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

Ben Evans does not call platform shifts lightly. For two decades, he has explained how the technology industry moves PCs to web, web to mobile, mobile to cloud and what gets built, captured, and destroyed at each step.

His May 2026 AI Eats The World update is the sharpest read yet on the current cycle.

The headline figure: the big four hyperscalers plan to spend roughly $700 billion on AI infrastructure in 2026, more than twice global telecoms capex and approaching the scale of global oil and gas.

The framing is sharper than before: Evans now asks whether AI infrastructure will mirror mobile networks a trillion-dollar capex industry where little equity value accrued to the operators.

Evans does not run an AI fund. He is not long the trade.

His job is pattern recognition across cycles.

That makes his framework more useful than most for investors underwriting digital infrastructure exposure now.

Here are the twelve takeaways that matter most for anyone allocating to data center, power, or AI infrastructure capacity.

The Capex Question

1. The big four will spend roughly $700 billion on AI infrastructure in 2026 — more than twice global telecom capex.

The 2026 capex guidance for Microsoft, Alphabet, Amazon, and Meta approaches $700 billion combined. For context, global telecom capex is roughly $300 billion annually, while oil and gas spending is around $1 trillion. AI infrastructure is now scaling alongside the world’s largest industries, with spending up roughly 7x since 2022 as tech leaders prioritize the risk of underinvesting over the possibility of excess capacity.

2. Capex-to-sales ratios are at all-time highs across the largest software companies.

Meta now spends about 55% of revenue on capex, Microsoft 54%, Alphabet 45%, and Amazon 25%. In just three years, three of the world’s largest software companies have transformed into infrastructure-heavy businesses at a scale that rivals or exceeds traditional telecom operators. As Evans puts it, they are “challenging financial gravity” these are no longer asset-light companies.

3. US data center construction has overtaken US office construction.

US data center construction spending now runs around $50 billion at a seasonally adjusted annual rate, matching or exceeding US office construction for the first time on record. Office construction has been declining since 2020. Data center construction has been rising on a near-vertical curve since 2022. The buildings of the American economy have shifted from desks to servers in less than four years. The figure excludes the compute inside the buildings.

4. The funding mechanism has migrated from cash flow to “structure.”

Evans’s own term, lifted from the deck: “Here comes the structure.” Meta has formed a $27 billion joint venture with Blue Owl Capital to help fund its Hyperion data center, while private capital is increasingly being drawn into what Evans calls the “$3 trillion AI building boom.” After three years funded largely by operating cash flow, the next phase is increasingly being financed through external balance sheets.

5. OpenAI is the test case: $1 trillion per year of aspirational infrastructure deployment.

OpenAI’s stated commitments imply 30 GW+ of new capacity, ranging from roughly $600 billion to $1.4 trillion by 2030. The internal aspiration referenced in the deck targets 1 GW per week, or about $1 trillion annualized at ~$20B per GW. OpenAI cannot fund this directly. Evans describes the setup as “Other People’s Balance Sheets, circular revenue and a lot of plate-spinning.” The plate-spinning is the warning the structure is the test.

The Bottleneck Map

6. Three bottlenecks now constitute a “pig in a python.”

Evans identifies three stacked constraints: GPU and memory supply, multi-year power bottlenecks (except China), and rising political and community resistance to data center siting. Together, these sharply reduce announced capex before it becomes operational capacity. The key geopolitical signal is that China is the only major economy able to add power at AI speed, meaning power-ready regions elsewhere may have structural pricing power not yet reflected in land or assets.

7. Microsoft’s CTO has stated directly that capacity has been “almost impossible to build fast enough since ChatGPT launched.”

Kevin Scott’s quote is the strongest first-person admission yet that the supply side of AI infrastructure is structurally behind demand. The constraint is not pricing. It is delivery. Three years into the cycle, the largest hyperscale operator in the world is on record that the build has not kept up with demand and may not for some time.

The Commodity Problem

8. Frontier LLMs are converging within 10 to 15 points on benchmarks.

OpenAI, Anthropic, Google, Meta, and Chinese labs now cluster within 10 to 15 points of each other on aggregate benchmark scores. A new leader emerges every few weeks. Evans’s verdict is direct: “For most general and consumer use, models are very similar, and crucially there are no network effects.” The second clause is the heavier one. The models can be excellent. They cannot be defensible at the model layer.

9. Anthropic has crossed OpenAI on monthly revenue.

Anthropic’s gross monthly revenue is around $3.5B, surpassing OpenAI’s roughly $2B net monthly revenue, with the crossover occurring in early 2026. Private valuations now sit near $850B for OpenAI and $900B for Anthropic a combined scale that exceeds the total value of all U.S. venture-backed IPOs from 1995–2000. The model-layer race is proving not to be winner-takes-all, and much of the capital assumes a structure that may be misread.

10. Commodity infrastructure historically does not capture value up the stack.

This is the central counter-thesis in the May 2026 deck: global mobile data traffic is up ~20x since 2010, while the MSCI global telco index has stayed largely flat. Despite massive capex, value accrued mainly to platforms like Meta, Apple, Google, and Amazon, not operators. Evans asks whether AI infrastructure will repeat this pattern. The scarcity thesis around power, land, and interconnection remains unproven, leaving telecom as the key bear case for data center investors.

11. The provisional thesis: “Models will just be infra. Innovation will move up the stack.”

Evans’s stated thesis is five clauses, each a direct claim: chat is a terrible user experience; general use needs apps; the labs cannot build all the apps; models are commodities with no network effects; models will end as infrastructure. The conclusion is that innovation, and therefore equity value capture, will move up the stack to the products and workflows that consume the infrastructure rather than the infrastructure itself.

The Deployment Gap

12. Consumer use is “a mile wide and an inch deep.”

ChatGPT has ~900M weekly active users, but only ~5% are paying. Usage remains relatively light, with under ~5,000 prompts per year for top users and ~1,200–1,500 for the top 20%. Daily penetration is ~14% of US adults, 26% of teens, and 22% of ages 14–29. Even so, this doesn’t yet justify a ~$700B capex buildout on monetized demand alone. Enterprise adoption shows a similar gap: Accenture books ~$2.2B per quarter in GenAI contracts, while most firms still report pilots far exceeding production use outside software development.

Strategic Implications for Digital Infrastructure

Three implications survive for anyone underwriting digital infrastructure exposure now.

First, the mobile network comparison is the key bear case the bull side hasn’t answered: history suggests commodity infrastructure rarely captures long-term value, meaning much of the equity upside may flow to application and service layers rather than data center infrastructure itself.

Underwriting should price both scenarios the scarce-input case and the commodity-infra case and discount accordingly.

Second, the funding shift is the cycle’s leading indicator of stress. Cash-flow-funded capex is a different asset class from joint-venture or debt-structured capex. The Meta-Blue Owl JV, the Oracle debt repricing, and the OpenAI plate-spinning structure are early signals of how the funding migrates.

The rate of that migration matters more than any single deal. Investors who model that migration explicitly will hold a different view of cycle duration than those who do not.

Third, power scarcity remains the key structural moat. Markets that can deliver reliable, dispatchable power within 24–36 months command a premium, especially where permitting is fast and grid capacity already exists. Most emerging markets, however, still lag on both fronts.

The ones that do are the addressable universe for the next phase of digital infrastructure capital formation.

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

If you’re not a subscriber, here’s what you’ve missed so far:

1. Q1 2026: The Quarter AI Infrastructure Became Energy-Constrained [How power, capital, and compute converged to redefine the global AI buildout.]

2. Where Is Capital Flowing in the Global AI Data Center Buildout? [February’s 2026 global data center deals reveal how capital, power, and platforms are determining where the next wave of AI compute capacity will scale.]

3. 19 key takeaways from Jensen Huang’s NVIDIA GTC 2026 keynote [Inside Jensen Huang’s GTC 2026 keynote: how AI factories, inference economics, and system design are reshaping data centers and shifting value to compute productivity.]

4. 10 Reports Shaping Global Data Center Strategy in Q1 2026 [A synthesis of the research defining AI infrastructure expansion, capital deployment dynamics, and the structural forces shaping the next phase of global data center growth.]

A structural transformation is underway in how AI infrastructure is financed, contracted, and scaled across global markets.

Capital is no longer anchored to standalone project finance or fragmented development pipelines.

It is consolidating into hyperscaler balance-sheet expansion and platforms bundling capital, power, and hardware into repeatable deployment systems.

AI infrastructure is now financed as secured capacity tied to energy, interconnection rights, and long-term offtake agreements.

This week made the shift explicit across capital and platform design.

In North America, financing is scaling through hyperscaler debt and platforms. KKR’s Helix launched with $10B+ backed by sovereigns, NVIDIA, and utilities. Amazon raised $17.5B in debt, while Oracle targets $70B in capex against a $638B backlog plus $40B more. Applied Digital, Switch, and Blue Owl expand secured-capacity financing.

In Europe, capital is concentrating around power-secured sites. Microsoft locked 470 acres in Finland for energy access, while Spain and the Nordics build gigawatt renewable pipelines. EdgeMode’s deal highlights grid stability as the key constraint.

In Asia-Pacific, growth is driven by megadeals and sovereign capital. DayOne raised $4.5B with sovereign backing, Blue Owl is exploring a $30B APAC exit, and Meta–Reliance is building in India, with expansion and monetization happening in parallel.

In South America, grid access is the bottleneck. Voltalia secured a 322MW interconnection in Brazil, turning power into financeable capacity, while Rio AI City reinforces that secured energy now drives development value.

The broader signal for investors is that financing, energy access, and compute supply are converging into a single infrastructure stack.

The key variable is no longer capital, but whether it’s tied to energized, contract-secured infrastructure that bypasses grid and permitting bottlenecks.

Winners will control interconnection rights, secured power, and platform-scale financing before constraints reprice global supply.

THIS WEEK BY REGION The week’s biggest moves — what happened and what it signals.

North America

Debt is becoming the primary fuel for US AI and data center expansion, and this week made the scale clear. Amazon secured a $17.5B loan backed by Citibank, alongside a C$14B Canadian bond. Oracle projected $70B in FY2027 capex and a $638B backlog while raising $40B in debt and equity.

KKR launched Helix Digital Infrastructure with $10B+ committed. Applied Digital, Switch, and Blue Owl also tapped debt markets for multi-billion-dollar financing tied to major data center buildouts.

For operators without an investment-grade balance sheet, the lesson is direct: secured project debt and committed offtake are now the price of entry, not a financing option.

Europe

Power access is driving European data center growth more than land. Microsoft acquired 470 acres in Finland for grid capacity and climate advantages. Ignis is building DayOne’s 300MW campus in Spain’s Aragon, while Grupo Fotones launched a €1.15B project in Murcia.

EdgeMode is also acquiring a stake in Ibersun to secure battery storage, underscoring Spain’s appeal due to renewable power and available grid connections.

For real estate and infrastructure investors, the European screen now starts with the power contract and works backward to the site.

Asia-Pacific

APAC capital flows split between fundraising and exits. DayOne raised a $4.5B Series C led by Coatue and Hillhouse with sovereign backing to expand across key Asian markets. Blue Owl is exploring a $30B sale of Stack Infrastructure assets across Australia, Japan, and Malaysia.

Meta and Reliance also partnered on an AI data center in India, highlighting hyperscaler expansion, while investors increasingly raise new equity as sponsors exit mature assets at peak valuations.

For investors, APAC is now a place to both deploy and exit in the same quarter, and the timing of each decision is the whole game.

South America

Brazil’s constraint and ambition became clear this week. Voltalia secured a 322MW grid connection for data centers, locking in scarce capacity ahead of construction.

The proposed Rio AI City aims to be Latin America’s largest campus, with Brazil’s renewable energy and rising hyperscaler demand making grid access the key bottleneck.

For investors weighing Latin American exposure, the secured interconnection is the milestone that converts a render into a fundable project.

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

KKR — Helix Digital Infrastructure, over $10 billion committed

Helix is a standing operating company backed by KKR, KIA, NVIDIA, and Vistra, investing across data centers, power, transmission, and fiber as a unified platform. It aligns capital, chips, and energy to streamline hyperscale buildouts.

For infrastructure funds, this sets a new competitive bar: the buyers hyperscalers prefer will be integrated platforms that deliver capital, silicon, and power together, not single-asset lenders.

Amazon — $17.5 billion senior unsecured delayed-draw term loan

Amazon secured a $17.5B delayed-draw facility led by Citibank, plus a C$14B bond, to fund capex. The hyperscaler is increasingly relying on bank debt and bonds as free cash flow turns negative against ~$200B annual spending.

For credit investors, hyperscaler paper is becoming a core allocation, and the spread compression on Amazon’s facility shows the market will fund this buildout cheaply for the strongest balance sheets.

Applied Digital — $1.59 billion senior secured notes

Applied Digital is raising $1.59B in secured notes for Polaris Forge 1, plus a 210MW lease at Delta Forge 2 and a $550M credit facility. The debt is backed by contracted capacity to enable large-scale development.

For neocloud and merchant operators, this is the executable path: secured notes against contracted capacity, not unsecured corporate debt, are how the next tier of the market funds gigawatt expansion.

THE WEEK IN THREE SIGNALS

The financing of AI infrastructure is now a dedicated asset class.

KKR’s Helix assembled a sovereign fund, NVIDIA, and Vistra into a standing >$10 billion vehicle built solely to underwrite hyperscaler infrastructure.

Investors should expect the integrated-platform model to become the structure hyperscalers prefer, which raises the capital and coordination bar for any single-asset competitor.

Hyperscalers have shifted from cash-funded to debt-funded capex.

Amazon’s $17.5 billion term loan and Oracle’s $40 billion raise against $70 billion of guided capex confirm the largest balance sheets are now borrowing to build.

