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

Table of Contents

1. Executive Summary

2. The Size and Shape of the US Data Center Market

3. The Operator Layer

4. Demand Is Not the Problem

5. Money Is Not the Problem

6. Power Delivery Is the Gate

7. PJM’s Auctions Price the Shortage

8. The Utility Layer

9. The Neocloud Layer

10. Recommendations

Executive Summary

One thing decides how fast US data centers grow in 2026.

It is the ability to get electric power delivered to the site not money, land, chips, or demand.

Two gates control power delivery.

The first is the interconnection queue. This is the waiting line a project joins to get permission to plug into the electric grid.

The second is the equipment supply chain. Transformers and switchgear are the machines that take high-voltage grid power and convert it into power a data center can use. Both gates are jammed.

The numbers show how jammed. Goldman Sachs Research (2026) estimates only 60% of next year's scheduled capacity will arrive on time, dropping to 50% over the following two years.

Bloomberg, citing Sightline Climate (May 2026), estimates 30–50% of the roughly 16 GW planned for 2026 will be delayed or canceled due to power availability, equipment shortages, and community opposition.

The constraint looks different in each market. Northern Virginia faces grid and transmission delays, Texas new curtailment rules, Phoenix power shortages, and Santa Clara transmission limits.

Nationally, large power transformers average 128-week lead times, while generator step-up units take 144 weeks (Wood Mackenzie, Q2 2025).

Money and demand prove the point from the other side.

The five largest hyperscalers plan roughly $602 billion in 2026 capital spending, with about 75% tied to AI (CreditSights, January 2026).

Vacancy across primary U.S. markets is a record-low 1.4%, and 81.5% of capacity under construction is already leased (CBRE, H2 2025).

The buyers are waiting. The capital is waiting. The power is not there.

One common read needs correction. The constraint is not regional. The way it shows up is regional.

The cause is national, running through the grid queue and transformer supply.

1. The Size and Shape of the US Market

The National Totals

The US market can be counted three ways.

By facilities: 17.2 gigawatts of active IT load across 2,396 data centers (ABI Research, 2025).

By revenue: a $72.37 billion colocation market led by Equinix and Digital Realty (United States Data Center Colocation Databook, 2026).

By electricity: data centers consumed 176 terawatt-hours in 2023 (4.4% of U.S. electricity), with LBNL projecting 325–580 terawatt-hours by 2028 and EPRI estimating 9–17% of U.S. electricity demand by 2030.

One Country, Many Markets

The US is not one market in any useful investment sense.

It is a collection of distinct grid regimes, utility tariffs, transmission constraints, and planning processes.

PJM, ERCOT, and CAISO each have different interconnection rules, while utilities such as Dominion, Oncor, APS, Georgia Power, and AEP Ohio set their own terms for large loads.

Where a project sits in this system determines how quickly it gets power, what it pays, and the obligations it carries.

Concentration at the Top

The market is heavily concentrated. Northern Virginia alone has 4,039.6 megawatts of colocation inventory, about 43% of the 9,432-megawatt U.S. primary-market total (CBRE, H2 2025).

Loudoun County, its core, has 200 operating data centers and 117 more in development (October 2025).

Operator concentration is similar: Digital Realty holds the largest U.S. leased data center power share at 15%, followed by Equinix at 4.9% (Q4 2024).

The Hub Map

The build-out runs through a small set of hubs, each defined by its grid and its utility.

The table below compiles the structural marker for each hub from utility and grid disclosures.

2. The Operator Layer

The table below ranks the top 25 US operators by operational IT capacity in 2025.

These are operational figures. Announced pipelines are far larger, and the notes beneath the table cover the gap.

_{Source: ABI Research, “Top 25 Largest Data Center Companies in the U.S. by Active IT Capacity,” February 25, 2026. Total US active IT load: 17.2 gigawatts across 2,396 data centers.}

Three notes on reading the table.

First, ABI’s hyperscaler megawatt figures appear conservative relative to capex and build activity.

ABI itself reports AWS added about 3.8 gigawatts of capacity in 2025.

Second, operational figures and announced pipelines are different things.

CloudHQ’s Ashburn LC campus is planned for more than 1.7 gigawatts at full build-out. STACK is launching a 1-gigawatt Virginia campus.

CyrusOne is building a 760-megawatt Texas campus. DataBank is developing a 480-megawatt campus in Red Oak, Texas, and a 192-megawatt campus in Culpeper, Virginia.

Third, methodology varies by tracker. CBRE’s H2 2025 market data, which counts colocation operating inventory, puts Northern Virginia alone at 4,039.6 megawatts, about 43 percent of the 9,432-megawatt US primary-market total.

3. Demand Is Not the Problem

The Forecasts Point One Direction

The demand estimates vary by source, and all of them are estimates built on utility and grid forecasts that the forecasters themselves flag as uncertain.

LBNL estimates data centers used 176 terawatt-hours in 2023 (4.4% of U.S. electricity), rising to 325–580 terawatt-hours (6.7–12%) by 2028.

Grid Strategies (December 2025) projects the five-year U.S. peak-load forecast reaching 166 gigawatts by 2030, with about 90 gigawatts driven by data centers.

S&P Global likewise projects data center electricity load more than doubling from 61.8 gigawatts in 2025 to 134.4 gigawatts by 2030.

The Market Data Confirms It

Vacancy tells the cleaner story. Primary-market vacancy fell to a record low of 1.4% at year-end 2025, even as supply grew 36% to 9,432 megawatts.

Net absorption reached a record 2,497.6 megawatts, with Northern Virginia accounting for 1,102 megawatts, while 81.5% of capacity under construction was already pre-leased (CBRE, H2 2025).

A 1.4% vacancy rate is what a supply squeeze looks like from the demand side.

We examined the supply side of the same squeeze in Land Is the New Bottleneck.

Under-construction capacity fell for the first time since 2020 from 6,350 to 5,994 megawatts not because demand weakened, but because permitting, zoning, and power-procurement delays slowed new development.

4. Money Is Not the Problem

The five largest hyperscalers reached about $602 billions of planned capital spending for 2026, up 36 percent year over year, with about 75 percent, roughly $450 billion, tied to AI (CreditSights, January 2026).

The Big Four alone guided up to $630 billion for 2026, about 62% above the record $388 billion spent in 2025, led by Amazon, Google, Meta, and Microsoft.

The longer arc is larger. Goldman Sachs Research (June 2026) projects Meta, Microsoft, Amazon, and Alphabet will spend a combined $5.3 trillion from 2025 to 2030, up from its previous $4.5 trillion forecast.

Private capital matches the corporate wall of money. Infrastructure funds held more than $1.7 trillion in assets, including about $400 billion in dry powder as of September 2025 (Goldman Sachs).

The roughly $40 billion Aligned acquisition by BlackRock, Nvidia, and Microsoft underscores strong exit liquidity, while project finance, private credit, and asset-backed securities remain widely available.

Capital is chasing power. Power is not chasing capital.

5. Power Delivery Is the Gate

The Queue Is Measured in Years

The federal target for signing an interconnection agreement is 8 to 11 months. The actual national average wait is about 25 months.

In ERCOT, the Texas grid, it is about 20 months. In PJM, the grid serving the mid-Atlantic, it is about 40 months (Carbon Direct, “AI Meets the Grid,” May 14, 2026).

Those are averages. In the zones where data centers are actually growing, active projects wait 36 to 48 months in both PJM and ERCOT.

In Northern Virginia, Phoenix, and Dallas, grid interconnection waits run 4 to 7 years (Sightline Climate via Bloomberg, May 2026).

The national queue is enormous. LBNL’s “Queued Up: 2026 Edition,” with data through the end of 2025, counts about 2,060 gigawatts of generation and storage actively seeking interconnection.

For projects built in 2025, the median time from request to commercial operation exceeded 5 years.

The Transformer Is the Deeper Gate

Even a project with a grid connection cannot energize without equipment.

Wood Mackenzie’s Q2 2025 survey found standard power transformers average 128-week lead times, while generator step-up transformers take 144 weeks.

Some specialized units now require up to four years, and substation transformer lead times have exceeded 160 weeks.

Prices moved with the wait times. Since 2019, power transformer prices have risen 77% and generator step-up transformer prices 45%, with some distribution transformers up as much as 95% (Wood Mackenzie, August 2025).

Demand has surged alongside prices, with power transformer demand up 119% and generator step-up demand up 274%.

The equipment market itself is scaling to meet the build.

Wood Mackenzie projects the U.S. data center electrical equipment market growing from about $20 billion in 2026 to $65 billion by 2030 as data center capacity expands from roughly 24 gigawatts to 100 gigawatts.

The Gate Has Moved

In 2023 and 2024, the gate was chips and capital. In 2025 and 2026, the gate is power delivery and electrical equipment.

This is the thesis we laid out in Power Is the Moat: power availability not land, not fiber, not capital is the gating factor for AI infrastructure. The mid-2026 data confirms it.

Bloomberg’s April 2026 reporting frames the binding constraint as physical infrastructure.

Transformers, switchgear, and batteries, much of it dependent on Chinese supply, now set the timeline.

One industry source put it plainly: “When equipment availability becomes the gating factor instead of capital or permitting, your entire expansion timeline gets dictated by supply chains you can’t control” (via IndustrialSage).

6. PJM’s Auctions Price the Shortage

A capacity auction is the market where a grid operator buys promises of future power supply.

When the clearing price spikes, supply is scarce. PJM’s auctions now read as a scarcity alarm.

The clearing price climbed from $28.92 per megawatt-day for the 2024/25 year to $269.92 for 2025/26, then $329.17 for 2026/27, then $333.44 for 2027/28 (PJM, Utility Dive).

The last three auctions all cleared at the FERC-approved price cap. The 2025/26 jump alone was an 833 percent single-year increase.

Data centers drove the move. Monitoring Analytics (January 2026) found that data centers caused 63 percent of the 2025/26 price increase, about $9.3 billion in added costs, and accounted for 40 percent of 2027/28 auction costs, $6.5 billion of $16.4 billion.

The December 17, 2025, auction for 2027/28 crossed a line.

For the first time, the entire regional grid fell short of its reliability requirement, by about 6,625 megawatts.

The Dominion and BGE zones cleared at premiums of $444.26 and $466.35 per megawatt-day in 2025/26.

Without the price cap agreed with Pennsylvania, PJM estimates the 2027/28 price would have reached about $530 per megawatt-day, roughly 60 percent higher.

7. The Utility Layer

Dominion and PJM: The Queue Plus Transmission Case

Dominion Energy in Virginia holds about 40 gigawatts of data center capacity under contract and about 70,000 megawatts of data center interconnection requests in queue.

The queue is nearly three times the system’s peak load, growing by about 10 requests per month, though Dominion expects about 25,000 megawatts to connect by 2031.

PJM projects the Dominion zone as its fastest-growing region, with the proposed 525 kV transmission line adding 3–4 gigawatts no earlier than 2029–2032.

Meanwhile, Virginia’s SCC is reviewing the large-load interconnection process.

Across PJM, the timeline from interconnection application to commercial operation now exceeds 8 years, versus under 2 years in 2008 (RMI).

