chips of 2030?

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

Higher interest rates hit data centers through one gap: the time between when rent resets and when debt resets.

An owner whose rent moves first can absorb this cycle. An owner whose debt moves first carries it.

The rates are real. On September 16, the Federal Reserve raised its target range to 3.75% to 4.00%, its first hike in three years (Federal Reserve).

The 10-year Treasury closed at 5.29% on September 30, up from 4.19% on January 2 (FRED). The Fed’s median projection holds the policy rate at 4.1% through the end of 2027 (Federal Reserve).

The money is still moving. Bonds and securitizations backed by data center leases reached $17 billion in the first half of 2026, up 29%, JLL reported (JLL).

Vacancy in major U.S. markets sits at a record low of 1.4%, CBRE reported (CBRE).

Lenders are still lending. They are pricing the calendar.

The Buildout Is One of the Pressures the Fed Is Watching

Fed officials have tied part of the rate picture to AI spending. In July, Governor Lisa Cook named “increased capital expenditures tied to the buildout of AI infrastructure” as one of two unanticipated shocks of 2026 (Federal Reserve).

In September, Chair Kevin Warsh said hyperscalers “are out in the market raising funding, and so the competition for capital is real” (Reuters).

Keep the weight right. TD Economics estimates that about 50 of the roughly 80 basis points the 10-year has risen since late February came from expectations of more Fed hikes (TD Economics).

Reuters lists tariffs and higher energy prices beside AI (Reuters). AI spending is one of several forces the Fed is leaning against.

For an owner, the shape of the pressure matters most.

The Fed’s median shows no cut in 2027. Plan around rates near today’s level and find which side of your balance sheet feels them first.

Debt Reprices First

Construction loans feel a rate move almost at once. They float over SOFR, the overnight rate most construction lending prices from, which stood at 3.90% on September 28 (FRED).

JLL reports construction loans for investment-grade hyperscale tenants at spreads in the low 200s over that rate, for up to 85% of cost (JLL). That puts the all-in cost near 6%.

Bonds show the same pressure. In April, a Microsoft-leased QTS campus priced bonds at 5.70%. In August, another Microsoft-leased QTS deal priced at 7.228%, according to the CCIR deal ledger, which draws on press reports (CCIR).

The 10-year rose about 40 basis points between those dates. Most of the remaining gap came from wider spreads on AI-linked debt. The two bonds differ in structure and term, so read the comparison as directional.

Fixed-rate debt delays the hit. It does not remove it. Digital Realty carries a weighted-average coupon of 2.9%, and 90% of its debt is fixed, with maturities peaking in 2029 and 2030 (Investing.com).

Its guidance for new long-term debt moved from 4.0% to 4.5% in February to 4.5% to 5.5% in July (Digital Realty). The higher cost arrives one maturity at a time.

Leases Reprice on Their Own Schedule

Rent moves on two tracks: the yearly escalator and the renewal.

The escalator is the slow track. The three Northern Virginia hyperscale buildings Digital Realty agreed to buy on June 29 carry 15-year leases with rent rising 3.6% a year (Digital Realty 8-K).

The Fed projects 3.7% PCE inflation for 2026 and 2.3% for 2027 (Federal Reserve). This year, that escalator roughly keeps pace with prices.

If the 2027 forecast holds, it pulls ahead. A fixed escalator protects the owner as long as inflation stays near it.

The renewal is the fast track. Digital Realty’s second-quarter renewals came in 25.4% higher on a cash basis (Digital Realty). Leases above one megawatt renewed 66.7% higher. Leases under one megawatt renewed 5.2% higher (Investing.com).

CFO Matt Mercier put it plainly: “our expiring rates continue to drop over the next several years, while market rates continue to march up” (earnings call).

Short contracts reset most often. Equinix’s revenue contracts carry initial terms of one to five years, then renew automatically in one-year increments, its quarterly filing shows (Equinix 10-Q). Each renewal is a chance to price at today’s market.

New lease terms move cost too. CBRE reports take-or-pay minimums of 60% to 85% of allocated power. It also reports colocation operators beginning to adopt triple-net leases, where tenants pay taxes, insurance and upkeep directly (CBRE).

Both shift cost inflation from the owner to the tenant.

Four Positions on Two Clocks

Set the lease clock beside the debt clock and four positions appear.

The steadiest pairs short or frequently renewing contracts with long fixed-rate debt.

Rent resets to market often. Interest cost stays put. Higher rates arrive late, and renewals can outrun them.

Next is a long lease with escalators, funded with long fixed debt. Income and interest are both locked. This position barely notices a rate move until refinancing, when a new coupon meets an escalator set years earlier.

Tighter is a long lease with a fixed escalator, funded with floating construction debt.

Rent is set for 15 years. Interest moves with every Fed decision. The escalator has to cover the gap until the project refinances into fixed debt.

The most exposed is construction without a credit tenant. JLL puts those loans 200 to 300 basis points wider than investment-grade deals, at 70% to 80% of cost (JLL).

With SOFR near 3.90%, that implies 8% to 9% or more all-in. Until a lease is signed, only the debt clock runs.

Most real portfolios hold all four. What matters is the weight on each.

For a live case of that mismatch, see a GPU cloud whose borrowing outlasts its customer deals by two years.

Value Shows the Gap First

Prices move before cash does. Digital Realty expects an initial stabilized cap rate above 6.5% on its June purchase (Digital Realty 8-K).

A cap rate is a building’s yearly income divided by its price. On signing day, the 10-year was 4.38%, a cushion of at least 212 basis points. On September 30, the cushion was at least 121.

If cap rates rise to rebuild that cushion while income holds flat, values fall. A move from 6.5% to 7.0% cuts value by about 7%. A move to 7.5% cuts it by about 13%.

Rent growth is the offset, and here the two clocks matter again. An owner who can re-sign leases 25% higher can grow income into a higher cap rate. An owner on a 3.6% escalator waits for the years to add up.