For credit investors, hyperscaler debt is becoming a core fixed-income allocation, and the cheap spreads on this paper signal the market will keep funding the buildout until returns are questioned.

Co-located power is separating fundable projects from stalled ones.

Stark Power’s 5.6GW gas-paired US pipeline and Voltalia’s 322MW secured Brazilian grid connection both show that the power contract, not the land, now gates the project.

Operators and developers without a committed generation or interconnection path will watch capital route around them to the sites that solved power first.

Global Data Center Hub is published weekly. If you found this newsletter useful, share it with a colleague. Have a great week.

— Obinna

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

Join Global Data Center Hub

When most readers think about AI infrastructure, fiber does not make the headline.

Power, capital and chips do.

Fiber sits in the background a commodity assumed to be solved.

It is not.

Fiber is the constraint that decides whether AI infrastructure scales, where it scales, and who captures the returns when it does.

Power can be sourced from new generation. Capital can be raised from new pools.

Fiber cannot be conjured.

A subsea route takes three to five years to build.

A metro ring takes eighteen months to permit.

A cross-connect inside a carrier-neutral facility requires the right neighbors in the right building.

When fiber is missing, the data center cannot operate at scale, no matter how much capital is behind it or how much power is available.

This is the layer of AI infrastructure that gets discussed least and matters most.

This article teaches you to read it.

It is built around three questions, each one a different cut of the same problem:

Where is fiber the binding constraint?

Where does it generate returns?

How does it decide which markets and operators win?

Each question points to specific articles in the Global Data Center Hub Resources where the answer is developed in detail.

If you are new to this publication, this is the orientation.

Every subsequent piece you read here about fiber, connectivity, or data center positioning assumes the lens you are about to learn.

Where Is the Bottleneck That Investors Are Mispricing?

Investors and operators routinely audit power capacity, grid interconnection, and offtake credit before committing capital.

Fiber rarely gets the same scrutiny.

It is assumed to be present, assumed to be sufficient, assumed to be cheap to provision.

It is none of those things.

A subsea cable takes three to five years (sometimes longer) to plan, route, finance, and lay.

A metro fiber ring requires permits across dozens of municipalities, each with its own timeline.

Dark fiber inventory inside a metro is finite and often privately controlled, which means availability is a negotiating position, not a market price.

And subsea routes increasingly carry geopolitical risk.

The same cables that carry hyperscale traffic are now strategic assets that states monitor, contest, and occasionally sever.

When fiber is the binding constraint, capital does not arrive. Projects stall in due diligence, or they energize but underperform because the available connectivity cannot support the design load.

The mispricing is upstream. Investors assume fiber is solved, then discover during construction that it is not.

Reading this question correctly means treating fiber availability as a primary risk before treating capital as the primary variable.

It also means recognizing that fiber-related friction, when it can be navigated, becomes a moat the operators who solve it have something competitors cannot replicate quickly.

Articles in the Resources library that answer this question:

• What If the Real AI Bottleneck Isn’t Power, But Fiber? — the foundational reframe.

• Fiber in the Crossfire: Why Subsea Networks Are the Next Geopolitical Battlefield — the geopolitical layer of the constraint.

• Fiber & Connectivity Bottlenecks (Part 2): How to Evaluate Dark Fiber Availability Before Acquiring Data Center Land — the diligence framework for assessing it.
  
  

Where Fiber Creates the Moat

The question of where fiber generates returns is different from the question of where it constrains capacity.

The first is about value creation. The second is about risk.

Fiber generates returns at concentration points.

A carrier-neutral facility with hundreds of network operators on-net commands cross-connect economics that pure colocation cannot match.

A peering exchange with sufficient density becomes a settlement point that other infrastructure must connect through.

A fiber-to-data-center route that aggregates traffic from multiple hyperscalers across one corridor becomes a toll road that compounds in value as adjacent capacity builds out.

The economics are counterintuitive.

Fiber-to-DC projects are capital-intensive and often unprofitable on a single-route basis. They still attract billions because the unit economics are not the relevant frame.

The strategic position is.

The operator who controls the fiber that connects three hyperscale campuses to a regional internet exchange does not need to win on cost. The position itself is the return.

Competitive intensity does not compress these returns the way it compresses returns in commodity layers.

Fiber moats are local, physical, and slow to replicate.

A competitor cannot replicate a buried metro ring or a subsea landing station with capital alone. The license, the right-of-way, the existing customer base, and the time-to-build all conspire to protect incumbents.

Reading this question correctly means understanding that fiber is not just plumbing. It is a positional asset, and positional assets generate returns differently than capacity assets do.

Articles in the Resources library that answer this question:

• How Fiber, Peering, and Cross-Connects Create Moats in Data Centers — the architecture of the moat.

• Why Fiber-to-Data Center Projects Cost So Much And Still Attract Billions — the returns puzzle resolved.

• How Can Fiber Investors Thrive in a Market Full of Competition? — the competitive dynamics that protect incumbents.

Which Markets and Operators Win on Fiber

The first two questions were diagnostic. This one is decisive.

Fiber positioning determines which markets attract hyperscale capacity, which operators capture the returns inside those markets, and which jurisdictions get bypassed entirely.

The patterns are visible in three places.

Mature markets with hidden fragility, emerging markets making deliberate fiber bets, and operators using fiber positioning to rewrite regional playbooks.

In the United States, the fiber map looks robust on the surface hundreds of carriers, dense metro rings, established interconnection economics. Underneath, the network is brittle in ways that only become visible during stress.

A single subsea cut, a metro outage, or a permitting delay can ripple across hyperscale workloads because the redundancy that looks adequate on paper has not been tested under AI-era traffic loads.

The mature-market risk is the risk of false robustness.

In emerging markets, the dynamics invert. Markets like Brazil’s Fortaleza are gaining hyperscale attention because they sit at the intersection of multiple subsea cable landings, low-cost power, and underutilized capacity.

A fiber position that would be unremarkable in a saturated market becomes a strategic asset in an emerging one.

The operators who recognize this early move fast and capture position before institutional capital arrives.

The operator-level pattern is the most actionable.

Operators like Digital Realty are visibly repositioning around fiber-dense corridors in Malaysia, Indonesia, and other Southeast Asian markets not because the immediate returns justify the build, but because the fiber positioning will compound over five to ten years as adjacent capacity builds out and traffic concentration grows.

Reading this question correctly means watching fiber moves at the operator level. They predict where the next phase of capital deployment will land.

Articles in the Resources library that answer this question:

• The Hidden Risk Inside U.S. Fiber Networks — mature market fragility.

• Is Digital Realty Quietly Rewriting the Data Center Playbook in Malaysia? — operator-level positioning in an emerging market.

• Fortaleza’s Rise: The Strategic AI Infrastructure Blueprint Emerging from Brazil’s Edge — market-level positioning enabled by fiber.

How to Read From Here

These three questions are how Global Data Center Hub reads the fiber layer of AI infrastructure. Every article published here that touches on connectivity whether it discusses subsea cables, metro fiber, cross-connect economics, or regional positioning answers one of them, or layers them together.

New to the publication means you are still building the habit. The first time you read a fiber announcement and ask which of the three questions it answers, it will feel effortful. By the tenth time, you will be reading the constraint, the returns potential, and the competitive positioning simultaneously.

That is the analytical discipline separating the investors and operators who treat fiber as a strategic asset from those who treat it as commodity plumbing.

The Resources section is organized around these questions.

The Infrastructure Bottlenecks sub-section pulls together the constraint pieces. The Connectivity Economics sub-section drills into the returns. The Data Center Connectivity sub-section applies both to specific markets and operators.

Start with whichever speaks to your portfolio or focus. Each article is useful on its own. Read in the sequence your focus demands, and the publication becomes a curriculum.

The daily cadence is how you train your eye to track fiber positioning as it happens which is the only real-time signal this layer produces.

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

Cooling has become the variable that decides which data center assets hold value and which become liabilities.

The old underwriting paradigm treated cooling as an operational line item, a fixed percentage of facility energy that varied little across deals.

That paradigm is now wrong, and continuing to apply it means buying stranded assets at prices that assume they can host the workloads they cannot.

The Density Inflection

The driver is thermal density. Each generation of AI compute dissipates more heat per rack than the last, and the step-change is not incremental.

A standard enterprise rack five years ago drew 5–10 kilowatts. A modern AI training rack built on NVIDIA’s current architecture draws an order of magnitude more.

The NVIDIA B200 carries a thermal design power up to 1,200 watts per GPU in liquid-cooled configuration. The GB200 NVL72 rack system approaches 140 kilowatts.

That is roughly ten times the heat load of the enterprise rack that legacy facilities were engineered to absorb.

Air cannot move that heat. This is physics, not preference.

Air has low specific heat capacity, while AI rack heat is concentrated directly on the chips, not dispersed across room airspace.

Standard air-cooled infrastructure tops out near 40 kilowatts per rack before throttling and thermal failure begin.

Liquid is approximately four times more thermally conductive than air for the same volume.

Some manufacturers have already abandoned air-cooled variants of their newest hardware entirely.

The market has not finished absorbing what that means.

The Retrofit Trap

The industry has long assumed that cooling architecture is a retrofit decision that an air-cooled facility can be upgraded when demand arrives.

That assumption held when rack densities crept up slowly and air handled the load with margin to spare. It does not hold now.

Converting an existing air-cooled facility to liquid runs on the order of two to three million dollars per megawatt. That is a stranding cost, not an upgrade cost.

It compresses returns on every legacy asset in a portfolio, and it arrives precisely when the highest-value tenants are walking past those assets.

Here is the counterintuitive part. Liquid cooling does not carry a structural capital premium at high density.

Capital cost analysis of comparable air and liquid designs shows that at high rack densities, liquid-cooled builds come in at or below air-cooled cost.

Eliminating chillers, room air handlers, and containment offsets the cost of pumps, piping, and coolant distribution.

The premium is not in building liquid-ready. The premium is in not building it, then paying to convert later.

Consider a regional operator that built a high-quality air-cooled facility on a twenty-year asset life assumption. The building is sound, power is contracted, and the location is strong.

Then the tenant pipeline shifts to AI workloads requiring 100-kilowatt racks, and the facility cannot host them without a multi-million-dollar-per-megawatt conversion that was never in the model.

The asset did not fail. The underwriting did.

Time Is the Binding Cost

Supply-side friction makes this worse, and time is the binding cost.

Hardware refresh cycles run 12 to 18 months. Physical facilities are underwritten on twenty-to-thirty-year lives.

The mismatch is structural.

A facility designed for 30–50 kW racks cannot easily host the next generation without a cooling retrofit, and traditional real estate underwriting does not capture this obsolescence curve.

The constraint is not whether liquid cooling works.

It is whether the asset was designed before the demand it now has to serve.

Three Lenses on the Same Constraint

For independent operators, the binding constraint is qualification. Can the facility certify support for 100-kilowatt-plus racks with liquid cooling? If not, the highest-density tenant category is closed, and the question is no longer electrical capacity alone.

Hyperscale customers now ask for water usage effectiveness numbers and liquid-cooling certification as core leasing criteria. An operator who cannot answer is structurally excluded.

For private equity and infrastructure investors, the decisive variable is stranding risk priced against schedule. A legacy air-cooled asset acquired at a multiple that assumes AI-tenant optionality is mispriced if that optionality requires undisclosed retrofit capital.

You must underwrite the conversion cost explicitly, model the schedule risk of executing it in a live facility, and discount the assets that cannot make the transition at any reasonable cost.

For public equity, the divergence is already forming between operators positioned for liquid-cooled density and those carrying air-cooled portfolios into an AI-tenant market.

The dispersion will widen as efficiency standards tighten and as regulatory water exposure converts into operating-license events.

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 signal in France is not the €93 billion announced at the June 2026 Choose France summit.

The real signal is the 11 GW of data center capacity sitting in RTE’s connection queue, against a transmission network where unrestricted connection capacity is already close to exhaustion.

Mainstream coverage interpreted the summit as France securing a lead in the European AI infrastructure race.

A more precise reading is that France has now committed more capacity than its grid can energize on the timelines investors are underwriting.

The widening gap between announced megawatts and deliverable megawatts is where capital will be made or lost over the next 24 months.

Power Access Is the Moat, Not the Headline Capital

France’s structural advantage is real and often misread.

The advantage is not the capital flowing into the market, but the electricity underpinning it.

Nuclear power provides more than two-thirds of French generation, with low-carbon sources accounting for roughly 95% of the total mix.

Grid carbon intensity sits at about 42 gCO₂e/kWh, compared to a global average of 472.

Electricity in France costs roughly one-sixth of Germany’s, creating a rare combination of low operating cost and low Scope 2 emissions for always-on AI compute.

The market is pricing the capital commitments, but it is underpricing the underlying driver that makes those commitments rational: cheap, low-carbon baseload power at scale.

Baseload matters because data center demand is highly stable, varying by only about 5% intraday.

France’s flat nuclear output matches that load profile, a pairing no other FLAP market can consistently replicate.

The Grid Connection Is the Constraint the Capital Cannot Buy Past

Power generation is not power access. This is the read the announced-capital coverage misses.

France generates abundant low-carbon electricity, but the transmission system that connects a data center to that electricity is the binding constraint, and capital does not clear it.

RTE’s unrestricted connection capacity is nearly exhausted. Connection lead times extend in some cases to a decade.

From 1 August 2025 the regulatory framework changed: RTE gained authority to reduce allocated capacity where actual usage falls below agreed levels, which imposes new dispatch and utilization obligations on developers who previously treated an allocation as a fixed asset.

The result is a two- to five-year lag between announcement and energization that no amount of committed capital compresses.

The 11 GW queue is not a measure of demand strength. It is a measure of the bottleneck. Every megawatt in that queue is a megawatt not yet earning.