Since 2023, PJM has processed more than 170,000 megawatts of generation requests, with about 30,000 megawatts remaining in the 2026 transition queue.

New programs such as Bring Your Own New Generation and Non-Capacity-Backed Load allow faster interconnection in exchange for curtailment risk, while developers on the expedited track pay 100% of network-upgrade costs.

ERCOT: Speed With a New Catch

Texas holds more than 233 gigawatts of large-load interconnection requests as of December 2025, more than 70 percent of them data centers, up about 300 percent in a year.

In Q1 2026 alone, another 198 gigawatts of large-load requests were submitted roughly equal to ERCOT’s entire peak load while the generation queue reached 421–432 gigawatts, about 77% from solar and storage.

SB 6, signed June 20, 2025, changes the rules.

It classifies large loads as 75 megawatts or more, imposes a minimum $100,000 study fee, requires curtailment protocols for new large loads, and allows ERCOT to dispatch on-site generation during emergencies.

Texas still offers fast grid connections, but that advantage comes with greater curtailment risk, while Oncor’s proposed 765 kV transmission line is not expected until around late 2028.

Georgia Power: New Generation, New Guardrails

Georgia Power projects about 8,500 megawatts of load growth over six years.

The Georgia PSC approved 9,885 megawatts of new generation in December 2025, with about 80% supporting data centers at an estimated lifetime cost of $50–60 billion.

Large-load customers above 100 megawatts face minimum billing and longer contracts, while Georgia Power will backstop costs through 2031 if demand falls short.

At the same time, local development moratoriums are expanding, including around the Atlanta BeltLine, MARTA buffers, Fayetteville, and Camden County.

Phoenix: Power, Not Water

Phoenix is the number two US market, with about 1,380 megawatts commissioned and more than 100 area data centers drawing about 1.4 gigawatts.

Water is a real but secondary constraint.

Phoenix holds a state-verified 100-year water-supply certification, and operators increasingly deploy closed-loop and air cooling, including Aligned’s Delta3 arrays and Meta’s reuse projects.

Power and transmission are the tighter near-term gate, running through new APS 230 kV lines and Salt River Project capacity.

The common assumption that water gates Phoenix is wrong.

Santa Clara: The Hard Limit

Silicon Valley Power in Santa Clara is the clearest hard-limit case in the country.

Digital Realty’s SJC37, built for 48 megawatts, and Stack’s SVY02A, at 48 megawatts, sit fully built but unpowered.

Nearly 100 megawatts of finished capacity waits on a $450 million system upgrade due in 2028.

Digital Realty’s site obtained planning permission in 2019 and remains unenergized six years later.

The city has 57 active or under-construction data centers.

Operators are working around the ceiling with on-site generation, including ECL’s 35-megawatt FlexGrid facility using grid, gas, and hydrogen fuel cells, and flexibility pilots run by SVP and Emerald AI.

8. The Neocloud Layer

A neocloud is a company that rents out GPU computing power built for AI workloads.

Most now report contracted or active power in megawatts and gigawatts instead of GPU counts, as Blackwell systems make GPU totals less meaningful.

Two distinctions matter: company-disclosed versus third-party estimates and contracted versus deployed capacity.

CoreWeave (public, CRWV) disclosed about 250,000 GPUs across 32 data centers and 360 megawatts active as of December 31, 2024, in its S-1.

It now reports 43 active data centers, more than 850 megawatts active, and over 3.1 gigawatts contracted at end-2025.

Revenue exceeded $5.1 billion with a $66.8 billion backlog, targeting 1.7 gigawatts active by end-2026 and 5 gigawatts by 2030. Microsoft accounted for 62–67% of revenue.

Crusoe (private) reports power. Its Abilene, Texas campus (Stargate Site 1) spans 1.2 gigawatts, expanding to roughly 2.1 gigawatts with a new Microsoft campus. Oracle expects the site to eventually host more than 450,000 GB200 GPUs.

Nebius (public, NBIS) reports megawatts and annual recurring revenue. It targets 800 megawatts to 1 gigawatt connected and 2.5 gigawatts contracted by end-2026.

U.S. expansion includes Kansas City and New Jersey, supported by major Microsoft and Meta contracts.

Lambda (private) runs roughly 15 to 20 US data centers with more than 320 megawatts committed, and targets 3 gigawatts by 2030 and more than 1 million NVIDIA GPUs long term.

Its Kansas City site hosts more than 10,000 Blackwell Ultra GPUs, and it signed a multi-billion-dollar Microsoft agreement in late 2025.

Voltage Park (private) built about 24,000 H100 SXM5 in 2023 for roughly $500 million and now runs about 35,000 to 36,000 GPUs across six U.S. facilities, with about 60 megawatts active following its January 2026 merger with Lightning AI.

Together AI (private) runs an aggregation model on partner infrastructure. It deployed a 36,000-GB200 cluster in 2025 and has secured access to more than 100,000 GPUs, though much of that capacity is leased rather than owned.

IREN (public, IREN) gives the most granular disclosure: about 23,000 GPUs as of September 2025 across multiple NVIDIA and AMD platforms.

It operates six data centers with 2,910 megawatts of grid-connected capacity and targets $4.4 billion in annual recurring revenue.

Vultr (private) runs 32 global locations, not US-only, offering NVIDIA and AMD AI platforms, with a $3.5 billion valuation established in 2024.

Nscale (Europe-headquartered with US operations) has about 200,000 GB300 contracted with Microsoft across Europe and the US as of October 2025.

The figure reflects contracted capacity, not deployed hardware.

9. Recommendations

Underwrite to the power-delivery date, not the construction date.

The gating variable is utility energization plus transformer delivery, not building shell completion.

For any US site, demand documented evidence of three things before crediting a delivery timeline: an executed or secured interconnection position, transformer and switchgear order slots with an assumed 24-to-48-month lead, and allocated substation capacity.

Treat announced pipeline as optionality, not supply. Only about one-third of 2026-announced capacity was under active construction as of mid-2026.

Prefer markets and structures that bypass the queue.

ERCOT retains a genuine speed-to-power edge, now tempered by SB 6 curtailment exposure.

Behind-the-meter and on-site generation, along with PJM’s Bring Your Own New Generation track, offer routes around the interconnection queue.

Secondary markets with spare substation capacity de-risk timelines. Discount saturated grids.

Northern Virginia is transmission-constrained until roughly 2029 to 2032, and Santa Clara until 2028.

Monitor four thresholds as constraint-shift signals.

First, large power transformer lead times need to fall below roughly 80 weeks from today’s 128.

Second, PJM capacity prices dropping below about $300 per megawatt-day would signal improving supply.

Third, interconnection reforms must shorten the current 3–4 year data center wait, with PJM’s new process and ERCOT’s SB 6 rules rolling out in 2026.

Fourth, a cut in hyperscaler capex guidance would indicate demand not power is finally easing.

Until at least two of these change, power delivery remains the binding constraint.

Underwrite counterparty bifurcation in the neocloud layer.

Hyperscalers and CoreWeave carry durable balance sheets and long-dated contracts.

Smaller neoclouds carry customer-concentration risk and GPU-depreciation risk, since a GPU financed over 5 to 6 years can sit against contracts that run shorter.

Always separate contracted power from deployed GPUs. The two diverge widely.

Nscale’s 200,000 GB300 is contracted, not live. CoreWeave’s 3.1 gigawatts contracted stands against 850 megawatts active.

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 debate over who wins AI infrastructure has fixated on hyperscaler capex totals.

That lens misses where the more interesting structural work is happening: at the converter tier, where former Bitcoin miners are turning secured power into contracted data center cash flow.

The binding constraint across the entire cycle is energized megawatts.

The scarce capability is financing them without diluting equity or taking recourse risk onto the parent.

Hut 8’s Q1 2026 print, read correctly, is a case study in solving the second problem against the first.

Read at the headline, the quarter looks weak.

Hut 8 posted a $253.1 million consolidated net loss, missed on revenue, and missed on reported EPS.

Read at the capital structure, it is one of the cleaner infrastructure repricings of the cycle.

The company closed the quarter with a mining balance sheet and, within weeks, financed two hyperscale AI campuses in the investment-grade bond market on non-recourse terms.

The income statement and the enterprise value moved in opposite directions, and the equity followed the enterprise value.

The Signal: Accounting Noise Over a Structural Rerating

Consolidated revenue reached $71.0 million, up 226% from $21.8 million a year earlier, with $66.0 million from the Compute segment.

The headline $253.1 million net loss was driven primarily by $295.7 million in unrealized, non-cash mark-to-market losses on the Bitcoin treasury under fair-value accounting after Bitcoin fell about 22% during the quarter.

Beneath the accounting loss, gross margin expanded to approximately 64% from 14% a year earlier, reflecting stronger energy procurement and utilization rather than treasury volatility.

The market recognized the distinction: despite missing revenue and EPS expectations, the stock rallied as investors increasingly valued Hut 8 on contracted future cash flow instead of quarterly earnings.

That shifts the analytical focus.

The quarter provides context, but the investment case lies in the post-quarter capital structure.

The key value drivers, including leases, bonds, refinancings, and divestitures, were all executed after March 31.

Focusing only on quarterly results captures an accounting loss while missing the underlying business transformation.

Infrastructure Strategy: Power-First, Tenant-Backed, Density-Engineered

Hut 8 organizes its multi-gigawatt platform across Power, Digital Infrastructure, and Compute, but its strategy centers on developing power assets into long-term data center leases.

It isolated its mining business in publicly listed American Bitcoin Corp., where it retains an 80% stake, repositioning the parent as an energy and AI infrastructure developer serving investment-grade tenants.

River Bend, a 330 MW Louisiana campus with 245 MW of IT capacity, is fully leased to Fluidstack under a 15-year triple-net agreement backed by Google and Anthropic, with projected annual NOI of about $454 million at stabilization.

Beacon Point, a 1 GW Texas campus, has 704 MW of committed IT capacity following Phase 2 activation, supporting $19.6 billion in base-term contracts.

Its tenant is described as high investment grade but remains undisclosed.

The campuses are designed for NVIDIA’s DSX AI architecture, with liquid cooling from Vertiv and EPCM by Jacobs.

Beyond them, Hut 8 has an 8,375 MW development pipeline spanning diligence through construction.

The pipeline represents future growth, while River Bend and Beacon Point demonstrate the model already works.

Capital Allocation: The Part That Actually Reprices the Equity

Hut 8’s financing architecture distinguishes it from a mining company.

It raised $7.5 billion in non-recourse, investment-grade secured project notes across River Bend and Beacon Point.

River Bend’s $3.25 billion BBB- bond fully funded construction at roughly 95% loan-to-cost while returning $184 million of equity to the parent.

Beacon Point replicated the structure with a $4.25 billion issuance. Because the debt is non-recourse and amortized through contracted lease payments, the projects require no parent-level refinancing or equity dilution.

The parent balance sheet was strengthened alongside the financing.

Quarter-end cash and Bitcoin totaled about $1.3 billion, with $795.6 million held at the parent.

Parent-level recourse debt is limited to a $159.3 million convertible note.

Post-quarter, Hut 8 refinanced its Bitcoin-backed loan, released about 3,300 BTC as collateral, sold its 310 MW Ontario gas portfolio, and reduced crypto-related risk while retaining its 16,331 BTC treasury through a 22% Bitcoin decline.