New building still pays. Digital Realty’s 1.4-gigawatt development pipeline is expected to deliver an 11.5% stabilized yield (earnings call).

That is about 620 basis points over the September 30 10-year. Higher rates thin the margin on existing buildings. On new ones, margins are still wide.

The other side of that June purchase tells its own story, which we covered in why Blackstone let go of finished, fully leased capacity.

Outside the U.S., the Clocks Run on Other Rates

The same test applies abroad with different inputs.

Gulf central banks hiked with the Fed because their currencies are pegged to the dollar.

Saudi Arabia’s repo rate rose to 4.50% (Reuters). Dollar debt in the Gulf feels U.S. moves directly.

In India, dollar loans price off SOFR. AdaniConneX’s roughly $800 million construction loan priced at about SOFR plus 200 basis points, Bloomberg reported (Bloomberg).

Part of the revenue there comes in rupees. Two clocks become three: rent, debt and currency.

Nigeria cut its policy rate 350 basis points in September, to 23% (Central Banking).

For many operators there, energy is more than 35% of operating cost, Nairametrics reported (Nairametrics). Power sets the math more than the rate does.

What to Check Before the Next Deal

The rate picture will not settle by year-end. The Fed meets again on October 27 and 28. Investors are watching 6% on the 10-year as a pain point, Reuters reported (Reuters).

So put every rent reset and every debt reset on one calendar and find where debt resets first. Compare each escalator with the Fed’s inflation path for the next two years.

Then ask whether the lease moves power and operating cost to the tenant, through take-or-pay floors or triple-net terms.

One question stays open. Renewal spreads this wide depend on tight supply.

CBRE counts less than 1,500 megawatts of future capacity available across primary U.S. markets, about six months of leasing (CBRE).

If that supply loosens, the fast clock slows. Then the debt clock you set today is the protection you have left.

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 Q3 2026, AI labs signed some of the largest data center contracts of the year. The company signing for capacity was often not the one owning the campus, raising the debt or backing the payments. If you build, lend to or invest in this market, that changes the questions you need to ask.

New Buyers, More Backers

AI labs became direct buyers of data center capacity in Q3 2026. Anthropic agreed to about $45 billion of capacity over six years with Nscale.

OpenAI announced a $20 billion campus of its own in Georgia and is the named customer for 8 GW at SB Energy’s PORTS-Pike campus in Ohio.

The money around them came through more hands than before. Banks lined up $15 billion for an Anthropic campus in Texas, with Google’s credit rating standing behind Anthropic.

Nvidia gave credit support at PORTS-Pike and set a $500 billion financing target with six of the largest investment firms. In El Paso, BlackRock became majority owner of a Meta campus while Meta kept a guarantee on part of its value.

I read this as good news for capital formation. More partners make money easier to raise. The open task is to widen the base of demand behind that money, because many of these structures still lean on a small group of labs and suppliers.

Most of the capital in this quarter’s list is also attached to capacity that is still being built, powered or contracted. For you, the opportunity is real, and so is the work of checking what each commitment actually covers.

How to Read This Quarter

Five Structural Trends

1. AI Labs Became Direct Buyers

Anthropic with Nscale in West Virginia. OpenAI in Effingham County and at PORTS-Pike. OpenAI again through Firmus in Asia-Pacific. The labs signed for capacity directly, and one report in the dataset put Anthropic’s compute deals at $517 billion over 11 months.

They were not the only buyers: Google, Meta, ByteDance and SK Telecom all committed capital too.

The labs also want capacity sooner than gigawatt campuses can deliver it, and both OpenAI and Anthropic were reported to be seeking 20 to 30 MW sites in the meantime. If you own ready capacity at that size, you are in a good position this year.

2. More Partners Now Stand Between the Tenant and the Money

The company using a data center is often not the one owning it, borrowing for it or backing its payments, and those roles are not interchangeable. Nvidia’s $500 billion platforms are a fundraising target.

At PORTS-Pike, Nvidia gave formal credit support on land, power and shell. On the Nexus loan, Google backstopped Anthropic with its credit rating. In El Paso, BlackRock became the majority owner while Meta kept a residual value guarantee.

Each of these moves a different piece of risk. If you lend into this market, “who is the tenant?” is no longer enough. Your question is which payments are supported, by whom, on what conditions, and which risks stay with the project.

3. Bitcoin Miners Became AI Landlords

Hut 8, Bitdeer and CleanSpark signed long-term AI leases. Riot reached a deal with Anthropic, IREN added $2.8 billion in contracts, and Core Scientific agreed to give AMD access to more than 500 MW.

The miners spent years assembling power and land, and in Q3 several turned those sites into AI leases running 16 to 20 years.

The value depends on conversion: cooling, network links, reliability, and the cost of meeting AI customers’ specifications. Where those line up, this is one of the quickest routes to powered capacity.

4. Local Policy Moved Into the Financing

States are setting clearer rules for large data centers, and the approaches differ. New York paused new state permits for hyperscale projects of 50 MW or more for one year.

Texas paused grid-connection approvals for new data centers while regulators review a queue of about 474 GW of proposed demand, roughly 90% of it from data centers.

Pennsylvania made its development standards mandatory and tied state permits to local approval, without a moratorium. These rules give developers a clearer path, and they also feed into financing.

Some lenders are already factoring permitting timelines into how they price construction debt. If you are planning a campus, your permitting plan is now part of your financing case.

5. Resilience Became Part of Site Planning

The Gulf remains one of the fastest-growing data center regions, and this year it also showed why resilience planning matters. Regional security events affected some facilities, and AWS advised some customers in Abu Dhabi and Bahrain to consider alternatives while repairs continue.

Investment kept coming. Microsoft framed its $10 billion Gulf strategy around resilience, Microsoft confirmed its Saudi Arabia East region for November 2026, AWS and Humain committed to 50 MW in a Saudi AI zone by 2028, and Humain is seeking $2.5 billion for new capacity.

If you are building or financing in the region, redundancy and physical security now sit next to power and land in your plan.