Power Secured at Commitment Separates the De-Risked From the Merchant

The deals that matter structurally are the ones that solved the grid problem before announcement, not after.

SoftBank’s Dunkirk program embeds EDF as a direct partner for the Bouchain site, with power purchase and connection negotiated at the investment-commitment stage.

The MGX-Bpifrance-Mistral-NVIDIA Paris campus is structured with RTE as a direct partner.

These projects are de-risked on the one variable that cannot be bought in the open market.

Set them against the precedent the report carries: Amazon abandoned a French self-build after failing to secure local support.

The differentiator across the entire pipeline is no longer capital, land, or demand.

It is whether the developer holds a secured connection.

Operators with confirmed RTE agreements hold a durable advantage that compounds as the queue lengthens.

Merchant developers holding headline commitments and no connection hold a liability dressed as a pipeline.

Investor Action

Private Capital. Underwriting French data center exposure should focus on grid connection, not project commitments. The key test is whether RTE or EDF capacity is secured and at what allocation, as the 1 August 2025 rule adds utilization risk even to allocated power.

Favor operators with confirmed connections like Digital Realty, Data4, and RTE-partnered campuses, and discount MW pipelines without verified grid access to deliverable capacity.

The cost of inaction is underwriting announced capacity at announced timelines and discovering the two-to-five-year energization lag after the capital is committed.

Public Markets. Public equity and credit investors holding the named actors Brookfield, SoftBank, the listed operators, EDF, Schneider Electric should benchmark exposure against deliverable capacity, not announced capacity.

The structural winners are the grid-adjacent names: EDF as the power counterparty embedded in the largest programs, Schneider Electric as the supply-chain partner SoftBank localized in Dunkirk to defeat the equipment bottleneck.

The cost of waiting is that the deliverable-versus-announced repricing is not yet in consensus estimates. It will be once the first announced timelines slip.

Operators. Developers and hyperscalers acting as buyers should sequence French entry around connection availability, which points to Hauts-de-France over Paris for new hyperscale.

The northern corridor combines brownfield industrial land, RTE pre-qualified connection zones, and government fast-track permitting. Paris remains the enterprise and colocation market but offers less new connection headroom.

The cost of inaction is competing for the same exhausted Paris-region capacity that is already forcing pre-leasing before construction completion.

The Verdict

France has won the capital race and has not yet won the energization race, and the second race is the one that determines returns.

The market inflection is the point at which announced timelines begin to slip against grid reality, which the connection queue and the RTE rule change place inside the next 24 months.

Investors who treat secured grid connection as the primary underwriting variable will price French assets correctly. Those who instead anchor on capital commitments risk overpaying for megawatts without a viable path to energization.

The open question is which announced projects hold real connections and which hold press releases, and the report does not answer it for every name.

That disaggregation is the work, and it is the work that separates deliverable return from stranded commitment.

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

Philippe Laffont does not chase the consensus. He prices it before it forms.

For six years, the founder of Coatue Management has used the firm’s Public Markets Update to do one thing.

Reframe the trade that everyone believes they already understand.

The May 2026 edition did not lead with a model, an application, or a call on where the Nasdaq closes the year.

It led with a divide.

Coatue argues every player in the AI build-out is on one of two sides.

Selling into the shortage or buying the scarce equipment.

They operate on different profit models and timelines, but the market is currently favoring only one side.

The message was direct.

The AI trade is not one trade.

It is two and most investors are positioned on only half of it.

Here are fifteen takeaways from Coatue’s May 2026 update that smart capital is already acting on.

The Central Framework

1. There are two ways to participate in AI, and they are not the same business.

Coatue’s framework splits the AI trade in two. On one side are the sellers of the shortage turbines, memory, storage, optics, and chips feeding insatiable demand. On the other are the buyers, the operators who deploy that hardware and monetize it over 5–7-year contracts. Both are essential, but they are fundamentally different trades.

2. Sellers book the profit immediately. Buyers wait.

When a seller ships a chip or turbine, it books the full margin immediately. A buyer, by contrast, deploys that same asset today but earns returns slowly over a multi-year contract. The seller gets instant reward; the buyer gets delayed payoff. That timing gap, more than the tech itself, is what drives how the market prices both sides.

3. The re-rating is a multiplication, not an addition.

Coatue’s point is that pricing power doesn’t lift earnings linearly it compounds through revenue growth, margin expansion, and multiple rerating. Micron’s margins rising from ~16% to ~70% shows part of it. Seagate shows the full effect, re-rating from ~50 to ~770 as scarcity returned. The move only looks irrational if you think additively instead of multiplicatively.

The Semiconductor and Compute Reordering

4. The agentic era breaks the GPU’s monopoly on the stack.

This is the most consequential call in the update. Training-era AI was GPU-bound, as GPUs generate tokens. Agentic AI still uses GPUs but increasingly relies on CPUs for retrieval, memory, tool use, and orchestration. Coatue estimates the GPU-to-CPU mix could invert from ~4:1 to ~1:4, implying a ~16x shift toward orchestration chips.

5. Inference and agents, not training, are the forward demand driver.

Coatue marks the shift to late 2025, when AI moved from answering to executing tasks with limited human involvement the “agentic Big Bang.” For infrastructure, demand becomes continuous rather than episodic: training runs end, but agents persist, loop, and run in parallel, creating sustained data center load.

6. Token demand is no longer exponential. It is hyper-exponential.

A chatbot uses tokens in a simple Q&A exchange. An agent, however, plans, reads, uses tools, and spawns sub-agents multiplying token use at every step. Coatue argues tokens remain profitable to produce, implying no clear ceiling on demand across consumer apps, developer tools, white-collar automation, and physical AI.

7. Memory and context are now infrastructure, not features.

An agent that forgets is not a digital employee. To be useful, agents must hold long context, recall past work, and retain preferences, which places a structural floor under demand for memory and storage. This is the same scarcity that lifted the memory and storage names. The agentic architecture does not merely sustain it; it deepens it.

Power, Capital, and the Bubble Question

8. Power is a first-class scarcity trade.

Coatue placed turbine makers in the same category as memory and chip suppliers: sellers of the shortage. That framing matters. Power equipment is not a peripheral beneficiary of the AI build-out. It is one of its scarcest inputs, carrying the same pricing power and the same immediate-margin profile as silicon.

9. The buyers of this cycle are the best-capitalized companies on earth.

The strongest rebuttal to the bubble thesis is balance sheet math. Hyperscalers generate nearly $1T in EBITDA, growing ~10% annually toward ~$6T in operating cash flow over five years. With low leverage and additional capacity from sovereign capital, private credit, and neoclouds, the AI buildout is broadly funded and largely self-financing.

What the Tape Is Telling You

10. AI is reordering winners and losers at a historic magnitude.

Coatue measured the performance gap between the best-performing and worst-performing of the world’s thousand largest companies. That spread has reached a width seen only twice before: the 2008 to 2009 crisis and the Covid shock. AI is not lifting all boats. It is a sorting event.

11. Even the Magnificent Seven is being reordered.

Since the late-October peak, with the exception of Alphabet, most of the mega-cap leaders underperformed the Nasdaq, and several were down on the year. The lesson for allocators is that mega-cap technology is not a monolith, and concentration is not the same as safety.

12. Sector dispersion confirms where capital is flowing.

On a relative basis, semiconductors and energy led, while software, internet, healthcare, financials, and consumer staples lagged. Capital is moving toward the physical layer of AI, the chips and the power, and away, at least for now, from the application layer that sits on top of it.

13. The macro backdrop is more supportive than the headlines suggest.

Markets absorbed tariffs, an oil shock, and fading rate-cut expectations, yet still hit all-time highs. Coatue’s view is that cuts were priced out not by rising inflation, but by stronger-than-expected growth where growth, not inflation, is the real driver of higher rates.

14. The oil shock is being priced as intense but short.

Oil moved from the high-60s to roughly 110 dollars. Yet the forward curve collapses that premium quickly, narrowing the spread toward 15 dollars within twelve months and lower still within twenty-four. The market is treating the energy shock as a spike, not a regime change, which matters for anyone modeling long-dated data center power costs.

When the Advantage Shifts

15. The power will eventually shift from the sellers to the buyers.

Coatue frames this as a sequencing call. Today, markets reward sellers of scarcity while shortages persist, but that advantage eventually shifts to early-invested buyers with contracted compute revenue. The seller trade is now; the buyer trade is positioning ahead of the rotation.

Strategic Implications for Digital Infrastructure

The seller-and-buyer divide maps almost perfectly onto the digital infrastructure value chain.

Equipment vendors are the sellers.

Data center owner-operators, colocation platforms, and neoclouds are the buyers.

They commit capital ahead of contracted, recurring revenue, exactly the deferred-reward profile Coatue described.

Three implications follow.

First, the CPU shift is a design and procurement signal, not just a stock call. If workloads shift toward CPUs, it changes rack design, power density, and capital planning.

Second, power is the binding constraint and the key scarcity, so energy procurement must come first in site selection, not last.

Third, the financing base (sovereign capital, private credit, neoclouds) is where independent sponsors and structured vehicles can compete, since hyperscalers won’t own every megawatt.

The asymmetric position is to be early to the buyers’ side while the market still pays only the sellers.

The Forward View

Coatue’s update should not be read as a bubble warning or a bubble denial.

It is a sequencing argument.

The capital to build is available, the demand is compounding faster than the language used to describe it, and the constraint is physical rather than financial.

The open question is one of timing.

When does the market stop paying exclusively for immediate scarcity profits and start paying for durable, contracted compute cash flows?

That rotation, from the sellers of the shortage to the owners of the compute, is the trade the next cycle will be decided on.

The investors who position for it early will not need to call the top in the sellers.

They will simply already be standing on the other side.

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 Financing Choice Is The Signal, Not The Size

The number that matters in Alphabet’s $80 billion raise is not $80 billion.

It is the decision to fund it with equity.

A company sitting on one of the largest cash positions and strongest balance sheets in the world elected to dilute its shareholders rather than lever up to finance its AI infrastructure buildout.

The market read the headline as a capex story and discounted the stock on dilution.

The structural read is different.

Alphabet is signaling that the AI infrastructure cycle is large enough and long enough that it wants permanent capital against it, not refinanceable debt.

That choice tells you how Alphabet underwrites the duration and the risk of this cycle, and it should reset how every other allocator underwrites the same exposure.

Equity Funding Prices The Cycle As Structural, Not Cyclical

The first signal sits in the instrument selection.

Alphabet is layering equity on top of an earlier debt program that included very long-dated paper.

Management has stated the objective is to preserve balance sheet flexibility while still funding capex at the top of the hyperscaler range.

Read the choice through the lens of what equity costs versus what debt costs.

Debt is cheaper when you believe the asset will throw off contracted cash flow on a predictable schedule.

Equity is the instrument you choose when you want to absorb timing risk, obsolescence risk, and demand risk without a fixed repayment obligation hanging over the asset.

Alphabet is telling the market it is not certain enough about the timing of AI monetization to want hard debt service against it, but it is certain enough about the direction to commit permanent capital.

That is a structural bet financed with structural capital.

Over the next 12 to 24 months, expect peers to be benchmarked publicly on the same question.

Are they funding AI capex with instruments that match the risk profile of the asset, or are they borrowing against revenue that has not arrived?

The Berkshire Placement Is Validation, Not Capital

The second signal is the Berkshire Hathaway tranche and its function is not the $10 billion.

Berkshire deploying capital into a large-cap technology name at negotiated terms, priced slightly below market, is a public validation of Alphabet’s decision to finance capex through dilution.

Berkshire built its initial position in the third quarter of 2025 and chose to deepen it through a private placement rather than buying in the open market.

A negotiated private placement with a long-term anchor signals conviction in the duration of the thesis, not a trade.

For the market, this lowers the perceived governance risk of the equity strategy.

When the most patient large-cap capital in the world underwrites your dilution at negotiated terms, it becomes materially harder for other shareholders to argue the equity raise is a sign of distress rather than ambition.

Expect this to compress the dilution discount over the next several quarters as the validation is priced in.

Power And Land Are The Real Constraint The Capital Is Chasing

The third signal is what the capital is actually for, and it is not GPUs.

The binding constraint in this cycle is power and the contiguous land to site it.

Leading markets are already strained.

The competitive frontier has shifted to who can secure large, contiguous tranches of power and land in markets that can still accommodate them.

Alphabet’s equity war chest is a pre-commitment instrument.

It funds long-lead power procurement, grid upgrades, and large campus development before the demand is fully contracted.

Most new capacity in leading markets is pre-leased or self-developed before construction begins.

A larger pool of permanent capital lets Alphabet crowd out smaller buyers in constrained nodes and pre-commit to power that second-tier players cannot match.

The signal for the next 24 months is that capital access is converting directly into power access, and power access is converting into market position.

The allocators who treat this as a compute race are mispricing it.

It is a power and land race financed by capital markets.

Investor Action

Private equity, infrastructure funds, and family offices should reprice the cost of competing with hyperscaler capital in constrained power markets. The opportunity is to avoid direct competition in strained primary nodes and instead focus on secondary markets, alternative regions, and the power-development layer before hyperscaler capital arrives.

The cost of inaction is being outbid on every contiguous power tranche in primary markets by buyers with a structurally lower cost of capital.

Public equity investors should benchmark every hyperscaler on the match between financing instrument and asset risk, not on capex magnitude alone.

The action is to diligence whether peers are funding AI capex with equity, debt, or revenue that has not yet materialized, and to treat instrument selection as a leading indicator of management’s own confidence in monetization timing.

The cost of inaction is mistaking a dilution-driven price dip for a deterioration in fundamentals and selling into a validated structural commitment.

Data center developers and operators should read this as confirmation that hyperscaler self-development and pre-leasing will tighten further in constrained markets.