The market is still deciding how to value Hut 8. Analyst targets range from $165 to $226, versus an average near $127 and a share price around $91.

The wide spread reflects an unresolved question: whether Hut 8 should be valued as a crypto company or as a contracted infrastructure platform.

Competitive Positioning: A Converter With a REIT’s Cost of Capital

The competitive implication is that Hut 8 has achieved the financing profile of an established data center REIT without decades of operating history.

That shift rests on secured interconnection, investment-grade tenants, and take-or-pay triple-net leases that transfer operating, tax, and power-price risk to tenants.

The bond market validated this structure with a BBB- rating, and Beacon Point’s second phase shows it is repeatable.

Hut 8 now competes with established infrastructure platforms on speed to energization and contracted backlog quality.

Its strengths are a large development pipeline and the ability to secure asset-level financing without straining the balance sheet.

Its primary risks remain greenfield execution and Bitcoin treasury volatility, which will continue to distort GAAP earnings and contribute to share-price volatility.

The Binding Constraint: Grid Before Everything

The forward risk is no longer demand or capital, both are contracted.

It is the grid.

Hut 8’s pipeline is concentrated on greenfield sites outside traditional data center hubs, where transmission is limited and utilities are tightening large-load interconnection rules.

ERCOT and others are reassessing gigawatt-scale connections.

With $7.5 billion in project debt tied to asset-level performance covenants and Q2 2027 delivery targets, any interconnection, cooling, or construction delay could postpone lease conversion into stabilized cash flow.

Three signals now matter.

First, whether River Bend meets its Q2 2027 target, validating the financing model.

Second, whether the 8,375 MW pipeline converts into signed leases, proving a scalable origination platform rather than a two-campus project.

Third, whether investors value Hut 8 as infrastructure or continue discounting its Bitcoin treasury.

For policymakers, the challenge is already clear: AI infrastructure demand and capital are moving faster than the grid can connect new capacity.

Capital structure was the first defining variable of this cycle.

Hut 8 demonstrates it is not fixed.

By pairing a mining balance sheet with secured power and investment-grade tenants, it secured infrastructure-style financing within a single cycle.

The quarterly loss was an accounting outcome; the $26.6 billion lease portfolio and the bond market’s willingness to finance it on a non-recourse basis define the business.

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 market views Cipher Digital’s first quarter as a successful transition from Bitcoin mining to AI infrastructure.

While mining revenue and earnings disappointed, investors focused on its contracted data-center backlog.

That view is only part of the story.

The backlog is underwritten not by Cipher’s balance sheet, but by two investment-grade counterparties, financed through non-recourse project debt with one parent-level obligation outside that structure.

This is the distinction that determines whether the backlog is durable or merely large.

The AI infrastructure debate has spent two years on capital intensity, on which operator can clear the largest capex number.

For a developer of Cipher’s scale, the deciding variable is different.

It is the quality of the credit standing behind the leases and the architecture of the debt raised against them.

Cipher has assembled both. It has also concentrated both.

The Quarter the Mining Line Stopped Mattering

Cipher reported $34.84 million in revenue for the quarter ended March 31, 2026, down 28.8% year over year and 41.7% sequentially, with both revenue and EPS missing consensus.

The company posted a GAAP net loss of $114.3 million, an improvement from the prior quarter, while Adjusted EBITDA came in at negative $48.0 million.

The financial results, however, are not the main signal.

The revenue decline was intentional as Cipher accelerated its exit from Bitcoin mining before data-center lease revenue begins.

Interest expense rose sharply as project debt was issued ahead of rent commencements.

The quarter reflects a business absorbing financing costs during its transition, with investors largely valuing the contracted backlog over near-term earnings.

The more important development is Cipher's three contracted hyperscale data-center projects, totaling 700 MW of gross capacity and backed by $11.4 billion in contracted revenue over 10–15 years.

The portfolio is expected to generate $787 million in annual net operating income from late 2026 through 2036, marking a shift from volatile mining income to long-term contracted cash flows.

The critical question is not the size of the backlog, but the quality of the credit supporting it.

Two Tenants Behind Eleven Billion

Cipher’s backlog is concentrated around two counterparties.

Black Pearl, a 300 MW site in Texas, is backed by a 15-year triple-net lease with Amazon Data Services, guaranteed by Amazon.com (AA-), and valued at roughly $5.5 billion.

Stingray, a 100 MW Texas site, is also leased to Amazon under a 15-year triple-net agreement with industry-leading economics.

Barber Lake, a 300 MW project leased to Fluidstack, differs because the tenant is unrated.

Instead, Google provides credit support through a recognition agreement covering up to $1.73 billion in obligations in exchange for warrants representing a 5.4% pro forma stake.

The structure is straightforward: two of the three projects depend on Amazon, while the third relies on Google's credit backing.

Roughly $7.5 billion of the contracted backlog is tied to Amazon, with the remainder effectively supported by Google.

This is not a flaw securing investment-grade counterparties is sound infrastructure finance but it means the backlog should be viewed as concentrated exposure to two hyperscalers rather than broadly diversified revenue.

If either Amazon or Google changes its infrastructure strategy, that concentration becomes the key risk behind the headline backlog.

Five Tranches Built to Contain It

The debt architecture is the answer to the concentration, and it is the most sophisticated part of the story.

Cipher carries roughly $5.2 billion of total debt principal, and its structure matters more than its size.

The site notes are non-recourse and issued at the project entity: $1.73 billion at Barber Lake at a 7.125% coupon, $2.0 billion at Black Pearl at 6.125%, and $810 million at Stingray at 6.000%, the descending coupons tracking improving credit perception as the lease book built.

Each tranche is secured by first-priority liens on its own site and serviced only by that site’s rent. A default at one campus cannot cascade to the parent or to the other campuses.

The concentration exposure is real, and the project-finance structure is engineered precisely to wall off each instance of it.

This is the mechanism that separates Cipher from a levered miner.

The ring-fence converts single-tenant risk into contained, site-level risk.

It is why long-only infrastructure capital has rotated into a name that was recently a high-beta crypto proxy.

The structure earns the re-rating.

The Gap the Ring-Fence Does Not Close

The financing structure has one key weakness.

Above the non-recourse project debt sits $1.47 billion of parent-level convertible notes $173 million due 2030 and $1.30 billion due 2031.

Unlike the site debt, these obligations are not supported by lease cash flows but by Cipher’s corporate liquidity and, if necessary, equity dilution.

With mining winding down and rents yet to begin, this remains the company’s primary financial exposure.

Timing increases that risk.

Rent at Black Pearl and Barber Lake is expected to begin in late 2026, with Stingray following in 2027, while project debt starts accruing interest beforehand.

Although Cipher ended the quarter with $715 million in unrestricted cash and $3.51 billion in restricted construction cash, execution remains critical.

Amazon can terminate leases after prolonged construction delays, potentially leaving project debt without its intended cash flow.

The structure protects against tenant default—not construction delays.

What Resolves Next

Three signals will determine whether the backlog converts to the cash flow the market is pricing.

First, on-time rent commencement at Black Pearl and Barber Lake in late 2026; the 180-day termination grace period makes the delivery schedule a credit event, not merely an operational one.

Second, the parent convertible position, which must be refinanced or diluted through and is the one liability the ring-fence does not reach.

Third, conversion of the 3.3-gigawatt development pipeline into signed leases at comparable credit quality, which is the only path that diversifies the two-tenant concentration rather than deepening it.

For investors, Cipher should be valued on counterparty credit, not capacity.

For operators, it demonstrates how project-finance discipline can unlock institutional capital, though parent-level risk remains execution-dependent.

For policymakers, hyperscaler balance sheets are becoming the credit backbone of digital infrastructure, with grid access in Texas and PJM remaining the key bottleneck.

Cipher did not escape the volatility of mining by getting bigger.

It did so by borrowing two of the strongest balance sheets in technology, tranche by tranche, and accepting the one gap the borrowing left open.

Join Global Data Center Hub

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

Join Global Data Center Hub

TL;DR

● Cipher Digital and Hut 8 used non-recourse project bonds to convert Bitcoin mining into contracted hyperscale infrastructure. Cipher carries roughly $5.2 billion in site-level project debt against $11.4 billion in contracted backlog from Amazon and Google. Hut 8 raised $7.5 billion in investment-grade secured notes across two campuses.

● Cipher and Hut 8 diverge entirely above the ring-fence, at the parent level the site debt does not reach. Cipher holds $1.47 billion in parent convertible notes serviced by corporate liquidity, not lease rent. Hut 8 reduced parent recourse debt to $159.3 million against $795.6 million of parent cash.

● The non-recourse structure protects against tenant default but not construction delays. Project debt accrues interest before rent commences. Amazon can terminate Cipher’s leases after prolonged construction delay, and Hut 8’s $7.5 billion in notes carry Q2 2027 delivery targets tied to grid interconnection in Texas and PJM.

The AI infrastructure debate has spent two years focused on hyperscaler capital spending.

That lens is now outdated.

Amazon, Microsoft, Alphabet, and Meta have already secured demand, committed capital, and reached the same constraints in power and components.

The real question has shifted to operators without hyperscaler balance sheets that must create, rather than inherit, creditworthiness.

This converter cohort (former Bitcoin miners transforming energized power into contracted data center capacity) faces a different challenge.

The scarce resource is no longer capital or compute, but the ability to finance energized megawatts without diluting equity or adding recourse risk.

This quarter, Cipher Digital and Hut 8 both demonstrated that model through non-recourse, investment-grade debt.

Reading through their balance sheets and income statements, their results show that competitive advantage is no longer backlog size, but the architecture of the debt supporting it.

Cipher: A Backlog Walled Off Everywhere Except the Top

The market viewed Cipher’s first quarter as a successful transition from Bitcoin mining to AI infrastructure, looking past revenue and earnings misses to focus on its contracted backlog.

That interpretation is directionally correct but incomplete.

The revenue decline was intentional as mining winds down and interest expense rises ahead of lease commencement.

The earnings miss reflects transition financing, not weakening fundamentals.

The real signal is three contracted hyperscale projects totaling 700 MW, backed by $11.4 billion of contracted revenue over 10–15 years and expected to generate $787 million of annual NOI.

Black Pearl and Stingray are leased to Amazon, while Barber Lake is leased to Fluidstack with Google providing credit support.

About two-thirds of the backlog is tied to Amazon, with the balance effectively supported by Google.

Cipher financed these assets with approximately $5.2 billion of non-recourse project debt secured solely by each campus, containing risk at the project level and supporting its institutional re-rating.

The remaining exposure sits at the parent company. Cipher carries $1.47 billion of convertible notes due in 2030 and 2031, backed by corporate liquidity rather than lease cash flows.

With rents not expected until late 2026 and 2027, construction delays remain the principal risk.

The ring-fenced structure protects against tenant default, not delayed delivery or parent-level obligations.

Read the full Cipher analysis: $11.4B Backlog, Two Hyperscalers and the Debt Structure Behind AI Infrastructure

Hut 8: The Parent Recourse Nearly Extinguished

Hut 8 has completed the same transition as Cipher, but further along.