Where the 12 Announcements Stand

Capacity moves through stages: announced, site controlled, customer committed, power contracted, financed and under construction, energized, and finally accepted and earning revenue. The dollar figures in this list sit at very different points on that path, and knowing which point helps you compare them.

Where the Openings Are

1. Lending Where a Strong Backer Supports Specific Payments

The Google-backstopped Nexus loan moved toward bank sell-down within weeks, and an Anthropic-linked data center raised $1.3 billion in private credit. The opening is for lenders who read the support documents, not only the press release: what is backed, by whom, for how long, and what stays with the project.

2. Mining Sites That Can Be Converted at the Right Cost

Hut 8, Bitdeer, CleanSpark, Riot, IREN and Core Scientific all signed AI deals in the quarter. Your opportunity is in the sites that can meet AI customers’ needs at a conversion cost and on a timetable the lease supports.

3. Public Listings for Builders

Nscale filed for a US IPO. Switch lined up a $10 billion listing at a potential $80 billion valuation, SB Energy filed publicly, Firmus courted investors for an ASX listing, and Blue Owl planned a listed data center REIT. These are filings and plans, not completed raises. As they price, early backers gain a route to liquidity and builders a new source of equity.

4. Emerging Markets With New Rules and Funded Projects

Brazil signed its Redata tax regime into law. Indonesia drew a $3.1 billion loan for Zankore’s 1 GW project, plus CoreWeave’s first APAC entry. Johor kept raising multi-billion-dollar financing, and Egypt signed a $1 billion data center and sovereign AI deal. In each case, policy arrived with money attached.

5. Mid-Size Capacity That Can Deliver Soon

OpenAI and Anthropic were reported to be seeking 20 to 30 MW facilities while larger campuses are built. If you own ready capacity at that size, you may have pricing power, provided your site meets the labs’ needs on power density, cooling, connectivity and location.

What to Plan For

1. Widening the Base of Demand

Different developers, lenders and financing structures can still depend on the same few labs and suppliers. More partners make capital easier to raise. Spreading the demand behind it is the next step, and it is worth tracking in your own portfolio.

2. Power Costs Became a Policy Question

California signed seven bills requiring data centers to disclose power and water use and to keep the cost of new power off other customers’ bills. The US House passed the Ratepayer Protection Act 417-3. It asks states to consider a standard that makes large users pay the full cost of new generation and transmission, and it now goes to the Senate.

Virginia ordered a review of backup generators and named Eastern Virginia a cooling-water-scarcity area. Thailand asked operators to pause 49 buildouts while it completes a legal framework.

Clear cost rules help you plan. They also mean your project should budget for its own power and grid costs from the start.

3. Credit Terms Tied to Ratings and Permits

Some utilities now ask large data center customers below a set credit rating to post collateral sized to the power plants and lines built for them, which can run into the billions for the largest projects.

The El Paso debt priced at about 7.5%, and some lenders are factoring permitting timelines into construction loans. If you are raising money, your credit rating and your permits now shape your terms as much as your tenant does.

4. Resilience in Fast-Growing Regions

Recovery timelines for some affected Gulf facilities are still being set, and the UAE has revised its AI data center plans. Redundancy and site choice in the region now carry a planning cost that earlier budgets did not include.

5. Gigawatt Campuses Need Time

Large campuses in several markets have taken longer than planned, and grid connection timing has shaped schedules in parts of the UK, where Ofgem has proposed reforms to prioritize ready projects.

The labs are signing faster than big campuses can be energized, which is why smaller, ready sites are in demand.

What Would Change My View Next Quarter

• Delivery dates. Whether Nscale’s West Virginia campus, Meta’s El Paso campus and OpenAI’s Effingham site hold their schedules.

• The terms behind the backing. Whether the Nexus loan closes, and what Google’s support covers once the terms are public.

• The state rules. What the Texas grid review concludes, whether the Ratepayer Protection Act passes the Senate, and how New York’s one-year permit pause ends.

• Demand breadth. Whether new lab and enterprise customers join the contracts, widening the base behind the money.

• Gulf recovery. Whether recovery dates are set and how new capacity in the region is designed for resilience.

The Full Ranked List: Top 12 Global Announcements (Q3 2026)

The 12 announcements below are ranked by capital scale, strategic signal and structural impact. The dollar figures measure different things, from contract value to financing targets, so each entry says which one it is.

1. Anthropic Agrees $45 Billion AI Data Center Deal With Nscale

Anthropic agreed to rent capacity from Nscale, a UK start-up, in a deal worth about $45 billion over six years. That is contract value, not money invested. It covers 460 MW at Nscale’s flagship campus in Mason County, West Virginia, with Anthropic’s access starting when the chips come online in late 2027. About a month later, Nscale filed for a US IPO, which gives public investors a view of the contract and the delivery plan behind it. Nscale’s next chapter centers on delivering this campus on time. [Read here]

2. Nvidia Builds $500 Billion AI Compute Financing Platforms With Six Major Investors

Nvidia partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on financing platforms that aim to mobilize more than $500 billions of third-party capital. That figure is a target, not committed money. The firmer step came the same month at SB Energy’s PORTS-Pike campus in Ohio.