The action is to lock power and land positions in markets the hyperscalers have not yet saturated, and to structure offtake conversations before the largest buyers pre-commit the available capacity.

The cost of inaction is watching contiguous power get absorbed by self-developing hyperscalers and being left to compete for the fragments.

The Verdict

Alphabet’s raise is the moment the AI infrastructure cycle was publicly repriced as a structural asset class rather than a capex line item.

The instrument choice, equity over debt, says management is underwriting a multi-year buildout with permanent capital because the duration and the demand timing do not fit a fixed repayment schedule.

The Berkshire placement says the most patient capital in the world agrees. The use of proceeds says the contest will be won on power and land, not chips.

Market inflection is the normalization of ultra-high, equity-funded capex as the price of staying in the top tier.

The open question is the one Alphabet itself cannot fully answer.

If the monetization of AI in search and cloud arrives slower than the build curve assumes, does permanent capital protect Alphabet or simply lock in the loss?

The allocators who answer that question first will position correctly.

The ones who wait for the answer to be obvious will pay for 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

Alphabet announced a plan to raise up to $80 billion in new equity during 2026 to fund artificial intelligence and data center expansion.

The capital is being raised across a private placement, underwritten public offerings, and an at-the-market program.

Proceeds are designated for general corporate purposes with explicit emphasis on capital expenditures for AI infrastructure, data centers, and global compute capacity.

The announcement was made on June 1, 2026.

The Three Funding Channels

The $80 billion is structured across three distinct issuance programs rather than a single block.

A $10 billion private placement goes to Berkshire Hathaway, split between Class A and Class C shares.

Approximately $30 billion comes through underwritten public offerings, divided between depositary shares tied to mandatory convertible preferred stock and common equity in Class A and Class C form.

The remaining $40 billion runs through an at-the-market program for Class A and Class C shares, expected to begin in the third quarter of 2026.

Goldman Sachs acts as placement agent on the Berkshire leg.

Goldman Sachs, JPMorgan, and Morgan Stanley serve as joint bookrunners on the underwritten public offerings.

Berkshire Hathaway Anchor Terms

Berkshire Hathaway’s $10 billion is structured as two equal tranches.

Five billion dollars is allocated to Class A shares at approximately $351.81 per share.

Five billion dollars is allocated to Class C shares at approximately $348.20 per share.

Both tranches are priced slightly below the prior closing market prices.

The investment is incremental to a stake Berkshire began building in Alphabet in the third quarter of 2025, which had grown to roughly $20 billion in market value before this transaction.

Capex Context And Capital Stack

Alphabet had already lifted its 2026 capital expenditure guidance by approximately $5 billion, to a range of roughly $180 billion to $190 billion, tied to AI infrastructure and custom silicon.

The company is choosing equity over incremental leverage, citing the objective of maintaining a healthy balance sheet while funding the capex program.

Its mix of common equity, mandatory convertible preferred stock, and at-the-market issuance spreads pricing and market-timing risk across multiple instruments rather than concentrating it in a single overnight block.

The at-the-market program gives Alphabet the option to match future issuance against the cadence of its AI infrastructure commitments through the second half of 2026.

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

If you’re not a subscriber, here’s what you’ve missed so far:

1. Q1 2026: The Quarter AI Infrastructure Became Energy-Constrained [How power, capital, and compute converged to redefine the global AI buildout.]

2. Where Is Capital Flowing in the Global AI Data Center Buildout? [February’s 2026 global data center deals reveal how capital, power, and platforms are determining where the next wave of AI compute capacity will scale.]

3. 19 key takeaways from Jensen Huang’s NVIDIA GTC 2026 keynote [Inside Jensen Huang’s GTC 2026 keynote: how AI factories, inference economics, and system design are reshaping data centers and shifting value to compute productivity.]

4. 10 Reports Shaping Global Data Center Strategy in Q1 2026 [A synthesis of the research defining AI infrastructure expansion, capital deployment dynamics, and the structural forces shaping the next phase of global data center growth.]

A structural transformation is underway in how AI infrastructure is financed, contracted, and scaled across global markets.

Capital is no longer anchored to greenfield development cycles or speculative buildout pipelines.

It is moving toward hyperscaler balance-sheet expansion and contract-backed platforms secured by power, leases, and existing generation.

AI infrastructure is now financed as capacity tied to energy availability and long-term offtake.

This week made that split visible across regions.

In North America, hyperscalers scaled balance sheets and power-secured builds in parallel: Alphabet sought $80B, Google and Intersect broke ground on Meitner in Texas, Crusoe signed a ~750MW deal, Switch is raising at $50B+, and Amazon filed a nine-building Indiana campus.

In Europe, platform financing scaled further: SoftBank committed 5GW in France and Brookfield added €10B. Polar raised an €800M Nordic bond, while atNorth, Arcem, and others advanced large Nordic sites.

In Asia-Pacific, hyperscaler-scale capital concentrated into single sites: AirTrunk committed $21B in Mumbai, Iren targeted an 800MW Australian campus, and BDx reportedly secured a 1.2GW power commitment in Indonesia.

In the Middle East and Africa, Gulf capital signaled intent to expand outward. Damac Digital outlined a 6GW expansion across four continents. Asprofin Bank unveiled a $12bn UAE proposal. Neither has reached financial close.

The broader signal for investors is that infrastructure sourcing is becoming the binding constraint in AI deployment.

The key variable is access to power-secured or energized infrastructure that bypasses grid and permitting delays.

Winners will control energized sites, grid capacity, and brownfield conversions before capital reprices supply.

THIS WEEK BY REGION The week’s biggest moves — what happened and what it signals.

North America

The constraint surfaced as Alphabet sought $80B in fresh capital and US build timelines slipped. The response was to fuse power and compute: Google and Intersect began the Meitner Energy Center in Texas, while Crusoe secured a ~750MW deal with Bergen Engines.

Switch is reportedly raising billions at a $50B+ valuation, while Amazon filed a nine-building Indiana campus. Meanwhile, a Virginia project was rejected and ~$670B of planned sites face storm-risk concerns.

For operators, the firms securing power and permits this week are buying schedule certainty their capital-rich competitors cannot.

Europe

Europe competed on site control and power, not capital availability. SoftBank committed 5GW in France and Brookfield added €10B, alongside Ardian and Verne’s Île-de-France campus.

The Nordics pushed a power-first model with Polar’s €800M bond, atNorth’s Norway site, BW Group’s 250MW build, and Arcem’s 500MW Finnish campus. In the UK, McLaren secured a £1B Corscale campus and ILI filed a 600MW project amid pushbacks.

For real estate and infrastructure investors, the Nordic land-and-power model is pulling ahead of UK markets where grid access and planning friction now set the ceiling.

Asia-Pacific

APAC capital is concentrating into large, power-anchored sites. AirTrunk committed $21B near Mumbai, a major bet on India’s hyperscale demand, while Australia saw momentum with Iren targeting an 800MW campus and Stockland filing a Melbourne site.

In Southeast Asia, BDx reportedly secured a 1.2GW power commitment in Indonesia, highlighting that power access not permits is the key constraint.

For data center investors, India and Indonesia are where AI infrastructure exposure is being priced now, ahead of the public-market narrative.

Middle East and Africa

Gulf capital signaled intent at scale this week, though both moves sit at the proposal stage. Damac Digital outlined a 6GW expansion across four continents, and Asprofin Bank unveiled a $12bn proposal for UAE data centers.

Neither is a closed financing, and the gap between announced ambition and committed capital remains the region’s defining question.

For investors tracking Gulf deployment, the discipline is to wait for financial close before treating these figures as deployable capital.

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

Polar — €800 million Nordic bond issue

Polar raised €800 million via a Nordic bond to fund European AI infrastructure, using public debt markets instead of private credit or hyperscaler-backed financing. It signals rising institutional demand and a broader funding base beyond banks and infrastructure funds.

For operators, a repeatable public-bond route lowers the marginal cost of capacity expansion and rewards the developers with the cleanest power and land story to underwrite against.

Alphabet — ~$80 billion AI capital raise

Alphabet is raising ~$80B in external capital for its AI buildout, despite having strong internal cash generation. It signals AI infrastructure capex is now so large that even hyperscalers are turning to external financing instead of self-funding.

For public-markets investors, the signal is that hyperscaler free cash flow is being absorbed by build-out faster than consensus models assume, and the financing overhang is now a sector-wide input.

Google / Intersect — Meitner Energy Center co-location

Google and Intersect are co-locating a data center with new generation in Texas, tying compute and power into a single asset. This removes interconnection queue delays and internalizes power supply instead of relying on the grid timeline.

For operators, co-located generation is becoming the premium execution model, and the developers who can pair compute with captive power will set the delivery benchmark the rest of the market is measured against.

THE WEEK IN THREE SIGNALS

Power is now bought before compute is built.

Crusoe’s ~750MW Bergen Engines agreement and the Google-Intersect Meitner co-location both put the power deal ahead of the data center build.

For operators and infrastructure investors, the interconnection queue and the generation plan are now the first underwriting question, and a financed gigawatt with no firm power is a stranded asset risk, not an asset.

Capital depth is no longer the differentiator.

Alphabet sought ~$80 billion the same week US build-out timelines slipped, proving money is abundant while delivery is scarce.

For public-markets and PE investors, the edge has moved from who can fund a project to who can actually energize and deliver it on schedule.

Emerging-market AI infrastructure is being priced ahead of the narrative.

AirTrunk’s $21 billion Mumbai commitment and BDx’s reported 1.2GW Indonesian power deal show private capital underwriting India and Southeast Asia at hyperscale before public equities reflect it.

For data center investors, the 6-to-18-month lead that private deal flow holds over public narratives is now widest in these two markets.

Global Data Center Hub is published weekly. If you found this newsletter useful, share it with a colleague. Have a great week.

— Obinna

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

Join Global Data Center Hub

This month’s stories reflect a market separating durable infrastructure from speculative capacity.

Hyperscaler concentration is reshaping returns across fiber and power, while retrofit limits and offtake economics are pushing capital toward purpose-built, contracted assets.

Winners will be defined by control of contracted demand, alignment of capital with compute economics, and the discipline to build only what offtake supports.

In case you missed any of the analysis, here is the full roundup of what we published this month.

Substack

Deep Dives

Detailed breakdowns on risks, strategic models, and long-term shifts.

• Capital Is Not the Bottleneck in AI Infrastructure. Here Is What Is. – [Read here]

• In AI Infrastructure, the Offtake Agreement Is the Asset – [Read here]

• The Retrofit Problem: Why Legacy Data Centers Cannot Serve AI Workloads – [Read here]

• Hyperscaler Dominance: The Real Threat to Fiber Investor Returns – [Read here]

Big Market Shifts

Major strategic moves by hyperscalers and what they signal.

• Microsoft And Alphabet’s $375B Q1: The AI Capex Industrial Phase – [Read here]

• Why NVIDIA Took A $2.1B Warrant In IREN, Not Just A Contract – [Read here]

• Why Meta’s $26B Leaseback Rewrote AI Infrastructure Financing – [Read here]

• Why Blackstone’s BXDC Is Credit Risk Priced As AI Growth – [Read here]

Infrastructure Fundamentals

Core constraints and capabilities shaping AI-ready compute.

• CoreWeave Is Not a Cloud Provider. It Is a Compute Factory. – [Read here]

• How Hyperscale Data Centers Move From Site Discovery to FID – [Read here]

Join Global Data Center Hub

LinkedIn

• NVIDIA Anchors IREN With $3.4B Contract And $2.1B Equity Warrant – [Read here]

• What KKR And Blackstone Just Signaled About AI Infrastructure Capital? – [Read here]

• How Hyperscale Data Centers Move From Site Discovery to FID – [Read here]

• Why Meta’s $26B Leaseback Rewrote AI Infrastructure Financing – [Read here]

• Demand Is Not the Constraint in Data Center Development. Power Delivery Is. – [Read here]

• Amazon And Meta’s $335B Q1: The Free Cash Flow Constraint – [Read here]

• The Grid Is the Gatekeeper: Why Most Data Center Pipelines Never Energize – [Read here]

• NVIDIA Posts $81.6 Billion Quarter And Splits Into Two Reporting Platforms – [Read here]

Twitter/X

• Italy’s data center boom isn’t about capital, it’s about power.

Terna’s 50 GW connection queue, bottlenecks in Milan and Rome, and the shift toward southern Italy show why grid access, energy partnerships, and interconnection capacity not capital determine which projects get built. This thread explains why power has become the real gatekeeper of Italy’s data center market. [Read here]

• AI infrastructure finance isn’t about data centers, it’s about contracts.

CoreWeave’s $7.5B debt raise backed by Microsoft, the growing importance of offtake agreements, and the shift toward counterparty-driven underwriting show why contract duration, credit quality, and revenue certainty not the physical asset determine financing outcomes. This thread explains why the offtake agreement is the asset, and the data center is the collateral. [Read here]

• AI infrastructure isn’t about data centers, it’s about financialization.

Blackstone’s $1.75B BXDC listing, the separation of development returns from stabilized yield, and the creation of a permanent institutional buyer for data center assets show why capital structure, transfer pipelines, and exit liquidity not asset ownership define the next phase of AI infrastructure investing. This thread explains how data centers are evolving into a fully financialized asset class. [Read here]

• CoreWeave didn't build a cloud. It built an AI compute factory.

CoreWeave’s evolution from crypto mining, Microsoft-backed offtake agreements, and the shift from diversified cloud models to specialized AI production systems show why committed demand, industrial economics, and power access not software-style scaling determine value creation. This thread explains why compute factories should be valued like infrastructure assets, not cloud providers. [Read here]

Thanks for catching up with this month’s roundup.

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

Cross-border fragmentation now sets the ceiling on returns for global fiber, and most capital still underwrites the asset as if borders were neutral.