The headline quarter looked weak a $253.1 million net loss alongside revenue and EPS misses but the loss was driven primarily by $295.7 million of unrealized Bitcoin mark-to-market losses.

Beneath that, revenue rose 226% year over year to $71.0 million and gross margin expanded to 64%, leading investors to value contracted future cash flows rather than quarterly earnings.

The company separated its mining business into American Bitcoin and repositioned the parent as an AI infrastructure developer.

River Bend is fully leased to Fluidstack under a 15-year triple-net agreement backed by Google and Anthropic, while Beacon Point supports $19.6 billion of company-stated contracted revenue with an undisclosed investment-grade tenant.

Together with an 8,375 MW development pipeline, these projects underpin Hut 8’s infrastructure strategy.

The financing is what differentiates the company.

Hut 8 raised $7.5 billion of non-recourse, investment-grade project debt across River Bend and Beacon Point, funding construction without parent-level refinancing or equity dilution.

Unlike Cipher, which still carries $1.47 billion of parent-level convertibles, Hut 8 has reduced parent-level recourse debt to just $159 million while holding approximately $1.3 billion of cash and Bitcoin.

The result is a financing profile that increasingly resembles a contracted infrastructure platform rather than a crypto miner.

The remaining challenge is valuation: Bitcoin treasury mark-to-market movements will continue to distort GAAP earnings until investors fully price Hut 8 on its contracted infrastructure cash flows.

Read the full Hut 8 analysis: $26.6B Lease Book and the Financing Shift From Mining to AI Infrastructure

Same Fence, Opposite Exposure: What the Two Structures Reveal

Cipher and Hut 8 share the same financing model but differ on the factor that determines long-term durability.

Both used non-recourse, single-site project bonds backed by triple-net leases to investment-grade hyperscalers, transforming tenant concentration into contained project risk while securing infrastructure-grade financing that traditional mining balance sheets could not access.

The difference lies above the ring-fence.

Project finance isolates site-level risk but does not address parent-company obligations.

Cipher still carries $1.47 billion in convertible debt that must be serviced before lease revenue begins, while Hut 8 has reduced parent-level risk with $795.6 million in cash against just $159.3 million of debt.

The key distinction is the parent balance sheet, not the backlog.

Both also face the same external constraint: the grid.

Their largest projects depend on timely power delivery in Texas and PJM, where interconnection has become the limiting factor.

Because project debt is tied to delivery schedules, energizing capacity on time is now a credit event.

The ring-fence protects against tenant risk, but it cannot protect against delays in grid access.

What Resolves Next

Three signals will determine whether the converter cohort holds.

First, on-time rent commencement at Cipher's Black Pearl and Barber Lake (late 2026) and Hut 8's River Bend (Q2 2027).

Delivery timelines are now credit events, not operational milestones, and any delay will test whether non-recourse financing holds its value.

Second, Cipher's parent-level convertible debt remains unresolved, unlike Hut 8's, making it the key liability outside its ring-fenced structure.

Third, both companies must convert their multi-gigawatt pipelines into investment-grade leases to diversify beyond a small number of hyperscaler customers.

For investors, the focus should be counterparty quality and parent-level capital structure rather than backlog size.

For operators, these deals show that disciplined project finance can unlock institutional capital, but only if parent-company risks are managed separately.

Capital structure defined the first phase of this cycle.

The next differentiator is what sits above the ring-fence and whether the parent can bridge the period between debt issuance and rent commencement.

That will separate durable infrastructure platforms from miners that improved project financing while leaving parent-level risk exposed.

Join Global Data Center Hub

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

Join Global Data Center Hub

TL;DR

• S&P and Fitch disagree by one notch on identical paper because they disagree about what secures it. S&P assigned Sopaipilla Investor, LLC’s $12.3 billion notes a preliminary A+; Fitch assigned an expected AA-(EXP). The notes are secured by the issuer’s accounts, revenues, and equity interests in the joint venture, and exclude a direct mortgage on the land and buildings.

• Meta holds 20 percent of the El Paso joint venture and absorbs the construction, operating, and residual obligations that determine whether bondholders are repaid. Meta is contractually obligated to begin paying base rent in December 2028 whether or not the buildings are complete, and a residual value guarantee sized to outstanding principal triggers if Meta declines to renew.

• The El Paso structure is the second execution of a template, not a one-off, and the next hyperscaler SPV financing will be priced against how these two agencies resolved the same question. Meta’s Louisiana Hyperion campus used an identical 80/20 joint venture with Blue Owl Capital, raising roughly $27 billion through Beignet Investor.

The Security Package Is The Signal

Two rating agencies looked at the same $12.3 billion of paper and arrived one notch apart.

S&P Global Ratings assigned a preliminary A+ with a stable outlook.

Fitch Ratings assigned an expected AA-(EXP).

Most coverage recorded the split and moved on to the headline figure.

The split is the most useful disclosure in the transaction.

It is a disagreement about what actually secures these notes and by extension, what any subsequent hyperscaler SPV bond is worth.

Miss it, and you are pricing the next one off a number rather than off a structure.

The Missing Mortgage Explains The Notch

The notes are secured by the issuer’s accounts, revenues, and equity interests in the joint venture.

They exclude a direct pledge on the physical land and buildings.

A bondholder in default does not foreclose on a data center. They take the equity interests in the vehicle that owns one.

S&P’s A+ sits one notch below Meta’s own AA- corporate rating.

That discount is the price of the absent asset pledge.

Fitch reads the same structure and lands at AA-, level with Meta, on the view that the contractual transfer of construction, operating, and power supply risk to Meta makes the physical security question close to academic.

Both reads are defensible, which is what makes the split informative.

It marks the exact boundary of the off-balance-sheet template.

If the market ultimately prices toward Fitch, hyperscaler SPV paper trades as a synthetic corporate obligation of the tenant, and sponsors can raise larger amounts against thinner security. If it prices toward S&P, every subsequent structure carries a notch of drag that compounds across a $12 billion issue.

Over the next twelve to twenty-four months, the volume of paper coming from this template will settle the question in the secondary market rather than in the rating committee.

Meta Pays Rent Whether Or Not The Buildings Exist

The lease terms tell you where the risk actually landed.

Meta’s tenant subsidiary is obligated to begin paying base rent in December 2028, regardless of whether the campus is complete or operational.

Deferral applies only under force majeure, landlord funding defaults, or landlord-directed work stoppages.

Abatement under a landlord-directed delay is capped at $218 million and offset by delayed start-up insurance.

Meta funds construction cost overruns above 105 percent of the fixed construction budget.

Under the triple-net lease, Meta bears property taxes, utilities, insurance, routine maintenance, capital repairs, and structural replacements.

The leases carry no service level agreements tied to uptime or server performance, so there is no operational trigger that would let Meta withhold rent.

At lease expiry, a residual value guarantee sized to cover outstanding principal triggers if Meta declines to renew.

Read the package together and the credit becomes legible.

Bondholders are exposed to Meta’s willingness and ability to honor a lease, and to almost nothing else.

Construction risk sits with Meta. Operating risk sits with Meta. Obsolescence risk sits with Meta through the residual guarantee.

What the bondholder holds is a long-dated claim on the covenant of a single investment-grade technology company, wrapped in project finance documentation.

That is a materially different instrument than the infrastructure debt it resembles, and it should be underwritten as corporate credit with a real estate overlay rather than the reverse.

Investor Action

Private Capital. Infrastructure and credit funds evaluating hyperscaler SPV paper should underwrite the tenant covenant first and the asset second, because the security package makes the asset a secondary recovery path rather than a primary one.

The diligence question is not what the campus is worth in 2035 but what Meta’s credit is worth in 2035, and whether the residual value guarantee is enforceable against the entity that actually holds the cash.

Funds that continue to underwrite these as infrastructure assets will misprice the correlation in their books: five hyperscaler SPV positions across five sponsors are not diversified if three of them name the same tenant.

Waiting for a secondary market to establish where this paper trades means buying the next primary issue without a comparable.

Public Markets. Equity investors tracking Meta should treat the SPV lease obligations as debt-equivalent commitments that do not appear as debt.

The December 2028 rent commencement is a fixed cash obligation independent of whether the asset produces anything, and the residual value guarantee is a contingent principal-sized liability at expiry.

Credit investors holding Meta corporate paper are now structurally subordinated in a specific sense: the SPV bondholders have a direct lease claim on assets Meta operates, plus a guarantee, while corporate bondholders hold a general claim.

Model the aggregate lease and guarantee exposure across Hyperion and El Paso before the next campus is announced, because the disclosure will lag the commitment.

Operators. Data center developers and operators competing for hyperscaler tenancy now face a sponsor set that includes BlackRock, Blue Owl, and the institutional credit market rather than only their traditional capital providers.

The template rewards operators who can accept triple-net terms with no service level agreements and fund overruns above a fixed threshold.

Operators that cannot absorb that risk transfer will find themselves bidding against balance sheets that can.

The strategic decision is whether to compete for this capital or position around the assets it will not finance, and that decision gets harder each quarter the template is repeated.

The Verdict

The El Paso financing is the second execution of a structure that was novel eighteen months ago and is now standard.

Meta ran the same 80/20 joint venture in Louisiana with Blue Owl Capital, raising roughly $27 billion through Beignet Investor.

Two sponsors, two vehicles, one tenant, one template.

The inflection is not the size of any individual issue.

It is that hyperscaler capital expenditure has migrated from corporate balance sheets into a market where institutional credit investors hold the debt and the technology company holds the obligations.

That market now has to decide what the paper is.

The S&P and Fitch split is the first formal disagreement about the answer, and it will get resolved in secondary trading over the next twelve to twenty-four months, not by argument.

The open question is what happens to this structure when a tenant’s credit moves the wrong way.

Every protection in the El Paso package points back to Meta. None of them is worth more than Meta is.

Join Global Data Center Hub

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

Join Global Data Center Hub

TL;DR

• Sopaipilla Investor, LLC launched $12.3 billion of senior secured notes on July 23-24, 2026, to finance Meta’s El Paso data center campus. The issuer is a holding vehicle tied to Project Sopaipilla Holdings, LLC, the joint venture that owns the campus. Meta itself is not the issuer.

• BlackRock’s Global Infrastructure Partners and HPS Investment Partners hold an 80% equity stake in Project Sopaipilla Holdings, while Meta retains the remaining 20%. Meta serves as developer, operator, and sole tenant of the campus.

• S&P Global Ratings assigned the notes a preliminary A+ rating with a stable outlook, and Fitch Ratings assigned an expected AA-(EXP). JPMorgan Chase and Morgan Stanley ran fixed-income investor calls beginning July 22.

Sopaipilla Investor, LLC launched $12.3 billion of senior secured notes on July 23-24, 2026, to finance Meta Platforms’ data center campus in El Paso, Texas.

JPMorgan Chase and Morgan Stanley were mandated to run fixed-income investor calls beginning July 22, with pricing following in the days after.

Proceeds fund a roughly 1,000-acre campus in northeast El Paso targeting approximately 1 gigawatt of capacity by 2028. Meta broke ground on the site in October 2025.