There, Nvidia agreed to give credit support on the land, power and shell buildout for an initial 4.25 GW of IT capacity, with an option on the rest, and to be the site’s AI compute supplier. OpenAI is the named customer for 8 GW. The chip supplier is now helping fund the buildings its chips go into. [Read here]

3. Banks Line Up $15 Billion for Google-Backed Anthropic Data Center

In late July, Morgan Stanley was leading advanced talks to lend $15 billion to Nexus Data Centers for an Anthropic campus in Hubbard, Texas. Google agreed to backstop Anthropic with its investment-grade credit rating. By early August, the banks were reported to be preparing to sell the debt on to other investors. If you are a lender, the key is the exact scope of Google’s support, meaning which obligations it covers and on what conditions. [Read here]

4. Brookfield and NextEra Plan $100 Billion Campus in Paducah, Kentucky

Brookfield Asset Management and NextEra Energy are reported to be developing a $100 billion data center campus in Paducah, Kentucky. That is a project estimate from an unnamed source, and neither company has confirmed it in the dataset, so I treat it as pipeline. The firmer Brookfield move came earlier: it expanded its Bloom Energy partnership fivefold, to $25 billion, to build and finance fast on-site power. [Read here]

5. OpenAI Announces $20 Billion Campus in Effingham County, Georgia

OpenAI announced a $20 billion data center development in Effingham County, Georgia. That is planned investment. About a month later, a 3.2 GW power deal for the site was approved. Acting as its own developer gives OpenAI more control over timing and design, along with responsibility for site, power and community relations. [Read here]

6. Meta and BlackRock Form El Paso Data Center Venture

Meta and BlackRock formed a venture to own a 1 GW campus under construction in El Paso, with about $14 billion in total development costs. Funds managed by BlackRock own 80% and put in about $4.9 billion of cash, partly funded by a $12.5 billion debt raise priced at yields of about 7.5%. Meta keeps 20%, leases the whole campus for an initial four years with options out to 20, and manages construction.

Meta also gives residual value guarantees with a threshold of about $13 billion: under certain conditions in the first 16 years, it covers any shortfall between the campus’s value and that level. It is a useful model for you to study: ownership moves to institutional capital, while the user keeps a share of the value risk. [Read here]

7. Hut 8 Fully Leases 1 GW Beacon Point Campus at $19.6 Billion

Hut 8 signed a second 352 MW IT lease, which fully leased its 1 GW Beacon Point campus and lifted campus-level base-term contract value to $19.6 billion. That is lease value over the base term. The source headline does not name the tenant. Bitdeer signed a $4.7 billion, 16-year lease in Norway, and CleanSpark a 20-year lease in Georgia. Mining sites bring power and land, and the payoff comes from converting them to what AI customers need, at a cost and on a timetable the lease supports. [Read here]

8. ByteDance Begins $38.44 Billion Data Center in Brazil

ByteDance began building a $38.44 billion data center in Brazil, the largest single Latin American commitment in the dataset. That is planned investment. Two more signals followed. Brazil signed its Redata tax regime into law, and Alibaba Cloud opened its first South American region there. Brazil is drawing global platforms at a scale that gives it a real say in how the next wave of capacity is built. [Read here]

9. SK Telecom Sells 49% of SK Horizon to KKR and IMM for $2.2 Billion

SK Telecom launched SK Horizon, its AI data center company, and agreed to sell 49% to KKR and a Korean private equity consortium including IMM for 2. 2billion.That is the price of the stake. Earlier in the quarter, SKToutlinedUS91.5 billion of new Korean data center investment and a plan for 15 GW of AI capacity. Bringing in a large minority partner before construction puts outside equity in at the start of the program. [Read here]

10. Google Commits €13 Billion to AI Infrastructure in Finland

Google committed €13 billion over two years to AI infrastructure in Finland. It had company. Cerebras and Compute Nordic announced a 165 MW site in Mikkeli, TikTok expanded its Finnish capacity, and Microsoft started pre-construction in Vaasa. Finland’s power and climate are drawing a cluster of global builders at once.[Read here]

11. Mitsubishi Estate Enters Data Centers With 2.5 GW and $9.3 Billion

Mitsubishi Estate is entering the data center market with 2,500 MW of planned capacity and a $9.3 billion planned investment. That is a first move at gigawatt scale. Goodman also signed the first hyperscale lease at its 1 GW Tokyo campus, and Keppel DC REIT agreed to buy stakes in two Tokyo data centers for $1.05 billion. Japan’s big property owners are moving into data centers as a business line. [Read here]

12. Firmus Passes 900 MW Contracted and Adds OpenAI as Anchor Customer

Firmus passed 900 MW of contracted capacity, signed OpenAI as an anchor customer and expanded into Malaysia. A month earlier it had closed a fully subscribed $2 billion strategic equity raise. By mid-September it was courting investors for an ASX listing. Contracts, equity and a listing process, all inside one quarter. [Read here]

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

Capital is moving toward AI chips. In one week, three large chip financings surfaced. Apollo is at the table on all three, alongside names like Goldman Sachs, Blackstone and PIMCO.

Chips are moving from company spending into something lenders will fund like infrastructure. And each deal solves the same problem with a different tool.

Oracle would rent its chips from an investor-owned company. Broadcom would stand behind its customer’s payments. SpaceX would borrow on its own name. All three still need the same thing at the end: a building with power.

What Was Reported, and What Was Not

All three deals are in talks. None is signed.

The Wall Street Journal reported that Oracle is in talks with Apollo and Goldman Sachs to fund a large chip purchase. Per that report, the likely structure is for investors to fund a separate company that buys the processors and leases them to Oracle over time. No amount was reported. Oracle is still talking to other partners.

The same report said Broadcom is arranging more than $50 billion so OpenAI can buy the custom AI chips the two companies are building together. Apollo and Blackstone are among the lenders approached. The talks are early. The size could change. The borrower has not been named.

The Financial Times reported that SpaceX is seeking about $40 billion to buy Nvidia processors. The package would combine about $10 billion of bank loans with $30 billion of investment-grade debt. Apollo is expected to lead and sell the debt on to other investors, and PIMCO is among the firms in talks. The FT’s sources said the deal should close in 2027. Bloomberg described the talks as early and said they could still end without a deal.

None of the companies has confirmed these terms. The filings below are the confirmed part.

The Structures Behind These Deals Already Work

These talks build on deals that have already closed.

In June 2026, Apollo led an initial $35 billion financing for Broadcom’s AI XPV Platform, alongside Blackstone and a group of banks. It supports Anthropic’s expansion of more than 1 GW of compute. The wider platform targets over 20 GW for frontier AI labs through 2028.