The old paradigm treated subsea and terrestrial cable as a defensive utility: stable demand, light regulation, low correlation to geopolitics.

That assumption is now the single most expensive error in the asset class.

The Bottleneck Moved to the Statute Book

The shift is structural, not cyclical.

Subsea cable carries the overwhelming majority of intercontinental traffic and clears trillions of dollars in transactions daily, so the routing of that infrastructure has become an instrument of statecraft.

The US-China rivalry has cleaved the supplier market into competing blocs. Regulators on both sides delay, deny, and reroute.

The question for capital is no longer where demand is growing.

The question is which routes will still be legal, insurable, and commercially open when the asset energizes.

One Route, Many Regulators

Start with how the constraint actually works.

A fiber route is not one asset. It is a chain of permits, landing rights, vendor approvals, and maintenance agreements, each governed by a different jurisdiction.

Every border the route crosses introduces a new regime that can veto, delay, or reverse the others.

AI-grade demand intensifies this because the highest-value routes are also the most contested corridors.

Consider a transpacific system that secured demand and financing, then lost two years to permitting friction across a single intermediate jurisdiction and a national-security landing review.

The capital was committed, but the revenue never arrived.

The IRR was set by the slowest regulator on the route, not by the strength of the offtake.

Engineering Risk Was Never the Real Risk

The industry has long assumed that fiber risk is engineering risk: lay the cable, qualify the route, collect the toll.

That assumption held when the network was governed by commercial peering and a broadly open permitting environment.

It does not hold now because the binding variable has moved from the seabed to the statute book.

A perfectly engineered route through a contested exclusive economic zone is a stranded asset if the landing license is denied.

Underwriting that prices construction risk while ignoring jurisdictional risk is mispricing the asset.

Time Is the Cost That Kills the Return

The supply side compounds the problem, and time is the cost that matters more than capital.

Cross-border projects carry longer lead times and higher overruns than domestic builds because each government approval is a serial dependency.

Right-of-way acquisition alone drives a large share of project delay in some markets.

The repair layer is thinner than the market assumes: a constrained global cable-ship fleet means a single cut in a contested corridor can suspend revenue for months, not days.

Insurance is contracting at the same moment exposure is widening, with some political-risk providers now excluding politically motivated sabotage entirely.

The investor without a contingency plan is watching risk-adjusted returns fall below hurdle through no failure of the underlying demand thesis.

Three Lenses, One Binding Variable

For the independent operator, the binding constraint is vendor and landing-station provenance. The qualification problem is no longer technical. It is whether the chosen supplier and landing partner will clear the regulatory screens in every jurisdiction the route touches, because one disallowed vendor can void approvals across the entire system.

For private equity and infrastructure investors, the decisive variable is schedule risk priced into the IRR. A route through contested water now demands a higher hurdle rate, scenario-based downside models, and explicit assumptions on rerouting cost, insurance gaps, and policy-reversal probability.

The deal that models bandwidth and demand but omits disruption timelines is underwriting a different, more optimistic asset than the one it is buying.

For public equity, the dispersion is between positioned and exposed cable holders. Capital is flowing overwhelmingly into hyperscaler-owned systems. Hyperscalers now hold stakes in dozens of operational cables, and a single one of them had more than thirty projects in or approaching service late in 2025.

As that capacity is built for private use, it withdraws from the open market. Several transatlantic systems approach retirement while new transatlantic investment is almost entirely hyperscaler-led.

The result is a capacity premium for independent buyers and an accelerated-obsolescence risk for legacy systems on routes a hyperscaler cable is about to overbuild.

The market has not fully repriced the legacy holders that lack contractual protection against route substitution.

Five Rules That Separate Resilient Capital

Translate this into discipline. Five rules separate resilient capital from exposed capital.

First, re-order your site and route filters. Jurisdictional and geopolitical screening moves to the front of the funnel, ahead of demand modeling. A route that fails the legal screen never reaches the financial model.

Second, model schedule and availability risk explicitly. Price the delay distribution, the rerouting scenario, and the repair-vessel wait as line items, not footnotes. Time-to-service is the dominant driver of fiber IRR, so underwrite it directly.

Third, secure access ahead of the market through long-term structure. Long-dated IRU agreements with creditworthy counterparties, with CPI escalators and regulatory cost pass-throughs written in, convert compliance cost into a recoverable expense rather than a margin leak.

Fourth, build redundancy into the asset itself. Diverse landing points, mixed submarine-terrestrial paths, and a satellite continuity hedge for high-value corridors reduce single-corridor exposure. Route diversity is an investment imperative, not an engineering preference.

Fifth, treat government relationship capital as an underwritable moat. Sustained access to permitting authorities and security reviewers buys early warning on policy shifts and faster approval timelines. In fragmented markets, that intelligence is a direct input to allocation timing.

You Are Underwriting the Map, Not the Cable

You are no longer underwriting a cable. You are underwriting a route through a contested regulatory map.

The engineering is the easy part now, and the market still prices it as if it were the hard part.

The discipline that wins screens the map first. It prices jurisdiction, delay, and repair wait before demand, then secures the route through long-dated structure.

That is not risk mitigation. That is the return.

The investors who move now own the resilient routes at today’s basis. The ones who wait will pay 2030 prices for decisions that belonged in 2026.

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 Number Is Not The Signal

Amazon did not commit $33 billion to Southeast Asia on May 21.

It aggregated four pledges it had already made over three years into a single regional headline.

The market read the figure as a fresh capital event.

The figure is a packaging decision, and the decision to package now is the actual signal.

When a hyperscaler converts scattered country commitments into one bloc-level number at a policy summit, it is not announcing capital.

It is positioning for a regulatory contest it expects to intensify.

Consolidation Timing Reveals A Defensive Posture

The four pledges underlying the $33 billion were announced separately between 2021 and 2024.

Singapore’s incremental program runs to 2028, Thailand’s to 2037, Malaysia’s to 2038, and Indonesia’s across a 15-year horizon. None of these is new.

What is new is the decision to present them as a unified blueprint at Asia Tech x Singapore, a policy-and-regulation venue rather than a customer or product event.

Amazon chose the regulatory stage to restate old capital as regional scale.

This is the move of an incumbent defending position, not a challenger deploying fresh money.

The next 12 to 24 months will test whether competing hyperscalers respond with their own consolidated bloc numbers, which would confirm that committed-footprint scale has become the competitive currency in ASEAN.

The Sovereign-By-Design Frame Is The Real Product

Zapolsky’s remarks reframed sovereignty as programmatic control rather than provider nationality.

This is the substance beneath the capital headline.

As ASEAN governments tighten data residency rules and pursue sovereign AI strategies, the binding risk for a foreign hyperscaler is regulatory exclusion, not insufficient capital.

Amazon is underwriting against that risk by embedding itself in the ASEAN Digital Economy Framework Agreement process and positioning AWS as a co-architect of cross-border data rules rather than a compliance target.

The $33 billion number is the credential that buys a seat at that table.

Capital scale is being deployed as political access, and the access is the asset.

Investor Action

Private capital should recalibrate how it benchmarks hyperscaler commitments in emerging markets. A consolidated bloc announced at a policy summit is not equivalent to a single site deployment and treating it as fresh demand inflates absorption assumptions that underpin regional data center development models.

Public markets invesotrs should read the consolidation as a signal about competitive defense, not expansion. The structural question for anyone holding hyperscaler equity or evaluating sector exposure is whether committed-footprint scale now functions as a regulatory moat in markets pursuing data localization.

If it does, the hyperscaler with the deepest committed regional footprint compounds an advantage that is political as much as commercial.

Position for the second-order effect: competing providers will be pressured to match the bloc-scale framing, and the capital that follows will be defensive, not growth-driven.

Operators should calibrate against the power constraint the headline obscures. The $33 billion cannot deploy without grid capacity, and Thailand alone implies quadrupling operational capacity within three years against extended lead times for high-voltage transformers and switchgear.

The Verdict

The $33 billion headline marks the moment hyperscaler competition in ASEAN shifted from capital deployment to regulatory positioning.

The bloc that becomes the world’s fourth-largest economy will not be won by whoever spends the most.

It will be won by whoever embeds deepest in the rules that govern where data sits and which models run.

Amazon has declared that committed-footprint scale is now the entry ticket to that contest.

The open question is whether scale of footprint actually converts to regulatory influence, or whether ASEAN governments extract the capital and write the sovereignty rules on their own terms anyway.

That answer decides whether the next decade of hyperscaler capital in the region earns a moat or merely funds someone else’s digital sovereignty.

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 Investment Committee

In early 2024, an investment committee in Abu Dhabi sat with a question that did not fit neatly into the standard infrastructure mandate.

The deal on the table was a compute factory GPU-dense, liquid-cooled, purpose-built for AI training workloads, located in the United States.

The financial case was straightforward.

The offtake structure In AI Infrastructure, the Offtake Agreement Is the Asset identified was in place. The counterparty was credible.

Returns, measured against the sovereign fund’s cost of capital, were competitive with the best infrastructure deals of the prior cycle.

The strategic case ran deeper. The committee was not evaluating a financial asset in isolation.

It was evaluating a position in the infrastructure the intelligence economy runs on.

Financial returns and strategic priorities pointed in the same direction.

The committee approved the allocation. Then it approved three more.

How Sovereign Capital Has Engaged Infrastructure

Sovereign wealth funds built their infrastructure positions over three decades through a consistent playbook. Airports. Ports. Toll roads. Energy pipelines.

Telecommunications networks.

The thesis in each category was similar: regulated or contracted cash flows, long asset life, inflation linkage, and strategic relevance to the sovereigns managing them.

Gulf funds in particular Mubadala, PIF, and ADIA often held positions that served both financial and strategic objectives. An Abu Dhabi stake in a European airport was a financial asset and a diplomatic relationship.

A Saudi position in a global energy infrastructure platform was a financial return and a geopolitical tool. The two motivations were never entirely separable.

The compute factory replicates that structure at higher velocity.

The asset produces a strategic output AI training capacity that governments have identified as the infrastructure imperative of the next era.

Financial returns and strategic positioning once again move in the same direction.

The sovereign capital that recognized this alignment earliest is establishing positions before the asset class has priced the strategic premium.

What Changed in 2024

Gulf and Asian sovereign capital moved into AI infrastructure at scale in 2024 through three distinct mechanisms.

Direct equity in compute factory operators. MGX, the UAE’s AI-focused investment vehicle, committed capital to AI infrastructure platforms operating across the United States and Europe.

Mubadala built positions in the operator layer the companies building and running GPU-dense facilities at scale. PIF initiated AI infrastructure commitments that extended its existing technology investment program into purpose-built compute capacity.

Anchor positions in private credit deals. The leaseback structure Why Meta’s $26B Leaseback Rewrote AI Infrastructure Financing identified created a natural entry point for sovereign capital.

Long-duration, investment-grade cash flows backed by hyperscaler offtake commitments fit sovereign fund mandates precisely.

Several of the major 2024 AI infrastructure private credit transactions carried sovereign capital as anchor investors alongside commercial private credit firms.

Domestic infrastructure development commitments. Saudi Arabia, the UAE, and Singapore each announced domestic AI infrastructure investment programs compute capacity built on home soil, for sovereign AI development programs, under sovereign ownership.

The financial return on these commitments is secondary to the strategic output they produce.

The Anchor Capital Concept

Sovereign capital changes the compute factory financing stack in a way that goes beyond the capital it contributes directly.

When a sovereign wealth fund takes an anchor position in a compute factory deal equity or debt it changes the risk profile for every other capital provider in the structure.

The sovereign’s long investment horizon, its institutional credibility, and its tolerance for non-financial return dimensions compress the risk premium required by commercial capital alongside it.

This is what anchor capital means in industrial financing terms.

The anchor investor absorbs the first tranche of uncertainty regulatory, geopolitical, offtake quality that commercial capital prices at a premium.

The anchor’s presence signals that the uncertainty has been evaluated at the highest level of institutional sophistication and found acceptable. Commercial capital prices accordingly.

In AI Infrastructure, the Offtake Agreement Is the Asset established that counterparty creditworthiness is the primary credit variable in compute factory underwriting.

Sovereign anchor capital introduces a second variable the standard framework does not capture: the institutional validation of the capital stack itself.

A compute factory backed by Mubadala equity and Apollo debt is a different credit from the same facility backed by a regional family office and a domestic bank even when the offtake agreement is identical.

You should price that difference when evaluating any compute factory deal. Anchor capital is not just a source of funds.

It is a signal about institutional validation that changes every other investor’s underwriting.

The Complication

Sovereign capital’s non-financial return requirements create a complication the standard offtake framework does not resolve cleanly.

A sovereign fund investing in a compute factory in the United States is making a financial investment. The return is measured in dollars. The offtake agreement, the counterparty quality, and the asset’s design efficiency are the right variables.

A sovereign fund investing in a compute factory on its home soil is making an infrastructure policy decision. The return is partially financial and partially strategic domestic AI capacity, technology transfer, local employment, national security positioning. Standard discounted cash flow analysis does not capture those returns.

The credit analysis becomes more complex when the sovereign is simultaneously the investor and the offtake counterparty when the government commits to purchase the compute output of a facility it is also financing.

The primary credit variable from C2 collapses into a single balance sheet. The counterparty’s creditworthiness and the capital provider’s commitment are the same entity.

Underwriting that structure correctly requires a framework the compute factory asset class has not yet fully developed. It is the question Series 3 will address directly.

Three Positions on Sovereign Capital in the Compute Factory

For private credit investors co-investing alongside sovereign capital, the anchor dynamic warrants a lower required return than the standard model produces independently.

The sovereign’s presence signals institutional validation. Its long investment horizon reduces liquidity pressure on the overall structure. Its tolerance for non-financial return dimensions absorbs complications that would require commercial capital to be restructured.