Ownership Structure And Sponsor Composition

Funds managed by BlackRock’s Global Infrastructure Partners and HPS Investment Partners divisions hold an 80 percent equity stake in Project Sopaipilla Holdings, LLC.

Meta retains the remaining 20 percent. Meta serves as developer, operator, and sole tenant of the campus.

The structure keeps the debt off Meta’s own balance sheet while GIP and HPS carry the equity risk.

BlackRock acquired GIP for $12.5 billion and HPS for $12 billion, a combined roughly $24.5 billion.

The El Paso transaction brings both platforms into a single financing.

Ratings And Syndicate Composition

S&P Global Ratings assigned Sopaipilla Investor, LLC’s $12.3 billion senior secured notes a preliminary A+ rating with a stable outlook.

Fitch Ratings assigned the proposed notes an expected AA-(EXP) rating, one notch above S&P.

As of the marketing stage, full details on maturity, coupon, and final pricing had not been publicly disclosed.

Campus Scope And Power Supply

Meta broke ground on the El Paso campus, its 29th data center globally and third in Texas, in October 2025 with an initial $1.5 billion commitment.

That figure rose to more than $10 billion in March 2026 as the target capacity grew to roughly 1 gigawatt by 2028.

The site is planned across five phases.

In July 2026, Meta filed to expand the campus with 12 additional buildings totaling nearly 26 million square feet, with completion pushed to February 2029.

El Paso Electric supplies power alongside 813 onsite modular gas generators totaling 366 MW, a $473 million line item slated online by 2027.

The build is expected to draw over 4,000 construction workers at peak and support around 300 permanent jobs.

Meta has committed to matching 100 percent of the facility’s power consumption with renewable energy and to a closed-loop liquid cooling system.

The Louisiana Precedent

The financing follows the template BlackRock and Meta used in Louisiana.

Meta’s Hyperion campus was financed through an 80/20 joint venture with Blue Owl Capital, structured as roughly $27 billion in bonds anchored by PIMCO, rated A+ by S&P, maturing in 2049, plus about $2.5 billion in equity through an SPV named Beignet Investor.

BlackRock purchased over $3 billion of those Hyperion bonds when they priced in October 2025.

Pricing on the El Paso notes follows the investor calls that began July 22.

Join Global Data Center Hub

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

Join Global Data Center Hub

In this week’s issue:

First, the landlord became the lender. BlackRock is selling over $12 billion of bonds to fund Meta’s one-gigawatt El Paso campus through Project Sopaipilla Holdings, where BlackRock’s GIP and HPS units hold 80 percent and Meta holds 20.

Two days earlier the same firm closed a $40 billion acquisition of Aligned Data Centers. If the cost of a gigawatt now sits in a project entity, what is hyperscaler capex guidance actually telling you?

Then, contracted revenue found a cheaper path. Hut 8’s second 352 MW Beacon Point lease, worth $9.8 billion, lifted campus contract value to $19.6 billion without a hyperscaler balance sheet. TECfusions took the brownfield conversion model public at a $4.0 billion valuation.

Finally, sovereign compute got a proof point. Z.AI completed a one-gigawatt campus on entirely Chinese-made chips, reportedly, while Microsoft committed billions to Mistral’s European buildout.

Let us get into it.

THIS WEEK BY REGION

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

North America

Ownership separated from occupancy this week. BlackRock marketed more than $12 billion of bonds to finance Meta’s roughly one-gigawatt El Paso campus, holding 80% of the project while Meta retains 20% as tenant. Two days earlier, BlackRock, AIP, and Abu Dhabi’s MGX acquired Aligned Data Centers at a roughly $40 billion valuation.

Hut 8 reached $19.6 billion of campus contract value without hyperscaler equity, while OpenAI committed $20 billion to Project Camellia, its first campus as principal builder.

For public markets equity investors, the mechanism is liability placement, because a campus financed at the project entity never reaches the parent’s capex line.

Asia-Pacific

Domestic silicon moved from policy to proof. Z.AI completed a one-gigawatt data center built entirely on Chinese-made accelerators, reportedly, and has begun partial operation training its GLM models on clusters of more than 10,000 domestic chips each.

The China Mobile-led SEA-H2X cable entered service as Malaysia’s infrastructure constraints drew scrutiny. The next GLM release will reveal whether its domestic stack can compete.

For tech and venture investors, the next GLM release is the evidence that settles whether the domestic stack competes.

Europe

Sovereignty is being purchased rather than legislated. Microsoft committed billions to Mistral’s European expansion, then contracted its France-based GPU capacity to serve Azure customers in regulated industries, without taking a new equity stake.

Mistral is building on Nvidia Vera Rubin systems, making the operating layer French while the silicon remains American.

For real estate and infrastructure investors, sovereign-designated capacity is now a distinct product, and the question is whether the label survives contact with the supply chain underneath it.

South America

Power is moving behind the meter because the grid will not arrive in time. Terra Innovatum signed a non-binding agreement to deploy up to 8 MWe of SOLO microreactors at data centers in Brazil. The deal is small, but the pattern is clear.

For tech and venture investors in emerging markets, behind-the-meter generation is becoming the entry condition where grid capacity cannot be contracted on a project timeline.

NOTABLE TRANSACTIONS

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

BlackRock: over $12 billion of bonds for Meta’s El Paso campus

The bonds finance BlackRock's 80% stake in Project Sopaipilla, with Meta retaining 20% as tenant. Capital markets are now financing AI campuses as project credits backed by the asset and lease, not the parent's balance sheet.

For public markets investors, the mechanism is liability placement, and the spread at which these bonds clear will show whether the market prices Meta’s tenancy or the project itself as the credit.

AIP, MGX and GIP: $40 billion acquisition of Aligned Data Centers, plus $5 billion growth capital

The consortium acquired Aligned from Macquarie-managed funds at a roughly $40 billion enterprise value, with $5 billion for expansion across 51 campuses and 6.4 GW. Sovereign capital is now entering at platform scale, not through minority asset stakes.

For infrastructure funds, the mechanism is holding-period mismatch, because capital without a fund life competes for the same assets on a different clock.

Hut 8: $9.8 billion second lease commercializing the 1GW Beacon Point campus

The 15-year triple-net lease doubles contracted capacity to 704 MW with a 3.0% annual rent escalator. A secured interconnect and one investment-grade tenant created $19.6 billion of campus value without hyperscaler equity.

For operators, the mechanism is interconnection scarcity, and the party holding firm utility capacity sets the terms regardless of the size of the balance sheet across the table.

When the Owner and the Occupant Come Apart

The company running the AI does not own the building.

BlackRock proved that twice this week.

It is raising over $12 billion of bonds for a one-gigawatt Meta campus in El Paso.

The debt sits in Project Sopaipilla Holdings.

BlackRock’s GIP and HPS units hold 80 percent of that entity. Meta holds 20 percent and takes the compute.

Two days before, the same firm closed on Aligned Data Centers alongside AIP and MGX at approximately $40 billion, then committed $5 billion more.

Read at the surface, these are two large financings in one week. That reading misses what changed.

The cost of a gigawatt now sits at the project entity, not the parent’s capex. A Meta investor sees rent. A bondholder sees a one-gigawatt obligation. Both are accurate. Only one is complete.

The structure is not new. Meta used the same 80/20 joint venture at Hyperion, where the project raised a reported $27 billion of debt. El Paso follows the same template, with BlackRock now in the sponsor seat.

Now set that against Hut 8. Beacon Point reached $19.6 billion of base-term contract value this week on a second 352MW triple-net lease.

The tenant is investment-grade and doubled its footprint to 704MW. Hut 8 holds a 1,000MW interconnection agreement with AEP Texas.

No hyperscaler equity sits in that structure.

The contracted revenue is comparable in duration. The scarce input is the interconnect.

So here is the discipline.

Stop reading capex guidance as commitment. Find the project entity, ownership split, and lease term.

When underwriting an operator, ask what it controls that capital can’t replicate. A balance sheet can be raised in a week. A 1,000MW interconnection cannot.

The El Paso bonds will price shortly. If they price on Meta’s credit, the structure is presentation. If they price as project credit, the transfer is real and bondholders assume gigawatt-scale residual value risk.

Which one does the market believe, and what happens to every copy of this template if the answer is the second?

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

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

Have a great week.

— Obinna

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

Join Global Data Center Hub

TL;DR

• US data center power demand is projected to more than double from 31 GW in 2025 to 66 GW by 2027, while US generating capacity falls 50 to 80 GW short by 2030. Goldman Sachs projects data centers reaching 8.5 percent of US peak summer demand; BCG identifies gas-fired generation with carbon capture, up to $80 billion in US capex, as the most viable near-term firm-power bridge.

• The announced data center pipeline is materially inflated: historically only about 72 percent of facilities scheduled within a four-quarter window come online on time, and developers file across multiple regions while building only one site. Utilities estimate power delivery 1.5 to 2 years later than developers expect. CBRE reports Northern Virginia vacancy at an all-time low of 0.3 percent even as constraints stretch timelines.

• Brookings finds data center job creation is overstated by roughly a factor of three in industry-sponsored studies, and no report resolves whether AI demand converts to revenue. Colocation facilities deliver almost no local information-sector gains, which makes many local tax incentives poorly targeted. The deciding question for investors is how much announced capacity comes online on time, powered, and profitable.

We reviewed the 12 most consequential reports, investor analyses, and market outlooks shaping the global data center sector this quarter.

Each was selected for its impact on how operators deploy infrastructure, how investors allocate capital, and how policymakers frame the next phase of the AI buildout.

Across the period, three signals defined the market:

• AI is the new power load.

• Power is the new bottleneck.

• Capital is the new variable.

Together, they describe an industry running at full throttle yet increasingly constrained by grid capacity, execution risk, community opposition, and unresolved monetization.

This is your Data Center Intelligence Briefing: a field guide to what the latest research reveals, where consensus is forming, and where critical uncertainties remain.

Here is what is inside:

• A synthesis of the 12 reports shaping global data center strategy and how they redefine infrastructure, capital formation, and regional competition

• The key trends driving capital discipline, power strategy, and the shift toward AI-centric, inference-led infrastructure

• The opportunities emerging from power-secured development, clean-energy sourcing, and the telco AI Grid

• The issues still constraining the next wave, from grid access and permitting to labor, water, and chip obsolescence

A clear view of where the market is heading and what investors, operators, and policymakers still need to solve to sustain the AI infrastructure buildout.

The 12 Most Influential Reports

Below are the most impactful research reports shaping global data center and AI infrastructure strategy this quarter, ranked by strategic relevance, capital impact, and long-term market influence.

1. Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out

Source: Goldman Sachs

This report establishes the financial frame for the sector, sizing the AI build-out in the trillions of dollars.

It argues the total scale of investment is not fixed but hinges on four supply-side assumptions: hardware lifespan, data center complexity, chip architecture, and logistical bottlenecks.

Depreciation and power density can shift capital estimates by hundreds of billions, independent of demand.

Its core insight: the durability of the boom depends as much on these technical realities as on consumer adoption.

2. US Data Center Power Demand Projected to Double by 2027

Source: Goldman Sachs

Goldman projects US data center power demand more than doubling from 31 GW in 2025 to 66 GW by 2027, reaching 8.5% of total US peak summer demand.