In the same month, SpaceX priced $25 billion in senior unsecured notes in five tranches, with maturities running from 2031 to 2056.

And NVIDIA has partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on platforms meant to mobilize over $500 billion of outside capital over time. That figure is a stated target, not money in place. It shows where the chip business is heading. Financing is becoming part of what gets sold with the chip.

We broke down how the GPU maker plans to channel investor money toward its own customers.

Oracle’s Structure Matches Chip Costs to Chip Revenue

The reported logic is timing. Oracle pays for hardware before it collects the cloud revenue that hardware produces. A lease spreads that cost into payments that run alongside the revenue.

Oracle’s quarterly filing for the period ending August 31, 2026 shows why that matters. It lists $288 billion of lease commitments that have not started yet, substantially all for data centers. They are expected to begin between fiscal 2027 and fiscal 2029 and run for 15 to 19 years. The same filing puts the estimated useful life of its servers and networking equipment at six years.

So the building and the chips run on different timelines. One building lease can span two or three generations of servers. A separate chip vehicle lets each hardware refresh be paid for on its own schedule, while the building lease runs on its own.

The terms that will shape this structure are not public yet. They include whether the lease can flex with demand, whether the processors are backed by a minimum resale value, and whether the vehicle can move hardware to another site. Those terms determine how Oracle and its investors share the work.

Broadcom’s Support Is What Lets Lenders Fund Custom Chips

Custom chips are harder to fund than standard ones. A lender has fewer obvious buyers if the hardware ever has to be sold. Broadcom’s answer is to back its customers.

Broadcom’s quarterly filing for the quarter ending August 2, 2026, explains how its existing platform works. A financial partner took over agreements to buy AI racks built on Broadcom-designed chips, along with the customer’s leases. Broadcom backs the customer’s lease payments over five-year terms.

If the customer defaults, Broadcom owes the gap between 85% of the amount still owed and what the racks bring in when sold. Broadcom can also take over the lease or arrange a sale. Its maximum exposure, once every rack is deployed, is about $29 billion. As of the filing, it had paid nothing, and the backstop’s fair value was not material.

Here is how that 85% line shares the load. The financial partner takes losses up to the first 15% of what is owed. Broadcom covers the shortfall below that, after whatever the hardware fetches. The chip designer becomes part of the financing, which gives lenders a strong party behind a new kind of asset.

The filing also describes a separate facility. Under certain conditions, the customer may issue Broadcom up to $42 billion of convertible notes to fund its lease payments. None had been issued as of August 2. The two figures should not be added together, because the filing does not show that both would ever be drawn at once.

OpenAI and Broadcom announced in October 2025 a plan to deploy 10 GW of custom accelerators starting in the second half of 2026. The new $50 billion package would help fund that plan. It has not been reported whether it uses the same 85% support.

A rival has already gone further, standing behind 15 years of rent at a landlord’s campus.

SpaceX Brings Its Whole Business to the Table

SpaceX’s reported package looks like a large company borrowing against its whole business, and SpaceX already has a record in the bond market. Its June notes carry coupons from 5.35% to 6.65%. Its second-quarter results reported $100 billion in cash, cash equivalents, and marketable securities as of June 30, along with 1.4 GW of compute capacity.

If the new debt is unsecured, lenders are relying on SpaceX’s launch, connectivity and AI businesses together. Repayment comes from the whole company, so the debt term and the chips' lifespan do not need to match. No one has reported whether any part of the $30 billion would be secured by the chips.

The Bankability Gap Now Sits at the Building

The Bankability Gap is the distance between announced capacity and capacity that can secure power, financing, permits, equipment, customers, and returns. These deals close much of that gap on the financing side. What remains sits at the building.

A financed chip earns its return once it runs. It needs firm utility power, finished substations, liquid cooling sized to the rack, fiber, and a team to commission it. The demand is there. The International Energy Agency projects data center electricity use rising from about 415 TWh in 2024 to about 945 TWh in 2030, with accelerated servers driving nearly half the increase.

So chip financing runs on four clocks. When the money is drawn. When the chips arrive. When the utility energizes the site. When the customer starts paying. When the four line up, the chip earns from the day the payments start.

That is good news for operators. A tenant with funded chips is a strong tenant. If you host one, tie the financing draw to energization in writing, and agree up front who handles late delivery, late power, cooling changes, or lower rack density than planned.

Regulators Are Building a Clearer Picture

The Federal Reserve Bank of Kansas City published research on October 5. It found that special purpose vehicles are the most highly leveraged entities in the build-out, with broad sponsor guarantees letting some finance 90 percent of their assets with debt. Sponsors support that debt through rent payments and residual value guarantees.

In March, the Financial Stability Oversight Council’s systemic risk committee recommended monitoring hyperscaler debt issuance and “debt-like or contingent obligations that might not be fully visible on balance sheets.” Treasury staff at the same meeting described bank exposure to private credit as contained, while noting the exact amount is unknown. The minutes of the Fed’s September meeting cited market commentary linking higher term premiums partly to heavy private debt issuance for AI infrastructure.

For policymakers, clear disclosure of leases and guarantees is what lets this kind of financing grow with confidence.

What to Look for in the Next Chip Deal

When the next deal is announced, read past the headline amount. You want five answers.

Who is the legal borrower?

Who stands behind the payments, and up to what cap?

What resale value the deal assumes for the hardware.

Whether the money is drawn on chip delivery or on power delivery.

And whether the debt term matches the chip's life.

If you invest, look at leases, guarantees and purchase commitments together, since that is how these deals are built. If you build, line up the financing draw with energization. If you regulate, ask for the backstops to be disclosed clearly.

The next filings will answer some of these. Watch whether the OpenAI package carries the same supplier support as the Anthropic platform. That single term will show how custom chips become a financeable asset.

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

The fastest new power plant on the largest U.S. grid may be one that is already built.

That is the lesson in the agreement Google and Constellation Energy announced on October 6.