A deal with credible sovereign anchor capital warrants a lower required return than the same deal without it and most commercial capital is not yet systematically pricing that adjustment.

For infrastructure equity investors evaluating operator-level positions, sovereign capital introduces both a structural advantage and an operational constraint.

The advantage: sovereign-backed operators access capital at a cost basis that independently financed operators cannot match, producing a durable edge in deal competition and asset development.

The constraint: sovereign capital’s non-financial return requirements domestic employment, technology transfer, local content obligations can introduce operational commitments that compress the equity return below what a purely commercial structure produces.

For compute factory operators seeking capital, sovereign anchor capital is the most strategically valuable capital in the current market precisely because it is the most patient, the most credible, and the most aligned with the long-duration nature of the infrastructure being built.

The operators who secure sovereign anchor capital in the current cycle hold a cost of capital advantage that compounds with every subsequent facility they develop.

The Question This Series Cannot Answer

The capital series has traced the compute factory’s financing stack from private credit through offtake structures to sovereign anchor capital. Each article has added one variable to the analytical framework this series is building.

What none of these articles has addressed is what happens when sovereign capital shifts from investing in compute factories built elsewhere to building compute factories at home when the sovereign is simultaneously the investor, the developer, the offtake counterparty, and the intended beneficiary of the AI capacity being produced.

That shift is underway. Saudi Arabia, the UAE, Singapore, and a growing list of sovereign actors are committing to domestic compute factory development at a scale that makes their US and European infrastructure investments look like positioning exercises.

Countries that missed the mainframe era and the cloud era do not intend to miss the compute factory era. The capital to build the domestic factory is sovereign capital.

The return the sovereign is optimizing for is not a financial return. It is the capacity to generate, store, process, and govern AI output on domestic infrastructure on sovereign terms.

That is where Series 3 begins.

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

Gavin Baker does not hedge. He builds a thesis, takes the position, and then explains the world in its terms.

Across a recent long-form interview on the Invest Like the Best podcast, Gavin Baker laid out the most complete public map yet of where the AI infrastructure trade actually sits: what is scarce, what is mispriced, and what could break it.

The conversation did not dwell on chatbots or model demonstrations.

It went straight to the physical layer: the watts, the wafers, the financing, and the silicon that intelligence runs on.

The message was clear: the AI buildout is not a bubble yet because two physical bottlenecks are constraining it, and value is accruing to owners of scarce inputs rather than application-layer builders.

One caveat governs everything that follows: Atreides is long most of the positions it discusses across AI, memory, and private silicon.

The framework is useful, but the conclusions reflect a vested view and are best treated as claims to be pressure-tested, not repeated.

Here are the fifteen takeaways that matter most, and where each deserves a second look.

The Bubble Question

1. The wafer bottleneck is the single mechanism preventing a bubble.

TSMC’s dominance in leading-edge wafers and disciplined capacity management help prevent bubble-like overbuild. Atreides estimates that meeting Nvidia’s demand could imply $2–3 trillion in annual GPU shipments, raising overbuild risk. The key signal is TSMC’s capacity decisions, assuming continued discipline and limited competition from Intel and Samsung.

2. This buildout is funded by cash flow, not debt, and the chips are fully utilized.

The strongest case against a bubble is structural: dot-com fiber was debt-funded and left ~99% of capacity unused, while today’s buildout is largely financed by hyperscaler cash flow with GPUs near full utilization. However, rising private credit and GPU-backed financing suggest a slow shift toward leverage, a dynamic that marked the turning point in the last cycle.

3. The binding constraint is now zoning and permitting, not power or chips.

Atreides expects capitalism to resolve the power bottleneck, with relief around 2027–2028. But the real constraint has shifted to permitting and land approvals bottlenecks that markets typically do not clear quickly, making that timeline uncertain.

The Underwriting Reframes

4. GPU useful life may stretch to ten to fifteen years.

A key underwriting claim is that splitting inference into phases could allow older GPUs like Hopper and Ampere to be repurposed beyond the typical 3–4 year cycle. If true, it would reduce financing costs, increase residual values, and strengthen GPU-backed deals. But it should be modeled both ways, since physical life isn’t the same as economic relevance and the claim aligns with a long position.

5. Orbital compute reframes terrestrial capacity risk.

The Atreides orbital compute concept uses SpaceX-style laser-linked, solar-powered orbital racks for inference while training stays on Earth. It could reduce terrestrial power demand, but its inference-only scope and unresolved deployment and repair limits constrain impact, leaving emerging-market builds more insulated than U.S. hyperscale demand.

Where the Value Accrues

6. Sellers of the shortage beat buyers, and so do owners of the scarce installed base.

Atreides argues that rewards go first to sellers of scarce equipment, while buyers wait for returns. It adds that in an agentic world, CPUs are more important than assumed because orchestration and tool calls are compute-heavy, and hyperscalers control the largest CPU fleets. Owning a scarce installed base becomes a key advantage.

7. The application layer has net-destroyed trillions in value.

Atreides argues that even AI-native firms like Cursor and Cognition show the application layer has destroyed net economic value, potentially in the trillions. Survival requires controlling the token path or building a defensible data moat early, with durable value shifting toward infrastructure and compute rather than applications.

8. Capital access is the real moat, so labs underprice their raises on purpose.

Atreides argues that steady investor returns create durable capital access. It cites SpaceX as the model, compounding by avoiding valuation peaks, and notes Anthropic and OpenAI raise below peak valuations for flexibility, with claims that Anthropic may match OpenAI’s revenue at about 80 percent lower burn. These figures should be treated cautiously.

9. Usage-based pricing is structurally bullish.

The key contribution is the pricing shift: from flat subscriptions to usage-based enterprise pricing, which ties revenue to consumption. Unlike subscription models that cap growth, pay-per-use scales with demand and capability. For infrastructure modeling, this means monetization tracks token usage rather than plateauing.

The Risks and the Tape

10. The biggest risk is the “bitter lesson” inverting.

The main risk, in Atreides’ view, is if algorithmic efficiency beats brute-force scaling, undermining the “more compute wins” thesis. If true, it would weaken the capex case at its core. The firm downplays a viral memory-optimization paper as a false alarm, though critics note Atreides is long memory, so the dismissal may reflect position bias rather than a neutral read.

11. Position on the Pareto frontier shifts fast and unpredictably.

Most model-layer value accrues to frontier tokens, where intelligence-per-cost leadership shifts quickly. Atreides argues Google lost cost advantage due to conservative TPU design choices. While specifics are unverified, the broader point stands: cost leadership at the frontier is unstable and can change within a single product cycle.

12. Cross-sectional valuations are internally contradictory.

Semicap equipment trades near ~40x next-quarter annualized earnings while memory sits in the mid-single digits, a divergence Atreides argues is inconsistent. Meanwhile, low-quality commodity players are outperforming, a pattern the firm reads as late-cycle behavior amplified by retail flows. The counterpoint is that the “memory is cheap” view reflects the firm’s own positioning.

13. Factor baskets are mispricing mis-categorized names.

When AI sub-sector correlations broke down, passive and quant indices misclassified companies using outdated labels. Atreides cites Astera Labs as misbucketed as a copper loser despite spanning both copper and optics, creating an inefficiency. But the example also reflects the firm’s positioning, since Astera Labs is reportedly a name Atreides has held since Series C.

The Competitive Map

14. The hyperscaler scorecard is diverging sharply.

Atreides sees the mega-caps as differentiated: Google lost its accelerator edge but keeps scale and data; Meta outperformed after going AI-first; Amazon is well-positioned through Trainium and robotics; and Microsoft pulled back on capex in early 2025, lost allocation, and is redirecting compute inward. These claims are fast-moving and should be verified.

15. The open-source prisoner’s dilemma will shape the model layer.

Atreides frames the model layer as a game-theory problem: closed frontier labs benefit from coordination, but any defection via open release brings share and advantage. The firm adds that open source isn’t free, as it still consumes compute and often involves revenue-sharing. In this context, Nvidia building a near-frontier model under the Nemotron line is a rational countermove, not a surprise.

Strategic Implications for Digital Infrastructure

Strip out the book-talking and three implications survive for anyone allocating to digital infrastructure.

First, GPU useful life is a key variable: if lifespans extend toward 10–15 years, financing terms, residual values, and returns all shift. Models should underwrite both extended and standard lifecycles and price the gap as risk.

Second, permitting speed should be a core site-selection metric. If approvals are now the binding constraint, jurisdictions that move faster gain a measurable structural edge, including in emerging markets.

Third, orbital and inference geography support an emerging-market thesis: inference disperses while training centralizes, and orbital compute doesn’t address sovereignty or latency. This suggests emerging-market inference capacity may be more insulated than assumed.

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

Google committed $15 billion to a new data center campus in New Florence, Montgomery County, Missouri, in an announcement dated May 19, 2026.

Ruth Porat, Alphabet’s President and Chief Investment Officer, led the announcement at a community event.

The campus will occupy more than 900 acres and support cloud and AI workloads.

Construction has begun, and the project will be delivered over multiple years.

Site And Scale

The New Florence site sits approximately 90 minutes west of St. Louis along Interstate 70.

This investment represents the largest single technology commitment in Missouri’s history.

The campus is built directly adjacent to Amazon Web Services’ Project Green, a $35 billion, 1,000-acre campus already under construction in the same county.

The two developments together create roughly 2,000 acres of combined hyperscaler campus infrastructure in a single rural county with a population of approximately 11,000.

Google is separately developing data center infrastructure in Kansas City as part of its Missouri strategy.

Power Procurement And Grid Commitments

Google contracted more than 1 gigawatt of new generation capacity in Missouri, with an additional 500 megawatts being developed through its partnership with Ameren Corporation.

Ameren Chairman, President and CEO Martin J. Lyons Jr. confirmed the campus is the largest economic development project in Ameren Missouri’s history.

The utility holds 2.2 gigawatts in aggregate binding contracts with multiple large data center users, equivalent to the full output of its largest coal plant, the Labadie facility.

Ameren is planning a 50 percent increase in generation capacity over the next four years.

Regulatory Framework And Cooling

Under Missouri Senate Bill 4, signed by Governor Mike Kehoe in 2025, Google is required to pay 100 percent of the power costs for its operations and any new infrastructure costs directly driven by those operations.

Google, Ameren, and Evergy brought the Capacity Commitment Framework to Missouri, binding large energy users to pay for their electricity and infrastructure needs and protecting residential ratepayers.

Google committed to air-cooling technology for the campus, limiting water consumption to domestic uses such as kitchens and bathrooms.

Jobs And Community Commitments

Google will fund the Construction Laborers and Contractors Joint Training Fund to train more than 2,300 laborers, including 1,500 apprentices, over the next two years.

Google estimates each direct data center role generates approximately nine additional local jobs.

The company announced a $20 million Energy Impact Fund to reduce household utility bills in Montgomery, Clay, Platte, and surrounding counties, with the Northeast Community Action Corporation as its first recipient.

Construction will unfold over multiple years, with permanent operational roles following completion.

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

If you’re not a subscriber, here’s what you’ve missed so far:

1. Q1 2026: The Quarter AI Infrastructure Became Energy-Constrained [How power, capital, and compute converged to redefine the global AI buildout.]

2. Where Is Capital Flowing in the Global AI Data Center Buildout? [February’s 2026 global data center deals reveal how capital, power, and platforms are determining where the next wave of AI compute capacity will scale.]

3. 19 key takeaways from Jensen Huang’s NVIDIA GTC 2026 keynote [Inside Jensen Huang’s GTC 2026 keynote: how AI factories, inference economics, and system design are reshaping data centers and shifting value to compute productivity.]

4. 10 Reports Shaping Global Data Center Strategy in Q1 2026 [A synthesis of the research defining AI infrastructure expansion, capital deployment dynamics, and the structural forces shaping the next phase of global data center growth.]

A structural transformation is underway in how AI infrastructure is financed, contracted, and scaled across global markets.

Capital is no longer anchored primarily to greenfield development cycles or speculative buildout pipelines.

It is now moving toward hyperscaler balance-sheet expansion and contract-backed independent platforms secured by power, leases, and existing infrastructure.

AI infrastructure is increasingly financed not as isolated assets, but as capacity tied to energy availability and long-term offtake agreements.

This week made that split visible across regions.

In North America, hyperscaler balance sheets and contract-backed infrastructure expanded in parallel. Google committed $15B and 1GW+ generation. Amazon outlined a ~$33B Southeast Asia plan. Rowan raised ~$550M for 300MW in Texas. Applied Digital hit 1GW contracted capacity, and TeraWulf acquired a 1+ GW campus.

In Europe, debt and platform financing structures continued to scale. Pure DC raised $2.7B, AtlasEdge secured €1.2B. HSCALE completed a second Milan campus with €2B+ committed. atNorth secured a 200MW grid connection in Sweden.

In Asia-Pacific, hyperscaler expansion and market reallocation moved in parallel. Amazon committed $33B+ to Southeast Asia. PDG is reportedly selling China data centers. Alceon and NEXTDC reflect rising institutional capital in the region.

In the Middle East and Africa, sovereign-linked capital continued to expand outward. Core42 secured $550M in HSBC trade finance for US and Europe AI capacity. A UAE–Oman–Italy consortium signed a 150MW Oman campus pre-financing deal.

At the same time, energy access is increasingly embedded in deployment outcomes. Google funded 1GW+ generation, Sweden secured 200MW, and TeraWulf bought energized sites power is the constraint.

The broader signal for investors is that infrastructure sourcing is becoming the binding constraint in AI deployment.

The key variable is no longer capital, but access to power-secured or energized infrastructure that bypasses grid and permitting delays.

Winners will control energized sites, grid capacity, and brownfield conversions before capital reprices supply.