It warns that supply chain shortages and labor deficits may delay nearly half of all planned facility activations.

The surge threatens regional grid stability and higher electricity costs, particularly in the Mid-Atlantic and Northwest.

3. Global Data Center Trends 2026

Source: CBRE

CBRE shows global supply reaching 16 GW across the 16 largest markets, up 25% year over year, yet average vacancy fell to 6.7% from 8.3%.

In the top four US markets, 80% of space under construction is already preleased, with Northern Virginia vacancy at an all-time low of 0.3%.

Power and grid constraints, plus public opposition, are curbing development timelines even as demand outruns record supply.

4. Global Data Center Market Comparison

Source: Cushman & Wakefield

Cushman reframes 2026 as a shift from expansion to managed growth, driven by financial self-regulation and stricter government oversight.

Developers increasingly pursue nuclear and private generation to bypass long utility wait times, while secondary and tertiary markets gain relevance on land and lighter regulation.

The report explicitly dismisses bubble fears in favor of disciplined, long-term profitability.

5. Solving the US Data Center Power Crunch

Source: Boston Consulting Group

BCG sizes a looming 50-to-80 GW US energy shortfall by 2030 and identifies gas-fired generation with carbon capture (CCUS) as the most viable near-term firm-power solution.

The approach lets developers secure fast, reliable, round-the-clock power while preserving low-carbon credentials, provided they move quickly on federal incentives.

It offers a roadmap prioritizing repeatable implementation and stakeholder engagement over one-off megaprojects.

6. 2026 Data Center Power Report

Source: Bloom Energy

Bloom Energy shows limited grid infrastructure fundamentally reorganizing the industry around a widening utility-developer timeline mismatch.

This is driving a strategic shift toward permanent onsite generation, gigawatt-scale campuses, and geographic reallocation toward power-rich regions like Texas.

It flags direct current distribution and other next-generation architectures as accelerating responses to the power boundary.

7. 2026 Asia Pacific Data Centre Trends & Outlook

Source: CBRE

CBRE examines how surging AI demand is reshaping APAC, with power scarcity and escalating construction costs as the defining bottlenecks.

It documents a geographic shift toward high-growth markets like India and Malaysia and the emergence of neocloud providers as a key demand source.

Advanced cooling and hyperscale expansion are reframing the region’s investment routes and locational dynamics.

8. The Outlook for Data Centers in Asia

Source: Goldman Sachs

Goldman identifies India, Japan, and the Philippines as high-potential growth markets while naming reliable power supply as the single largest obstacle to expansion.

It advocates public-private partnerships and innovative infrastructure designs that balance utility needs with sustainable cooling.

The report frames evolving customer requirements and regional regulation as the forces reshaping Asian capacity.

9. From Megawatts to Meaning: Why Cost per Token Matters in AI Data Centres

Source: Cushman & Wakefield

Cushman argues token-level metering is becoming the primary benchmark for performance and billing, replacing raw power capacity as the measure of work performed.

This gives operators a new competitive lever: tokens processed per megawatt, achieved through tighter integration of compute and networking.

The transition is complex given varying hardware efficiency but signals a structural repricing of the asset class.

10. New Evidence on Data Center Employment Effects

Source: Brookings

Using a synthetic control method, Brookings finds data center job creation is real but overstated by industry-sponsored studies by roughly a factor of three.

It draws a sharp distinction between hyperscale facilities, which build local technology ecosystems, and colocation facilities, which bring almost no local information-sector gains.

Its conclusion: local tax incentives are often poorly targeted, subsidizing the facilities with the smallest long-term benefit.

11. The AI Grid: A 2026 Playbook for Telco AI

Source: Nscale

Nscale outlines how telcos can evolve from data carriers into distributed AI infrastructure providers by converting last-mile networks and edge sites into a national AI Grid.

This solves the latency and data-sovereignty problems of centralized cloud, enabling secure, localized inference for regulated sectors like healthcare, defense, and finance.

It is a call to monetize existing real estate and power contracts through GPU-as-a-service and modular AI infrastructure.

12. Global Data Center Report 2026

Source: IDCA

IDCA provides the global baseline: capacity expanded to 67.7 GW, with data centers now consuming roughly 2% of world electricity and the US holding a dominant 43% share.

It reframes physical and digital security as equally vital against asymmetric threats in a geopolitically contested landscape.

The report serves as a strategic guide to a new era defined by energy constraints and immense capital intensity.

Key Trends

1. From Growth at All Costs to Capital Discipline

AI remains the primary driver of investment, but the narrative has shifted from scale to profitability. Investors are scrutinizing whether $4 trillion to $8 trillion in projected capex converts into repeatable free cash flow and pricing power.

These estimates are not fixed. They are highly sensitive to assumptions about chip lifespan and next-generation architecture costs.

In response, developers are recycling capital, spinning operational assets into fund and partnership structures to protect balance sheets and maintain liquidity.

2. Power as the Gating Variable

Access to reliable electricity has overtaken capital and land as the binding constraint on growth. Time-to-power is now the defining underwriting metric.

US generating capacity is projected to fall short by 50 to 80 GW by 2030, and utilities consistently estimate delivery will take 1.5 to 2 years longer than developers expect.

Developers are turning to onsite generation, storage, and hybrid energy, with more than a third of data centers projected to run on 100% onsite power by 2030.

3. The Phantom Pipeline and Execution Risk

The staggering figures for planned capacity are partly inflated. Developers file applications across multiple regions, then build only the single most favorable site.

Markets are suffering from elongation, the widening gap between capital deployment and live operations, driven by shortages of labor, transformers, and switchgear.

Historically, only about 72% of data centers scheduled within a four-quarter window actually come online on time. The headline pipeline deserves heavy discounts.

Key Opportunities

1. Power-Secured, High-Yield Development

Ground-up development offers the strongest returns where power can be secured early through onsite generation and long-term PPAs.

AI-focused facilities serving inference and regional demand are achieving higher yields despite greater capital intensity.

The clearest path to profitability currently runs through built-to-spec wholesale colocation, and in regulated markets, expanding existing facilities beats greenfield.

2. Clean Energy and the CCUS Bridge

The industry’s vocabulary shift from renewable to clean energy is a strategic admission that variable renewables cannot support 24/7 AI baseload.

Gas-fired power with carbon capture represents up to $80 billion in US capex, a natural opening for oil and gas majors seeking fast, reliable, low-carbon capacity.

Nuclear is emerging as the preferred scalable long-term solution, alongside immersion cooling and AI-driven workload optimization.

3. The Telco AI Grid and Sovereign Inference

Telcos hold the least-crowded opportunity in the set: converting distributed last-mile networks and edge sites into a national AI Grid.

Because they already operate trusted, regulated national networks, they are uniquely positioned to offer sovereign AI environments where sensitive data stays within borders.

The monetization window runs through GPU-as-a-service, localized inference platforms, and modular AI infrastructure.

Key Issues

1. Grid Deficiencies and Community Backlash

The 50-to-80 GW US shortfall and the 1.5-to-2-year time-to-power gap turn even funded projects into stranded capital.

US residential electricity rates rose roughly 32% between July 2020 and July 2025, largely driven by data center loads.

Local governments are responding with zoning restrictions, longer permitting, and moratoriums. The era of unchecked expansion is over.

2. Chip Obsolescence and Capital Risk

The short economic life of AI silicon creates transitional data centers that risk obsolescence within a few years of commissioning.

Facilities designed for 5-to-10 kW racks cannot support the 40 kW and higher densities modern AI workloads require.

This is why investors are repricing execution and demanding a credible path to free cash flow before committing capital.

3. The Unresolved Monetization and Jobs Questions

Every report assumes demand converts to revenue. None resolves what happens if enterprise adoption lags or model efficiency improves faster than expected, weakening utilization.

Brookings’ finding that job creation is overstated by 3x directly conflicts with industry studies and with more favorable reads on tax incentives elsewhere in the set.

The bubble question is the wrong question. The right one is narrower: of the capacity announced, how much will actually come online, on time, powered, and profitable?

Join Global Data Center Hub

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

Join Global Data Center Hub

TL;DR

• Data centers, the buildings where AI tools such as ChatGPT run, will require up to $3 trillion of investment by 2030, according to the real estate firm JLL. Morgan Stanley estimates that the technology giants can fund roughly $1.4 trillion of the bill from their own cash through 2028, leaving a gap of about $1.5 trillion for outside money to fill.

• Four kinds of money are building the AI boom: stock market investment, private buyout firms, loans secured against buildings and even chips, and direct investment by pension and sovereign wealth funds. Blackstone’s purchase of the Australian data center operator AirTrunk for more than A$24 billion and the cloud company CoreWeave’s borrowing against its NVIDIA chips show the second and third kinds at work.

• Many ordinary savers already own a piece of the AI buildout without choosing to. Index funds and public pension plans hold shares of Microsoft, NVIDIA, and data center landlords, and retirement funds such as Canada’s CPP Investments own data centers directly through joint ventures with operators like Equinix.

Every question you type into ChatGPT gets answered inside a building.

The building is real. It has a loading dock, and an electric bill closer to a small city’s than to an office’s.

Someone paid to put it there, and someone is paying for the thousands more now under construction.

The money behind those buildings comes from four sources that behave nothing alike.

The distinction is not academic.

The real estate firm JLL projected in its January 2026 outlook that the world will add nearly 100 gigawatts of data center capacity between 2026 and 2030, requiring up to $3 trillion of investment.

Morgan Stanley, in a July 2025 report, sized the bill through 2028 at roughly $2.9 trillion and estimated that the technology giants can cover about $1.4 trillion of it from their own cash, leaving a $1.5 trillion gap for everyone else to fill.

Even Microsoft, Google, Amazon, and Meta together can’t write a check that large.

Once you can tell the four kinds of money apart, the headlines about the AI boom become a map of who owns the physical layer under the AI you use every day.

1. The crowd’s money

Start with the money most people touch without noticing.

When savers buy shares of Microsoft, of NVIDIA, the company whose chips the AI boom runs on, or of Digital Realty, a landlord whose entire business is owning data centers and renting them out, they are pooling small amounts of money into the companies doing the building.

No single shareholder funds a specific site.

The crowd funds the company, and the company pours the money into concrete and chips.

The sums are hard to hold. Meta told investors in April 2026 that it expects to spend between $125 billion and $145 billion this year, most of it on AI infrastructure.

At the low end of that range, the company is spending more than $340 million every single day.

If you own an index fund through a retirement account, you never opted into the AI buildout, but you’re already in it. Part of your savings is the crowd’s money doing this work.

2. The building buyers

The second kind of money never touches a stock exchange.

Private equity firms pool money from large investors to buy whole companies, hold them for years, and sell them later.

A newcomer would reasonably assume the biggest data center deals involve the names everyone knows.

The largest one on record involved a company most people have never heard of.

In September 2024, the New York investment firm Blackstone, together with CPP Investments, the fund that manages the national retirement savings of Canadian workers, agreed to buy AirTrunk, an Australian data center operator, in a deal valuing the company at more than A$24 billion, roughly US$16 billion.

The purchase closed that December as one of the largest data center transactions ever completed. AirTrunk sells nothing to consumers.