The deal adds 890 MW of new nuclear capacity to PJM by upgrading existing reactors. It also gives 2,700 MW of existing nuclear supply a 15-year home. Together, the two contracts cover 3,590 MW.

So the deal does two jobs. It adds firm power on a timeline measured in years. It also keeps a large block of existing clean power on the grid as demand climbs.

What Google Actually Signed

The package has three parts, according to Constellation’s announcement and Google’s own post.

First, a 20-year power purchase agreement backs 890 MW of added output at 11 reactors in Illinois, Pennsylvania and New Jersey. Constellation says it will invest more than $4.3 billion in the work. The first uprate is expected by 2028. All 890 MW is due before the end of 2032.

Second, a separate 15-year supply agreement covers 2,700 MW from Constellation’s existing PJM fleet. Constellation says this keeps those plants on the grid.

Third, an expanded five-year technology alliance brings Google Cloud and Gemini Enterprise into Constellation’s operations, from uprate planning to plant output.

The deal also includes load shaping and demand response. Google can cut non-critical power use during high-stress grid events. The two companies also set up a framework to look at new generation, storage and demand response, with PJM first.

When you count new capacity, use 890 MW. That is about 25% of the 3,590 MW total. The other 2,700 MW is power that already flows, now under a long contract.

Headline figures often blur that line, which is why we wrote a guide to reading hyperscaler pledges number by number.

How Existing Reactors Add Power Faster

An uprate is an increase in the maximum licensed power of a reactor that is already operating. The Nuclear Regulatory Commission approves every increase. Utilities have used uprates since the 1970s.

Here, the upgrades cover turbines, steam generators and digital control systems, according to the American Nuclear Society. The plants already have land, operating licenses and grid connections. That saves much of the work that slows a new power plant.

Some steps remain. Each unit needs NRC approval. Each upgrade fits into planned outages at a working plant. And 11 units means 11 sets of engineering, permits and work crews, spread across four years.

Think of it like adding a floor to a house you already own. You still need a permit and a builder. But you skip buying land and running a new driveway.

Why PJM Needs the Megawatts

PJM needs this power. Its independent market monitor reports that the 2028/2029 capacity auction cleared 6,831.3 MW short of the reliability target. The shortfall was 208.7 MW in the two auctions before that.

The same monitor names data center load growth as “the primary reason” for those conditions. It estimates that adding existing and forecast data center load raised capacity market revenues by $29.37 billion across four auctions.

The monitor also cautions that data center forecasts carry real uncertainty. So grid planners face a hard task. They must add supply for demand whose timing is still moving.

Deals like this one help with that task. A buyer with a long contract funds new supply, and the new supply serves the whole grid.

The Rule That Decides Whether 890 MW Counts

PJM has proposed new rules for large loads, filed with FERC as Docket ER26-3515-000. Under the proposal, a large load that brings its own new capacity can avoid an interim service that may curtail uncovered loads during grid emergencies. PJM calls this Bring Your Own New Capacity.

As of October 9, PJM’s large load page marked the registry and the interim service as pending FERC approval.

The Google and Constellation deal aligns with that proposal's direction.

According to POWER magazine’s review of the filing, an uprate qualifies when it raises both installed capacity and capacity interconnection rights. Capacity that has already cleared a PJM auction generally does not qualify, which shapes how the 2,700 MW agreement fits.

The registration path for the 890 MW will become clear as the rules and the unit filings move forward.

Accreditation is the second step to watch. PJM rates each unit to decide how much accredited capacity it provides. That rating lines up against the 6,831 MW shortfall.

Texas is wrestling with the same question, and one regulatory filing there matters more than the headline pause.

The Terms That Will Shape the Economics

The commercial terms are private. The parties set the power price, the escalation formula, the treatment of clean energy credits, how capacity value is shared, and the timing provisions.

Those terms will decide the return on Constellation’s $4.3 billion and Google’s delivered cost of power. The 20-year term already tells you something useful. It aligns the revenue period with a long-lived capital program, making the work easier to finance.

The asset list will fill in over time. A Constellation spokesperson named six sites to the American Nuclear Society: Braidwood, Byron, LaSalle, and Quad Cities in Illinois; Limerick in Pennsylvania; and Salem in New Jersey. Those sites hold 12 reactors, and the deal covers 11. The unit list and the split of the 890 MW across units are the next disclosures to look for.

Salem has shared ownership. PSEG owns 57%, and Constellation owns 43%. How the co-owners share costs and added output is another detail to watch.

On cost to other customers, Google says the deal comes at no cost to other ratepayers. Constellation ties it to the federal Ratepayer Protection Pledge. Transmission work and capacity market results will show how that plays out across the system.

The demand response terms matter too. The number of megawatts Google can cut, for how long and on what notice, will show how much flexibility the deal adds in a grid emergency.

Uprate Deals Are Becoming A Pattern

I would read this deal as part of a series. One week earlier, on September 30, Amazon and Constellation signed a 20-year agreement at Calvert Cliffs in Maryland. It covers 690 MW, including a 190 MW uprate, backed by more than $3 billion of investment.

POWER also lists a 176 MW Walmart agreement at Dresden that includes 30 MW of uprates, and a proposed Google arrangement with Georgia Power for about 96 MW of uprates at Vogtle and Hatch, subject to state approval.

Google says its uprate and restart deals have enabled more than 1.5 GW of new nuclear capacity in the U.S. That sits beside a planned restart of the Duane Arnold plant in Iowa with NextEra and an advanced reactor program with Kairos Power. Each runs on its own timeline, and uprates come first.

The shape is clear. Large buyers sign long contracts, and plant owners use that revenue to add output at reactors already on the grid.

Leave a comment

The Lesson Travels Further Than the Reactors

The Bankability Gap is the distance between capacity that is announced and capacity that can secure power, financing, permits, equipment and customers. This deal closes much of that gap in a specific way. The 20-year contract answers the financing question. The existing sites answer most of the land and grid questions. Licensing and the PJM rule remain.

That is why the model travels in two forms.