THIS WEEK BY REGION The week’s biggest moves — what happened and what it signals.

North America

Balance-sheet capital and contract-backed financing moved in parallel across the US this week. Google committed $15 billion to Missouri and will fund over 1GW of new generation directly, reducing grid risk.

Rowan Digital Infrastructure closed ~$550 million financing for a 300MW Texas campus under a ~$4.4 billion framework, Applied Digital surpassed 1GW contracted capacity via a hyperscaler lease, and TeraWulf acquired a 1+ GW Kentucky HPC campus for AI conversion.

For operators, the signal is clear: investment-grade leases and secured power now unlock debt financing.

Europe

European capacity growth ran on debt this week, not equity. Pure DC secured $2.7 billion to accelerate AI infrastructure across Europe and the Middle East, and AtlasEdge raised €1.2 billion in loans for its European portfolio.

HSCALE closed its second Milan hyperscale campus with over €2 billion committed, concentrating Italian capacity in one operator. In Sweden, atNorth secured a 200 MW grid connection via an E.ON unit, locking power ahead of buildout.

For real estate and infrastructure investors in Europe, the constraint to underwrite is the grid connection date, so price the interconnection queue before the land.

Asia-Pacific

The region split between hyperscaler expansion and a contrarian exit. Amazon will invest over $33 billion in cloud and AI infrastructure across Southeast Asia through 2039, anchoring it as a key growth market.

Against that build-out, PDG is reportedly looking to sell its China data centers, a rare retreat from a market most operators are entering. The divergence is the signal.

For tech and infrastructure investors, capital is flowing into ASEAN and questioning China at the same moment, so treat the two as separate risk books rather than one regional bet.

Middle East and Africa

Gulf capital expanded as Core42 secured $550 million in structured trade finance from HSBC to deploy AI cloud and compute capacity in the US and Europe via non-dilutive funding tied to the US–UAE AI Acceleration Partnership.

Meanwhile, a five-party consortium from Oman, the UAE, and Italy signed a pre-financing agreement for a 150MW hyperscale campus in Oman, signaling growing cross-border interest in Gulf digital infrastructure hubs.

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

Rowan Digital Infrastructure: ~$550M green construction financing, 300MW Cinco, Texas

The facility, led by TD and Mizuho with ICBC, Huntington, and Lloyds, is backed by a 300MW contracted hyperscale project. It underscores sustainability-linked debt as the default financing route, with 40+ lenders competing for secured capacity.

For operators, the mechanism is a green framework that lowers the cost of capital and widens the lender pool, so the path to scale runs through a rated framework and a signed offtake, not an equity round.

Applied Digital: 1+ GW contracted on investment-grade hyperscaler lease

The milestone is a single investment-grade hyperscaler lease at Polaris Forge 3 that converts pipeline into bankable revenue, showing how lease quality now drives AI infrastructure valuations and can re-rate entire platforms.

For investors, the mechanism is counterparty credit substitution, where the tenant’s balance sheet underwrites the operator’s debt, so the lease covenant matters more than the developer’s own rating.

TeraWulf: acquisition of 1+ GW Eastern Kentucky HPC campus

TeraWulf acquired a gigawatt scale crypto mining campus and is repurposing its power for high performance compute. It shows that secured power, not land, is the constraint and mining sites offer the fastest path to energized AI capacity.

For operators and investors, the mechanism is power-asset arbitrage, where buying an already-energized site skips the multi-year interconnection queue, so the next wave of capacity will come from conversions before it comes from new builds.

THE WEEK IN THREE SIGNALS

Contract-backed financing is now its own asset class.

Rowan’s ~$550 million green close and Applied Digital’s investment-grade lease show that independent operators fund capacity against signed offtake and rated frameworks rather than equity.

For lenders and PE investors, the underwriting object has shifted from the operator’s balance sheet to the tenant covenant and the power contract behind it.

Power, not capital, sets the build ceiling.

Google funded 1+ GW of its own generation in Missouri, atNorth locked a 200 MW connection in Sweden, and TeraWulf bought an energized Kentucky site outright.

For developers and real estate investors, the interconnection date is now the gating variable, so a secured power position commands a premium over land or design.

Capital is entering ASEAN while questioning China.

Amazon committed ~$33 billion across Southeast Asia through 2039 in the same week PDG reportedly moved to exit China.

For tech and infrastructure allocators, the two markets now carry opposite risk signals, so a single Asia thesis no longer holds.

Global Data Center Hub is published weekly. If you found this newsletter useful, share it with a colleague. Have a great week.

— Obinna

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

Join Global Data Center Hub

Market feasibility and demand analysis is the first analytical stage of any data center development.

It determines whether a proposed project sits in a market with real, deliverable demand before capital is committed.

This primer explains how the analysis works in practice, why it differs from traditional real estate feasibility, and why much announced capacity never becomes operating capacity.

What Market Feasibility Means

Market feasibility is the assessment of whether a market can support a proposed project at a scale and price that justify the investment.

In simple terms, it asks whether: (i) demand exists, (ii) that demand can be served at the location, and (iii) the economics work over the life of the asset.

For most asset types the central variable is tenant demand, with supporting infrastructure assumed available.

Data center feasibility reorders the variables: demand is studied, but the limiting factor is rarely demand.

It is the supply of electricity to the site.

Why Data Center Feasibility Differs from Traditional Real Estate Feasibility

A data center is best understood as a power-delivery and compute-hosting facility rather than a conventional building.

Its value depends on bringing large amounts of electricity to a specific point on a specific timeline. Three differences follow.

The first is power intensity.

Modern AI racks draw 100 to 140 kilowatts, compared to 5 to 15 kilowatts for older systems. Since a megawatt equals 1,000 kilowatts, large campuses can consume more power than small cities.

As a result, feasibility must assess the full energy supply chain, not just the real estate.

The second is timeline mismatch.

A data center shell can be built in roughly 18 to 24 months, but securing grid power can take four to seven years (or more) in constrained markets.

In traditional real estate, the utility connection is rarely the gating item; in data centers it frequently is.

The third is cost intensity.

Standard hyperscale construction runs roughly $10 to $12 million per megawatt, and AI-optimized facilities can reach $20 million or more.

The capital at risk per megawatt is high, which raises the cost of a feasibility error.

The result is a shift in the central question.

Traditional feasibility asks whether tenants will come; data center feasibility asks whether the site can be powered on the tenant’s schedule.

Reading Demand Signals

Demand analysis begins with the signals that indicate how much capacity the market wants. Three are commonly used.

Announced megawatts are the planned capacity that developers, operators, and utilities publish through press releases, planning applications, and public filings.

This is the broadest and least reliable signal, because an announcement commits no one to build.

Hyperscaler capital expenditure refers to large-scale infrastructure spending by major cloud and AI platforms.

For 2026, total capex across the four largest US hyperscalers is projected at roughly $600–$710 billion, about 75% tied to AI infrastructure, with Amazon alone guiding around $200 billion.

Letters of intent are the third signal.

A letter of intent is a non-binding or partly binding statement that a tenant intends to lease capacity.

It names a counterparty, making it stronger than an announcement, but does not guarantee a signed lease.

Taken together, these signals point to strong, well-funded demand.

For example, North American vacancy is near a record low (~1.4%), and ~92% of under-construction capacity is already committed.

However, none of these metrics alone guarantees that a specific project will ultimately be delivered.

Bankable Demand: What Counts as Real

Bankable demand is demand a lender or investor can rely on when financing a project: the subset of total demand supported by binding, credit-backed commitments.

Separating it from announced demand is the core task of data center feasibility.

Three features raise demand toward bankability.

Pre-leasing is a lease signed before completion, often 24 to 36 months ahead of delivery, which converts a forecast into a contractual obligation.

Credit quality is the tenant’s financial strength; leases to investment-grade hyperscalers support higher leverage.

Contractual commitment covers take-or-pay terms, long durations, and limited termination rights, which make the income stream more reliable.

When these features are present, lenders can extend capital on favorable terms.

Senior debt can reach up to 80-90 percent of project cost, but typically only where demand is contractually secured and power is confirmed.

A pre-lease without confirmed power remains exposed to the largest delivery risk in the sector.

Power Deliverability: The Binding Constraint

Power deliverability is the assessment of whether electricity can reach the site in the quantity and on the timeline the project requires.

It is the variable that most often separates feasible projects from infeasible ones.

The starting point is the interconnection queue, the waiting list a project joins to connect a new load or generator to the grid. Federal guidelines target 8 to 11 months, but actual averages are longer.

The US national average is around 25 months. PJM, covering much of the eastern United States, averages about 40 months; ERCOT, the Texas grid, averages about 20. In the zones where data centers concentrate, waits of 36 to 48 months are common.

Completion rates compound the delay. Of projects that entered the ERCOT queue by 2020, about 40 percent reached an interconnection agreement or operation; in PJM the figure was about 24 percent. Most applications are withdrawn.

Grid constraints are often physical, not financial. In Northern Virginia, the utility cannot guarantee power for new data centers due to transmission limits, even when generation exists.

Local zoning rules, such as Loudoun County’s special exception requirement, further restrict development.

Because grid connections are slow, the industry is shifting to alternatives. About 30% of planned US capacity now uses behind-the-meter generation. For example, Meta’s $1.6B gas turbine deal for its Ohio campus helped drive a 27% rise in turbine orders in 2025.

A Simple Conversion Model: From Announced Demand to Energized Capacity

Demand analysis can be organized as a funnel with four stages, each narrower and more reliable than the one before.

Announced demand is capacity stated through press releases, planning applications, and letters of intent. It is the largest figure and the least certain; historically, 60 to 76 percent is never built.

Pre-leased demand is capacity covered by a signed lease, often on a building already under construction. It filters out speculative announcements but remains exposed to power-delivery risk.

Power-backed demand is pre-leased capacity that also holds a confirmed power arrangement, such as an executed interconnection agreement, a power purchase agreement, or a utility will-serve letter.

This is the bankable stage, where the main remaining risk is construction.

Energized capacity is capacity with power delivered and the facility operating, where lease income has commenced. This is the only fully de-risked stage.

The gap between the first and last stage is large. Applying the historical conversion rate of 24 to 40 percent to North America’s planned pipeline of about 31.6 gigawatts implies that roughly 7.6 to 12.6 gigawatts may reach operation by 2030.

The deliverable figure is a fraction of the announced figure, and a sound study reports it as the basis for decisions.

Practical Implications

For developers, the lesson is to secure the power path early. A project without a credible route to power carries the highest risk in the development cycle.

Controlling transmission-adjacent sites, advancing interconnection applications, and arranging on-site generation where appropriate move a project toward bankability.

For investors, the lesson is to underwrite demand by stage rather than by headline. A pipeline figure should be discounted to its power-backed and energized components before it informs a valuation.

Confirmed power, tenant credit, and contractual terms determine whether projected income is reliable.

For site selection, power availability is now the primary factor, ahead of land cost and often fiber. Markets with faster interconnection, available transmission capacity, and supportive zoning have a clear advantage, shifting development toward secondary markets and sites with on-site or dedicated generation.

Market feasibility and demand analysis ultimately produces a single output: an estimate of demand that can be reliably powered and delivered at a given site, within a defined timeline and cost. That estimate underpins every later stage of development.

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

This month's deal log highlights four transactions reshaping the global AI infrastructure landscape. Together, they show how power alignment, supported by capital access and platform execution, is determining where the next generation of AI capacity will be deployed.

Alberta — $10B Data Center and 1.4GW Power Project (North America)

The Alberta project, April’s highest-scoring global transaction, signals a structural shift in North American AI infrastructure planning.

At an estimated $10 billion, it is defined less by capital intensity and more by its pairing of a data center with a dedicated 1.4-gigawatt gas-fired power plant.

The defining feature is its power architecture. Instead of relying on grid capacity or interconnection queues, the project brings its own generation, sized to match expected compute demand.

Natural gas provides firm, dispatchable power, but also introduces a more material carbon footprint at this scale.

The capital stack is substantial but only partially disclosed. The filing states a ten-billion-dollar figure but does not identify an operator, anchor tenant, or financing structure. As a result, the transaction’s score is driven more by its integrated power-compute design than by proven demand fundamentals.

From a strategic standpoint, Alberta illustrates a broader shift:

Credible AI capacity is increasingly being sited where dedicated generation can be built, not where legacy data center ecosystems already exist.

The project rises to the top of the scoring because it directly addresses the binding constraint of the cycle:

Power availability.

The broader signal in North America is clear:

Competitive advantage is increasingly shifting toward projects capable of bundling generation and compute, even when much of that capacity remains in the proposed stage.

The broader North American signal is clear: advantage is shifting toward projects that bundle generation and compute, even when much of that capacity remains proposed part of the buildout now reshaping Canada’s data center market.

Mistral AI — $830M Paris Data Center Investment (Europe)

Mistral AI’s Paris investment is Europe’s highest-scoring transaction in April and the only European entry to reach a total score of seven. Unlike North America’s power-first deals, it is defined by sovereign AI alignment rather than integrated energy design.

The most uncertain element is power. The transaction specifies a 200-megawatt facility but does not disclose its energy source or procurement structure.

This missing layer distinguishes it from the North American entries where power is explicitly engineered into the deal structure.

The capital component is a corporate investment estimated at approximately $830 million, announced rather than fully closed. While meaningful, it remains below the multibillion-dollar scale seen in leading hyperscaler-led programs.

The strategic positioning is clearer than the infrastructure detail. The investment reflects a domestic European AI developer anchoring compute capacity within its home market, aligned with broader sovereign AI ambitions.

While no direct state capital is indicated, the framing is consistent with national-champion infrastructure development.

Europe’s signal this month is therefore asymmetric:

Strategic intent around sovereignty is strong, but the underlying power architecture remains less defined.