It builds the buildings and rents them out at enormous scale to companies such as Microsoft and Amazon.

That’s the quiet rule of this kind of money: the logo on the app is often only the tenant.

If you assumed the company whose name is on your screen also owns the building your data sits in, this is the correction, because much of the internet’s real estate belongs to owners whose names you’ve never seen.

3. The mortgage on the internet

The third kind of money is borrowed.

Data centers carry loans the way houses carry mortgages. The building, the long customer contracts inside it, and the rent they produce stand behind the loan as collateral, the property a lender can take if the loan is not repaid.

Morgan Stanley’s July 2025 report estimated that this kind of direct lending to data centers could reach $800 billion.

Then AI produced a stranger version. In 2023, CoreWeave, a company that buys NVIDIA chips and rents out their computing power, borrowed $2.3 billion by pledging the chips themselves as collateral, Reuters reported at the time.

By 2026 the practice had matured enough that CoreWeave closed an $8.5 billion loan package it described as the first chip-backed financing to earn an investment-grade rating, the credit score that marks a loan as relatively safe.

Sit with that for a moment.

A computer chip loses value quickly as newer models arrive, the way a new car loses value the moment it leaves the lot.

Lenders are betting that the rent those chips earn comes in faster than the chips age.

If lenders ever stop believing that, the buildout slows, and this is the part most people miss: the pace of the AI boom depends on lender confidence as much as on how many people use AI.

4. The savings of whole nations

The fourth kind of money skips the funds and the stock market and shows up at the negotiating table itself.

Sovereign wealth funds, the investment accounts nations use to grow their public savings, and the world’s largest pension funds now build data centers directly, side by side with the operators.

In October 2024, Equinix, one of the world’s biggest data center operators, formed a joint venture worth more than $15 billion with GIC, the fund that invests Singapore’s national reserves, and CPP Investments, the Canadian retirement fund again.

The two funds together own three quarters of the venture, which is adding more than 1.5 gigawatts of capacity in the United States, more electricity than a million homes draw.

Weeks earlier, BlackRock, the world’s largest money manager, had launched a partnership with Microsoft and MGX, an Abu Dhabi investment firm created to fund AI, targeting $30 billion for data centers and the power plants behind them.

By January 2026 it had raised more than $12.5 billion, and NVIDIA had joined as a partner.

AI reads as a California story. Follow the ownership and the map widens to Singapore and Abu Dhabi, and to the retirement savings of millions of Canadian workers, because their funds now hold the ground the internet sits on.

What Abilene, Texas traded for its giant AI campus

All four kinds of money eventually land in a real town.

In Abilene, Texas, a campus known as the Stargate site began as a $3.4 billion joint venture announced in October 2024, built around a 206-megawatt facility designed to run as many as 100,000 AI chips.

Public reporting now puts the campus at 1.2 gigawatts across eight buildings, close to the electricity draw of a million American homes, with OpenAI, the maker of ChatGPT, running its workloads there on hardware owned by Oracle.

Abilene’s side of the bargain has a documented cost. The city granted an 85 percent property tax break to land the project, and Business Insider reported in July 2025 that Oracle disputed the county’s valuation of the site to lower its bill further.

New York moved the other way: on July 14, 2026, the governor signed a one-year pause on new data centers of 50 megawatts or more, roughly the electricity draw of 40,000 homes, while the state studies their effect on electricity prices and the grid.

The same money that makes these projects possible is now meeting towns and states deciding what the projects are worth to them.

The foundation under the feed

None of this is visible from your phone. The answer arrives in a second, and the screen never shows the loading dock or the loan documents behind it.

These buildings are the physical foundation of the digital life you already live, and the four kinds of money are simply the world deciding, in four different ways, that the foundation is worth paying for.

The next time a headline announces a pension fund buying a data center company, or a state pausing construction, you’ll be reading the same story this piece just told: the financing of the physical world underneath your digital one.

A question for you: if part of your retirement savings is helping to build AI data centers, and it probably is, how do you feel about that?

Join Global Data Center Hub

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

Join Global Data Center Hub

TL;DR

• New York’s Executive Order 62 froze permitting for data centers above 50 MW while 9,682 MW of data center load sat in the NYISO interconnection queue. A state with available grid capacity chose to withhold it rather than accelerate it.

• The order shifts grid upgrade costs onto developers through beneficiary-pays allocation: data centers fund transmission, distribution, and network upgrades directly, ending the fifteen-year model that socialized interconnection cost across the ratepayer base and made cheap-power markets pencil.

• The repricing extends beyond New York: discretionary-permit projects face a one-year delay floor with no confirmed ceiling, and as the state positions the framework as a national model, the cost-allocation structure is positioned to spread to Virginia and Texas, where far more hyperscale capacity is at stake.

Permitting Risk Now Prices Alongside Power

The signal is not the pause. The signal is that a state with grid capacity chose to withhold it.

New York held roughly 9,682 MW of data center load in its interconnection queue and answered that demand with a permitting freeze, not an acceleration.

Mainstream coverage read Executive Order 62 as an environmental measure. It is a cost-allocation measure.

The states that follow will not copy the environmental language.

They will copy the mechanism that shifts grid upgrade costs onto the developer.

Firms that underwrote New York based on power prices alone now hold assets whose feasibility depends on a variable they never modeled.

The 50 MW Threshold Targets Capital, Not Carbon

Governor Hochul set the moratorium threshold at 50 MW.

The legislature’s competing bill set it at 20 MW.

The gap between those two numbers is the entire argument. A 50 MW floor insulates back-office financial services, hospitals, and academic load while capturing only hyperscale-scale demand.

That is not an environmental line.

Environmental impact does not begin at 50 MW.

The threshold is calibrated to isolate the specific projects that carry grid-upgrade cost and community-sentiment risk, and to hand local governments a state-provided framework to extract concessions once the freeze lifts.

The number tells you the policy targets capital, not carbon. Underwrite the next state’s threshold as the tell for who it intends to tax.

Beneficiary-Pays Ends the Ratepayer Subsidy

The order directs data centers to fund grid infrastructure directly and pushes toward a dedicated utility rate class.

Doc 2’s supporting material frames the mechanism as beneficiary-pays: the developer bears one hundred percent of transmission, distribution, and network upgrade cost.

For fifteen years the siting model assumed the utility socialized interconnection cost across the ratepayer base.

That subsidy is what made cheap-power markets pencil.

New York is removing it and calling the framework a national model.

The forward read for the next 12 to 24 months is that interconnection cost migrates from an externality onto the project’s own capital stack, and the markets that adopt the rate-class structure will reprice every pro forma built on socialized grid cost.

Investor Action

Private Capital. Infrastructure funds and private equity holding New York-exposed pipeline should diligence permit status now, not at the next milestone.

The freeze shields projects already holding complete permits or under construction.

Everything in the discretionary-permit stage carries a one-year timeline floor with no confirmed ceiling, because the lift depends on DEC finalizing standards, and the crypto precedent shows DEC missing its statutory deadline by more than a year.

Re-underwrite queued New York positions with a two-year delay case and price the beneficiary-pays upgrade cost into the equity return.

The cost of waiting is holding an asset at a power-cost basis the state has already decided to raise.

Public Markets. Public equity investors holding hyperscale and REIT names with New York disclosure should benchmark exposure against the 9,682 MW that now sits behind a regulatory floor.

The immediate earnings impact is contained, because New York was never a primary hyperscale destination.

The mispriced risk is contagion.

As sponsors position this as a national model, expect the cost-allocation framework to spread to major hyperscale markets like Virginia and Texas, where far more capacity is at stake.

Waiting for that repricing to appear in a filing means buying it after the market has.

Operators. Developers and hyperscalers acting as buyers must now underwrite legislative stability and community sentiment as siting inputs alongside grid access.

The end of regulatory uniformity, in DLA Piper’s framing, means power availability alone no longer determines feasibility.

Sequence new site selection toward jurisdictions that have declared their posture and treat undeclared states as carrying latent moratorium risk.

Structure community-benefit contributions into the base case rather than negotiating them as concessions later.

A proposed surcharge on the order of one million dollars per megawatt, drawn from Doc 2 and treated here as an estimate rather than a confirmed term, is the number to model as the emerging floor.

The Verdict

New York did not pause data centers. It repriced them.

The order converts a permitting process into a cost-allocation negotiation and packages the result as a template.

The long-term significance is that the AI infrastructure build-out just acquired a second binding constraint.

Power was the first. Permitting is the second, and unlike power it does not respond to capital.

The market inflection is the shift from underwriting energy cost to underwriting regulatory posture, and the firms that build that discipline now will diligence the next moratorium before it is signed.

The open question is which state declares next, and whether its threshold targets carbon or capital.

Join Global Data Center Hub

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

Join Global Data Center Hub

TL;DR

• Amazon is funding its AI buildout with debt: it raised $79 billion in 2026, lifting long-term debt to $119.1 billion, while free cash flow fell 95 percent to $1.2 billion. Amazon is the only hyperscaler financing the capex gap through the bond market; Alphabet self-funds and Microsoft relies on operating cash flow. AWS backlog nearly doubled to $364 billion, concentrating demand visibility in a handful of frontier AI customers.

• Meta added roughly $107 billion of non-cancelable purchase obligations in one quarter, bringing total commitments to about $238 billion across GPUs, cloud capacity, and energy. Meta halted buybacks after repurchasing $12.75 billion a year earlier, redirecting shareholder returns to infrastructure. The obligations remove Meta’s option to spend less if advertising, its only revenue engine, weakens.

• Capital structure, not capex scale, is now the deciding variable in AI infrastructure: Amazon borrowed its funding gap while Meta contracted its away, and the market priced the divergence — Amazon gained less than a point while Meta fell 8.6 percent despite both beating expectations. Combined 2026 hyperscaler capex approaches $700 billion; from 2027 that spending hits income statements as depreciation, a cost the market has not yet priced.

For two years the AI infrastructure debate fixed on a single axis: how large the hyperscaler capex numbers could get before the market flinched.

Amazon at $200 billion, Alphabet at $180 to $190 billion, Microsoft near $190 billion, Meta at $125 to $145 billion. Scale was the whole narrative.

That framing is now exhausted.

The cohort has converged on comparable spending, combined 2026 capex approaches $700 billion, and backlog has answered the demand question definitively.

What remains unresolved is more structural: not how much these companies will spend, but how they will fund it and which balance sheets can absorb a multi-year build.

Amazon and Meta answered that question in opposite ways.

Amazon financed the gap through the bond market, adding leverage.

Meta contracted it away through non-cancelable obligations, sacrificing flexibility.

Neither preserved its optionality.

That divergence, not the capex totals, is what the market priced when Amazon gained less than a point while Meta fell 8.6% despite both beating expectations.

Amazon: Free Cash Flow Went to Zero by Design

On the surface, Amazon delivered a standout quarter.

Revenue rose 17% to $181.5 billion, operating margin hit a record 13.1%, and diluted EPS of $2.78 beat expectations, boosted by a $16.8 billion pre-tax gain on its Anthropic stake.

Excluding that gain, operating income grew 29.6%.