In markets with nuclear fleets, like Canada, South Korea, Japan and parts of Europe, the asset itself may repeat. There may be output to add inside plants that already run, paid for by a long contract with a large buyer.

In most of the Gulf, Africa, Latin America and Southeast Asia, the contract is the part that travels. A creditworthy buyer signs a long offtake. That offtake pays for new supply on the shared grid. The buyer also agrees to cut load when the grid is under stress.

That second form matters most for emerging markets. The International Energy Agency reports that emerging and developing economies outside China hold about half the world’s internet users but less than 10% of global data center capacity. Reliable power is the condition it names first.

What to Watch Next

Watch FERC. Its ruling on PJM’s large load rules will decide how bilateral deals like this one count as new capacity.

Watch the NRC docket. Unit-level applications will show which reactors, how much added power each one gets, and when.

Watch 2028. The first uprate is the first real test of the schedule.

And when you read “new capacity” in the next power deal, check two things before you count it. Is the output added, or already flowing? And how will the grid operator accredit 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

Edition No. 72

THE WEEK’S THESIS: THE BUYER BRINGS THE POWER

The signed customer is now the asset that gets financed. Power, land and buildings matter. But this week, money moved fastest toward projects with a named buyer already locked in.

DayOne took its customer book to Nasdaq. CoreWeave committed to India as the only tenant on a new campus. Lenders are reportedly building vehicles that lend against chip buyers’ promises to pay. The tenant came first, and the capital followed.

The projects still waiting have everything except a buyer. TeraWulf has the power contract and no tenant. Applied Digital and atNorth have Finnish power and are still marketing it. Then there are the projects with a buyer and no permit, which is the constraint further down.

The structure behind DayOne’s filing is in The Mechanism below.

MARKET SIGNALS

Google pays for nuclear output it doesn’t own. Google and Constellation signed a 20-year deal to add 890MW through upgrades at 11 existing reactors. What counts as new supply on PJM’s books is still open.

Amazon builds beside a dedicated gas plant. AWS filed a preliminary plan for 36 data centers on 1,100 acres at the Homer City Energy Campus, about 50 miles east of Pittsburgh. A 4.5GW gas plant being built by Homer City Redevelopment will power the site. More than 1GW of that output goes back to the grid. The campus is sized to its power plant, and the grid gets the surplus.

Washington and Seoul pick a Texas power builder. The US Commerce Department and South Korea selected Related Companies, NextEra, and Lewis Energy Group to develop Project Star in Encinal, Texas. The $22.3 billion campus will provide 6.47GW of gas-fired power for an adjacent 5GW data center site. Korea and the US will jointly own the power campus under their trade deal. Sovereign money now funds the power block, and the data center follows it.

Onsite engines become a repeat purchase. Wärtsilä booked an order for 282MW of onsite power at a US data center built by an unnamed independent power producer. Fifteen gas engines arrive in 2028, with full operation in 2029. It is Wärtsilä’s seventh US data center order. Its US data center total now tops 3GW. Seven orders are a product line.

Finland’s old paper mills turn into AI campuses. Sesterce plans an AI campus worth more than €10 billion at the former Kaipola paper mill in Jämsä. The first phase is 200MW, rising to 600MW, and the company says an unnamed anchor customer is already committed. It also pledges 1MW of new renewable supply on the Finnish grid for every 1MW it uses. Old mills come with grid connections, and the connection is the scarce part.

CAPITAL WATCH

Broadcom, SpaceX, Oracle and the lenders. Broadcom is arranging more than $50 billion in debt to fund OpenAI’s custom chips, the Wall Street Journal reported. SpaceX is reportedly in talks to borrow ~$40 billion for Nvidia chips, split between ~$10 billion of bank loans and ~$30 billion of investment-grade debt. Oracle is reportedly weighing a vehicle that buys chips with investor money and leases them back to Oracle. Chips are leaving company balance sheets and moving into lease vehicles.

Firmus pulls its IPO. Nvidia-backed Firmus shelved its planned ~$5 billion Australian listing amid growing scrutiny of AI spending, Reuters reported. One Asia-Pacific operator stepped back from public markets in the same week another filed for New York. The next test is whether public buyers pay up for the one with signed tenants.

GLOBAL LENS: THIS WEEK BY REGION

North America: power contracts set the build order

TeraWulf signed an amended electric service agreement with Kentucky Power that doubles contracted capacity at its Muskie campus to 1GW. The second 500MW phase moves up from 2030 to 2029, pending state commission approval. The first 500MW starts ramping in 2028. No tenant has signed, so TeraWulf is selling a power date. In this market, the date is the product.

Brookwood’s city council voted to back the county development authority’s agreement with Applied Digital for its $3.2 billion Delta Forge 2 campus in Alabama. The vote followed weeks of local pushback over environmental impact and transparency. The Alabama Public Service Commission still has to approve the power agreement with Alabama Power. Local consent is done. The power contract is the last gate.

Europe: Finland’s power draws the developers

Applied Digital signed an agreement for access to up to 1GW of potential power in Finland, its first project outside the US. Initial power is expected from 2028. The company is already marketing the site to hyperscalers and says it structured the deal to limit early capital at risk. A US developer bought its first foreign option, and it bought Finnish power.

atNorth plans a new campus in Salo, west of Helsinki, at about €2 billion. atNorth has secured power for the first 75MW phase, with a path to 230MW through a Fingrid substation. Customer servers could add up to €10 billion more, by the company’s estimate. Securing first-phase power is what lets atNorth announce before naming a customer.

Asia-Pacific: tenants arrive before the buildings

CoreWeave plans to enter India with 240MW across three 80MW buildings at AdaniConneX’s Taloja campus in Navi Mumbai. It will be the sole tenant, with an option to double capacity. The first phase is expected in mid-2028 and will run NVIDIA Vera Rubin. A US AI cloud is committing to India at a single campus, at hyperscale size.