Amazon — Navi Mumbai Data Center Development (APAC)

Amazon’s Navi Mumbai project anchors APAC in April, though its ranking reflects relative positioning rather than absolute scale.

The region produced a seven-way tie at the top, and this transaction advanced primarily due to marginally stronger power disclosure.

The key differentiator is its mixed renewable power orientation, which provides at least partial visibility into the energy strategy supporting the facility.

This stood out against several larger regional announcements, including a $15 billion Google project in India, where power details were not disclosed.

The estimated investment is approximately $430 million, modest compared to other hyperscale commitments in the region.

The project remains pre-permit, with Amazon Data Services India still seeking environmental clearance, placing it firmly in the early execution stage.

APAC’s broader pattern is evident here:

Large announced capital pipelines are not yet matched by equally mature energy or permitting visibility.

Even the most credible deal remains dependent on regulatory progression before construction certainty emerges.

APAC's broader pattern is evident here: large announced capital pipelines outrun energy and permitting visibility. Even the most credible deal depends on regulatory progression before construction certainty the same gap shadowing Microsoft's $50B India AI bet.

Volt — 29MW Dubai Data Center Development (Middle East & Africa)

Volt’s Dubai project leads the Middle East and Africa region, though its position reflects low regional deal density rather than exceptional scale.

It is an early-stage development in a market where power structure definitions remain incomplete.

The power dimension is not disclosed. While the project specifies a 29-megawatt capacity, it does not identify energy sourcing or generation strategy, leaving its scalability and cost structure open-ended.

No capital figure is provided, consistent with its joint venture status, reinforcing the early-stage nature of the transaction.

From a regional perspective, this is the most concrete AI infrastructure development recorded in April across the Middle East and Africa dataset.

However, it also highlights the broader constraint:

Regional competitiveness is still contingent on translating energy availability into clearly structured, power-backed capacity.

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

Permitting Reform Exposed The Power Constraint

Italy has solved its data center permitting problem and exposed a larger one.

The February 2026 unified authorization decree compressed a multi-year, fragmented approval process into a single procedure with a 10-month statutory deadline.

Market coverage has read that reform as the green light for Italy’s buildout.

The reading is incomplete.

With permitting compressed, the binding constraint is now power.

Terna is holding more than 50 gigawatts of data center connection requests against a grid that can energize only a fraction of them in the near term.

The capital flooding into Italy will be sorted not by who can permit a campus but by who can power one.

The 50-Gigawatt Queue Is Mostly Virtual

The headline pipeline number overstates buildable capacity.

Terna recorded data center connection requests above 50 gigawatts across more than 300 projects by June 2025, up from 30 gigawatts six months earlier. Buildable near-term capacity is a fraction of that figure.

The substations serving Milan and Rome face a combined shortfall of roughly 150 megawatts by 2028 as metropolitan data center demand breaches 500 megawatts.

Terna has launched its Efficient Territorial Planning model specifically to screen out speculative and inactive requests and free saturated capacity for viable projects.

That tool exists because most of the 50 gigawatts is virtual.

Over the next 12 to 24 months, the queue will sort into projects with secured grid allocations and projects that were always speculative.

Investors underwriting Italian exposure should benchmark allocated and energized capacity, not announced pipeline.

The pipeline figure is a marketing number. The interconnection is the asset.

Investors should benchmark allocated and energized capacity, not announced pipeline. The pipeline figure is a marketing number. The interconnection is the asset the same delay now repricing data center markets across Europe.

Power Position Now Sorts The Market

Capital is not the scarce input in Italy. It is abundant and competing.

Microsoft has committed 4.3 billion euros to its Italian cloud region, Amazon Web Services 1.2 billion euros, and Gulf-backed capital has entered through the Eni and Khazna joint venture at Ferrera Erbognone.

The scarce input is a defensible power position.

Developers in congested northern zones now secure grid allocations 18 to 24 months before breaking ground and pre-fund substation upgrades that raise land costs by 10 to 15 percent.

Italy also carries among the highest wholesale electricity prices in Europe, which Polytechnic University of Milan has flagged as a direct risk to investment competitiveness.

The structural response is captive generation.

Eni is supplying its Ferrera Erbognone campus through gas generation paired with carbon capture. The utility A2A has dedicated 1.6 billion euros to data center development off the back of its own power capacity.

The market is bifurcating.

Projects with captive power or secured allocations will price and schedule on their own terms.

Grid-dependent projects carry a timing and cost discount the market has not yet applied.

The market is bifurcating. Captive-power and secured-allocation projects price on their own terms. Grid-dependent projects carry a timing and cost discount the market has not yet applied the pricing logic of clean power scarcity as the hidden constraint.

The Next Capacity Is Not In Milan

The geography of Italian data center capacity is about to shift.

Roughly 60 percent of grid connection requests are concentrated in Piedmont and Lombardy, the regions already in virtual saturation.

The buildable capacity is moving south and to the coast.

Southern Italy sits under the Single Special Economic Zone, which offers tax credits and expedited authorization across Campania, Sicily, Puglia, and other southern regions.

The Italian energy decree explicitly encourages southward distribution to access higher renewable and storage headroom.

The coastline carries the connectivity:

Sicily lands more than 20 submarine cables, and Genoa and Bari are emerging as carrier-neutral gateways to Africa, the Middle East, and India.

Terna is spending more than 23 billion euros through 2034 to lift cross-zone transfer capacity from 16 to 39 gigawatts, a plan built to move southern renewable power to northern demand.

Over 12 to 24 months, the investable Italian data center will increasingly sit outside Milan, in coastal and southern nodes with renewable headroom and subsea connectivity.

Capital that screens only the Milan market is screening the saturated half of the country.

The Investor Action Layer

Private capital should re-sequence diligence on Italian assets.

The first question is no longer the developer or the cost of capital. It is whether the asset holds a defensible power position, a secured grid allocation or captive generation.

Funds that screen on capital structure and yield before power access will allocate into projects that stall in the connection queue.

The cost of inaction is a portfolio of permitted-but-unpowered sites acquired at powered-site prices, and a southern opportunity ceded to faster movers.

Public markets hold several instruments tied directly to the constraint.

Eni, the utility A2A, and Terna are all listed. Terna’s 23 billion euro grid plan is itself an investable thesis on Italian data center growth, because the grid build is the gating input to the buildout.

Energy majors and utilities converting grid position into contracted data center revenue represent optionality not yet in consensus estimates.

Investors should benchmark which listed Italian operators hold powered land near demand, and which carry the grid mandate, rather than waiting for the revenue to surface in reported segments.

Operators should treat power partnership and geography as the first two decisions, not the last.

Entering the Milan or Rome corridor without an incumbent energy or utility partner means negotiating against the connection queue from the back of the line.

The February 2026 decree compresses permitting but does not resolve local zoning, and it does not add a megawatt of grid capacity.

Operators that sequence land first and power second will calibrate their delivery timelines to Terna’s queue rather than their own.

The Verdict On Italian Capacity

Italy has completed its transition from second-tier European market to a top destination for data center capital.

The capital is committed, the permitting is reformed, and the political alignment is in place. The constraint that remains is physical.

Power generation and grid interconnection now govern which announced projects become operational capacity and which stay on a connection list.

Italy is the clearest case in Europe because its queue is the most visibly saturated, but the same inversion is forming in every market where connection requests outrun buildable capacity.

The question for the next 24 months is not how much capital Italy can attract.

It is how much of the 50-gigawatt pipeline can actually be energized, and whether the capital now committing has underwritten the power, or only the campus.

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 Credit Committee

In 2023, a credit committee at a major infrastructure lending institution sat down to evaluate a compute factory financing.

The deal on the table was a purpose-built GPU-dense facility liquid-cooled, engineered from the component level for AI training workloads.

Built to the specifications The Retrofit Problem: Why Legacy Data Centers Cannot Serve AI Workloads identified as the qualifying threshold.

The facility had a single tenant, under a long-duration lease. Power density was roughly three times the committee’s standard data center benchmark.

Asset life, measured in GPU hardware cycles rather than building depreciation, produced a depreciation schedule the model had never processed.

Standard framework occupancy rate, diversified tenancy, cap rate, net operating income produced a number that felt wrong to everyone in the room.

The debate was not whether to do the deal, but whether the model needed to be rebuilt before anyone could correctly price it.

The model needed to be rebuilt.

How Data Center Credit Has Always Worked

The colocation model that dominated data center financing for two decades was built on a specific credit logic.

The building was the asset. Tenants were multiple. Leases were long but staggered.

The credit thesis was real estate: diversified cash flows from a pool of tenants, secured by an asset with residual value independent of any single occupant.

This model produced a predictable risk-return profile.

A colocation facility with twenty tenants, a weighted average lease term of seven years, and 95% occupancy was a well-understood credit.

Banks underwrote the building. Tenants provided credit support. If one tenant left, nineteen remained.

The asset retained value because it could be re-leased to another tenant at broadly similar economics.

Open Compute Project: The Hardware Moat Facebook Built and Gave Away introduced the design efficiency variable two facilities at identical capacity produce different returns when one was engineered at the component level for its workload.

The colocation credit model never needed to price that variable. The building was the asset regardless of what ran inside it.

The compute factory changed every assumption the colocation credit model was built on.

The CoreWeave Structure

CoreWeave raised $7.5 billion in debt financing in 2024. The credit structure was built around the Microsoft contract not the building.

Microsoft had committed to purchasing GPU-dense compute capacity from CoreWeave at defined terms, duration, and scale representing CoreWeave’s primary revenue base. Apollo and participating lenders underwrote Microsoft’s payment obligation.

The data centers were the collateral. The Microsoft contract was the asset.

That distinction is precise and carries significant implications. The lenders modeled Microsoft’s balance sheet. The covenant package was built around contract performance.

Credit assessment was anchored to the counterparty’s creditworthiness. The facility’s net operating income was secondary across all three dimensions.

The model was industrial project finance the same framework used to underwrite power purchase agreements, pipeline throughput commitments, and contracted manufacturing capacity. It was not a data center credit.

The committee that understood that distinction cleared the deal. The committee that applied the colocation framework to the same transaction would have priced the risk incorrectly in both directions.

The Offtake Agreement Defined

The offtake agreement is a contract in which a buyer commits to purchasing a defined quantity of output from a producer over a defined period at defined terms. The agreement exists before the factory is built. The factory is built because the agreement exists.

In power generation, the offtake agreement is the power purchase agreement. In pipelines, it is the throughput commitment. In contracted manufacturing, it is the supply agreement. In each case, the credit underwriting starts with the agreement.

The asset is what produces the output the agreement commits to buying. The agreement is the reason the asset has value.

The compute factory is an industrial facility. The GPU cluster produces AI training capacity. The offtake agreement, the committed customer contract, determines whether that capacity has a buyer.

Contract duration, counterparty creditworthiness, and pricing terms are the three variables that produce the credit profile. The building’s square footage is not in the primary equation.

You should apply this framework when evaluating any compute factory financing.

Start with the contract. What is the duration? Who is the counterparty? What does the pricing structure look like under stress?

The answers produce the credit profile. The asset analysis follows from there.

The Complication

CoreWeave secured its terms because the counterparty was Microsoft.

Its strong balance sheet reduced the risk premium to investment-grade infrastructure levels, and lenders priced the deal accordingly.

The next compute factory deal won’t always have Microsoft behind it.

Some will involve weaker or less established counterparties, including emerging operators, regional AI firms, or sovereign-backed buyers relying on political rather than corporate credit strength.

Credit frameworks adjust for weaker counterparties. A weaker offtake counterparty requires a higher risk premium.

Shorter contract durations increase the residual asset risk lenders must price. More specialized asset designs tied to a single workload also reduce residual collateral value if the contract fails.

The offtake agreement framework accommodates all three adjustments systematically.

Traditional colocation credit models cannot, because they were designed for real estate assets with diversified tenant bases, not industrial assets tied to a single committed buyer.

Three Positions on Offtake-Backed Credit

For private credit investors evaluating compute factory deals, the offtake framework requires a counterparty due diligence process that standard infrastructure credit teams are not always staffed for.

The technical evaluation of the compute factory asset GPU density, thermal architecture, power position is necessary but secondary.

The primary work is understanding the offtake counterparty’s financial position, the durability of their demand for the specific compute capacity being financed, and the enforceability of the contract terms under stress.

For infrastructure equity investors, the offtake framework changes the holding period analysis entirely. A compute factory with a ten-year offtake agreement from a strong counterparty is a different investment from the same facility with a three-year agreement and a renewal option.

The equity value is a function of contract duration and renewal probability. Valuing the equity on real estate comparable transactions produces the wrong answer and produces it consistently.

For sovereign and institutional capital allocators entering the compute factory asset class, the offtake framework identifies the entry point with the most attractive risk-adjusted return.

The deals where offtake quality is strong and the market has not yet priced the counterparty correctly represent the allocation window. The Meta-Apollo deal priced Microsoft-quality offtake at investment-grade terms.

The next tranche of deals will price a range of counterparty quality and the capital that can distinguish between them will allocate more precisely than the capital treating all compute factory debt as equivalent.

The Forward Question

The offtake framework resolves the credit question for compute factory deals with established counterparties. It is the correct model.

It is the industrial model. It is how every major capital-intensive industry has been financed when output must be committed before the factory is built.

The unresolved question is what happens when the offtake counterparty is a sovereign buyer whose commitment is real, but whose creditworthiness is measured by different instruments than a corporate balance sheet.

Sovereign capital changes the offtake framework in two directions simultaneously. It introduces a counterparty whose payment certainty is backed by fiscal authority rather than corporate cash flow.

And it introduces a return requirement that includes non-financial objectives (national capacity building, technology transfer, domestic employment) that standard credit analysis does not capture.

The framework holds. The inputs change. That is where this series goes next.

Join Global Data Center Hub