The New South Wales planning department approved Stockland’s A$630.4 million data center at Macquarie Park in Sydney. The 10-story building has 49MW of total power and 34.3MW of IT load. The consent bars its backup generators from selling power into the grid. Those generator and emissions caps are now the template other Sydney projects will face.

Share

CONSTRAINT OF THE WEEK: PERMISSION IS THE BOTTLENECK

Permission is now the scarcest input in this market. Finland’s Supervisory Agency ordered Google to suspend site work at planned data centers in Muhos and Kajaani. Regulators cite tree clearing, new roads, and drainage changes made before environmental assessments were done. The company said it fell short of its own standards. The sites stay frozen until those assessments are complete.

Raleigh’s city council voted 7-1 for a six-month moratorium on new data centers. Residents had asked for 32 months. Staff must write rules on noise, water and pollution by April 6, under a 2024 state law that bans downzoning. Six months sounds short. It still stops every new filing in the city.

In Melbourne, Monash council rejected Centuria’s 38MW, A$137 million expansion at Clayton. Its own planners had recommended approval. Councilors pointed to Victoria’s new plan, which proposes 150-meter buffers from homes and self-supplied renewable power. Those rules don’t start until 2027. The council applied them early anyway.

In India, the Andhra Pradesh High Court let Adani’s Vizag campus keep building for Google’s AI hub, at the developer’s own risk. A public interest case challenging state approvals is still open. The court said money spent now can’t be used to argue for relief later. The next bottleneck is legal risk on capital already in the ground, and lenders will price it.

THE MECHANISM

Two customers carry the book, and the listing funds the build

DayOne is a data center platform headquartered in Singapore. It was the international business of China’s GDS until it spun out in 2025, DCD reports. It runs campuses across Malaysia, Singapore, Indonesia, Thailand and Japan, with Finland and Spain in development. On Oct. 6 it filed a Form F-1 for a Nasdaq listing under the ticker DODC.

Revenue reached $484.3 million in 2025, up from $178.1 million a year earlier. The first half of 2026 alone brought $512.0 million. The company lost $367.1 million in 2025. It has 962MW in service and 1,328MW under construction.

Seven customers hold 2.3GW of bookings due by the end of 2028. Finishing them will cost about $11.4 billion, by the filing’s estimate. One customer paid 69.2% of first-half revenue. The filing describes it only as a global technology company with a leading short-form video platform. The second-largest paid 15.1%, and Malaysia supplied 87.0% of all revenue.

So two customers and one country carry a build already backed by $4.35 billion of long-term borrowings. Lenders funded the first gigawatt against those contracts. The listing asks public buyers to fund the next one on the same promise. The filing lists access to power as a core risk, and almost all revenue sits on Johor’s grid.

The share count, price, and proceeds are still blank. Reuters reported in February that DayOne aimed to raise ~$5 billion at a ~$20 billion valuation, per DCD. Customer contracts allow early termination, with penalties applying only after “a certain number of years.” The filing doesn’t say how many.

Leave a comment

NOTABLE TRANSACTIONS

Macquarie Capital: investment in Zerra, amount undisclosed

Macquarie Capital’s Infrastructure Principal team invested in Zerra, a hyperscale developer headquartered in Singapore. AGP, which founded Zerra, remains the majority owner. The amount wasn’t disclosed.

What makes it notable is the order of the steps. Macquarie first joined Zerra’s Western Downs Digital Park in Queensland as a consortium partner and adviser. Then it bought into the developer itself.

Western Downs plans up to 1.44GW across four stages and holds a long-term lease commitment from Anthropic. Zerra’s wider pipeline tops 2GW across Australia, Japan and India. This is Macquarie Capital Principal Investment’s fifth investment since April.

Watch the Western Downs development application, which is still under assessment. The tenant has committed, so the approval is the gate that binds.

Sabey Data Center Properties: Ares lifts its stake above $500 million

Ares Secondaries funds expanded their minority equity stake in Sabey Data Center Properties. Their total commitment now tops $500 million. Sabey is a joint venture between Sabey Corporation and National Real Estate Advisors. The follow-on builds on a stake first announced in July 2026. The tranche size wasn’t disclosed.

The buyer is what makes it notable. A secondaries fund is adding to a minority stake in a private platform, and the founding partners keep control.

Sabey owns and runs 275MW across US markets and expects its pipeline to reach about 737MW by 2033. National Real Estate Advisors manages money for the National Electrical Benefit Fund, a union pension trust. Pension money and secondaries money now share the same build risk.

Watch how the next 460MW gets funded. More partial stakes would show minority capital can carry a platform this size without a full sale.

AirTrunk: ~$1 billion more for its Tokyo TOK1 campus

AirTrunk, owned by Blackstone and CPP Investments, plans to invest another ~$1 billion in its TOK1 campus in Inzai, east of Tokyo. Reuters described it as an added investment, while DCD, citing Nikkei Asia, reported a ~$1 billion green loan led by SMBC, MUFG, and UOB.

What makes it notable is the repeat. AirTrunk closed a ¥191.6 billion green loan for the same campus in March, then the largest data center financing in Japan.

The new money goes to added capacity and liquid cooling for dense AI racks. AirTrunk puts its Japan capacity at 530MW today and targets more than 1GW. Its CEO said customers are looking to Japan for geopolitical reasons.

Watch for AirTrunk’s own release naming the lenders and terms. Two bank-led financings on one campus in seven months would mean Japan’s lenders treat AI capacity as core business.

THE FRAMEWORK: WHO PAYS FOR THE POWER

Each of this week’s power deals answers one question differently: who builds the supply, and who pays for it. The chart below lines them up.

THE QUESTION FOR NEXT WEEK

Will Google’s Finnish subsidiary satisfy regulators by Oct. 14? Tuike Finland must explain its site work and set out next steps by that date, or face enforcement. A clean answer puts Muhos and Kajaani back on the assessment timeline. A weak one tells every developer in Finland that early site work now carries a real cost.

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

Share Global Data Center Hub