The real signal was cash flow. Free cash flow fell 95% to $1.2 billion even as operating cash flow rose 30% to $148.5 billion, reflecting $44.2 billion in quarterly capex. Nearly all operating cash was reinvested into infrastructure.

The differentiation is at the silicon layer.

Graviton, Trainium, and Nitro reached a $20 billion annualized run rate, while Anthropic and OpenAI committed roughly 7 GW of future Trainium capacity, reinforcing Amazon’s proprietary-chip strategy.

Here the cohort diverges. Alphabet is self-funding, Microsoft relies on operating cash flow, and Amazon alone is funding the gap with debt. It raised $79 billion in 2026, increasing long-term debt to $119.1 billion.

Amazon has traded free cash flow flexibility for build speed.

AWS revenue grew 28% to $37.6 billion, while remaining performance obligations nearly doubled to $364 billion.

The backlog strengthens demand visibility but also increases reliance on a handful of frontier AI customers.

Meanwhile, AWS operating margin fell to 37.7% as accelerated infrastructure depreciation outpaced monetization.

Read the full Amazon analysis: Zero Free Cash Flow, $79B in New Debt and the Financing of AI Infrastructure

Meta: The Signal Is in the Commitments Footnote

Meta reported GAAP diluted EPS of $10.44, boosted by an $8.03 billion one-time tax benefit.

Excluding it, normalized EPS was $7.31, a 9% beat driven by advertising.

Revenue rose 33% to $56.31 billion, its fastest growth in nearly five years, as stronger ad pricing and engagement not user growth drove monetization.

The more consequential disclosure was contractual.

Meta added roughly $107 billion in non-cancelable purchase obligations, bringing total commitments to about $238 billion across GPUs, cloud capacity, and energy.

The significance is not higher spending but the loss of optionality to spend less.

The commitment structure highlights Meta’s priorities.

It is expanding custom MTIA accelerators with Broadcom while securing off-grid power and long-duration storage, signaling that power not capital is the binding constraint.

In AI infrastructure, energized megawatts determine deployment speed.

Capital expenditures rose 45% to $19.84 billion as full-year guidance increased to $125–145 billion.

Operating cash flow funded the spending, but free cash flow narrowed to $12.39 billion.

More telling, Meta halted buybacks after repurchasing $12.75 billion a year earlier, redirecting capital from shareholder returns to infrastructure while maintaining modest leverage.

Meta occupies a higher-variance position. Unlike Microsoft, Amazon, and Alphabet, which fund infrastructure with diversified cloud revenues, Meta relies almost entirely on advertising.

That creates a faster AI monetization loop but no fallback. If ad pricing weakens, there is no second revenue engine to support a $238 billion contractual load.

The investment case rests on a single demand curve.

Read the full Meta analysis: The $107B Quarter That Turned Buybacks Into a Liability

The Pattern: Optionality Is Being Contracted Away

The two quarters rhyme, but the common thread is not spending it's the loss of flexibility.

Alphabet turned to self-generated power, Microsoft capitalized component inflation, Amazon replaced free cash flow with leverage, and Meta locked in $107 billion of obligations.

Each turned uncertainty into fixed costs before returns were visible.

The constraint stack is the same for all four: advanced-packaging silicon, memory, liquid cooling, transformers, and grid interconnection.

Power remains the binding constraint, dictating deployment order.

As hyperscalers compete for the same scarce inputs, costs rise, driving repeated capex revisions, including Meta's memory-driven increase. Each capacity commitment makes the next increment more expensive.

For emerging markets, the implication is exclusionary.

Amazon's model adds a capital filter, requiring stable rates and deep bond markets, while Meta's adds a demand filter, relying on sustained pricing power from a single monetization engine.

As a result, buildout is concentrating in the United States, select European hubs, and parts of Asia-Pacific, where power, water, regulatory stability, and capital markets align.

Regions without secured grid capacity and financing access are increasingly priced out, reinforcing concentration rather than broadening global expansion.

What Resolves Next

Three signals will define the next several quarters.

First, whether Amazon’s free cash flow recovers against its debt trajectory.

Free cash flow should recover as AI workloads monetize and the build cycle normalizes; whether that lands in 2026 or 2027 determines how long Amazon carries elevated leverage.

If capex stays near $200 billion and debt issuance resumes, the market will treat the balance sheet, not demand, as the constraint.

Second, whether Meta’s monetization layer expands fast enough to service commitments that are now contractually fixed.

Roughly ten million weekly conversations across its business AI agents is a usage signal; a revenue line tied to them would be thesis confirmation, and its absence through 2026 would compress the multiple further.

Third, the depreciation curve from 2027 forward, when the combined effect of the cohort deploying roughly $700 billion in 2026 capex hits income statements simultaneously.

That accounting expression of the current cycle has not yet been priced.

The Bottom Line

The demand story is settled. A $364 billion backlog at Amazon and a $238 billion contractual load at Meta confirm it.

What remains is whether capital structure compresses returns faster than AI workloads monetize.

Capital structure was the first deciding variable in the Microsoft and Alphabet prints.

Amazon and Meta showed the two ways a balance sheet absorbs it: borrow the gap, or contract it away. Neither is free, and neither travels.

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 hyperscaler capital cycle has passed the point where spending totals carry information.

Combined 2026 AI infrastructure capex across the four largest builders approaches $700 billion, and the market has absorbed the headline.

What it has not priced is the mechanism by which that spending converts from discretionary budget into fixed obligation.

A capex plan can be trimmed, deferred, or re-phased in response to demand. A non-cancelable purchase obligation cannot.

Meta’s Q1 2026 results highlight the shift: despite strong growth and tax-boosted earnings, the company paused buybacks and added $107 billion in new multi-year commitments.

The moment matters not because Meta spent more, but because it removed its own optionality to spend less.

The Signal Is in the Commitments Footnote, Not the EPS Line

Meta reported GAAP diluted EPS of $10.44 for Q1 2026, beating consensus by more than 55%, but the result was boosted by an $8.03 billion one-time tax benefit under U.S. Treasury Notice 2026-7.

Excluding that benefit, normalized diluted EPS was $7.31, a roughly 9% beat driven by the strength of Meta’s advertising business.

The underlying performance was solid, but far less dramatic than the headline figure suggested.

Revenue reached $56.31 billion, up 33% year over year and 29% in constant currency, the fastest growth in nearly five years.

The data shows AI is directly improving Meta’s ad engine: ad impressions rose 19%, ad prices increased 12%, and revenue per user grew 27% to $15.66 despite only 4% user growth.

Monetization intensity, not audience expansion, is the primary lever. The engine is working.

The more consequential disclosure came elsewhere.

During the quarter, Meta added approximately $107 billion in new non-cancelable purchase obligations, bringing total contractual commitments to roughly $238 billion.

These legally binding, multi-year liabilities cover GPUs, cloud capacity, and energy procurement.

Unlike a buyback pause, commitments of this scale cannot be easily reversed, marking a shift from planned spending to fixed obligations.

Buildout Sequencing Signals That Grid Access Is Already the Ceiling

The commitment structure reveals where Meta believes the binding constraints lie.

Meta is co-developing its MTIA accelerator with Broadcom, adding over one gigawatt of custom compute through 2029 to complement NVIDIA and AMD while reducing GPU cost exposure.

The energy disclosures are the more revealing tell.

Alongside conventional buildout, Meta has reached for up to one gigawatt of prospective space solar capacity and more than 100 gigawatt-hours of long-duration terrestrial storage.

A company secures speculative off-grid generation only when it has concluded terrestrial grid procurement is insufficient at its planning horizon.

That is a statement about sequencing: power is solved first, ahead of land, silicon, and networking, because energized megawatts govern how quickly contracted compute converts to revenue.

Capital is the enabler; it is no longer the ceiling.

The Free Cash Flow Inversion

Capital expenditures reached $19.84 billion in the quarter, up 45% year over year and back-weighted against full-year guidance raised a third time in twelve months to $125–145 billion.

Operating cash flow of $32.23 billion covered the outlay, compressing free cash flow to $12.39 billion.

The near-term coverage is not the concern; the trajectory is.

Full-year 2025 free cash flow of $43.6 billion is projected by sell-side estimates to compress toward zero in 2026 and turn negative in 2027 as the $135 billion capex midpoint absorbs operating cash.

Against that, the buyback decision is the analytically important one.

Meta repurchased $12.75 billion of stock in the first quarter of 2025 and nothing in the first quarter of 2026.

Capital previously supporting the stock during downturns has been redirected into compute and is unlikely to return during the build cycle.

A modest $1.35 billion quarterly dividend remains, but buybacks are paused. With long-term debt stable at $58.7 billion, Meta has sacrificed shareholder-return flexibility to fund infrastructure growth.

A Single Monetization Vehicle Carrying the Entire Infrastructure Base

Meta occupies a distinct and higher-variance position within the cohort.

Microsoft, Amazon, and Alphabet deploy comparable capital against diversified enterprise cloud revenues, where third-party demand underwrites the buildout through contracted backlog.

Meta is deploying its entire infrastructure base against a single monetization engine: advertising. That concentration cuts both ways.

The ROI feedback loop is more direct than any peer’s, because better models measurably improve ad ranking, which lifts pricing and conversion, which funds more compute.

But the same concentration removes the fallback.

A cloud hyperscaler facing softer AI demand can lean on general-purpose compute and enterprise workloads to absorb depreciation.

Meta has no second revenue line at scale to cushion a $238 billion contractual load if advertiser return deteriorates or engagement inventory stalls.

The entire capital thesis rests on a single demand curve, and that curve answers to ad pricing.

The Pattern Across the Cohort: Optionality Is Being Contracted Away

Meta’s quarter rhymes with signals across its peers, and the through-line is not spending magnitude but the surrender of flexibility.

Alphabet paused buybacks and moved to internalize energy generation; Microsoft formalized component inflation as a distinct capital line; Meta locked $107 billion of obligations in three months and switched off repurchases.

Each converts a variable into a fixed cost placed before the return is observable.

The constraint stack they all bid against is identical: advanced-packaging silicon, memory, liquid cooling, transformers, and grid interconnection.

When four firms contract simultaneously for the same scarce inputs, component pricing inflates, which is what has driven successive capex revisions industry-wide, including Meta’s.

The act of committing capacity raises the cost of the next increment.

For emerging markets the implication is exclusionary: buildout concentrates in the United States, select European hubs, and parts of Asia-Pacific, where baseload power, water, and regulatory continuity exist.

Regions without secured grid capacity are priced out of the frontier by power scarcity and supply-chain allocation that favors incumbents with the balance sheets to pre-commit.

What Resolves Next

The forward question is no longer whether AI demand justifies the capex; Meta’s advertising results and the cohort’s backlog have answered it.

The question is whether the monetization layer expands fast enough to service commitments that are now contractually fixed.

Three signals will define the next several quarters.

First, actual second-quarter capex against the back-weighted guide, since the first quarter’s $19.84 billion implies a steep ramp.

Second, the free cash flow trajectory, which will show whether operating cash can buffer the build without material debt expansion beyond $58.7 billion.

Third, any pricing model for Meta’s business AI agents, now handling roughly 10 million weekly conversations, could signal a second revenue stream beyond advertising.

One forward-looking development is Meta potentially selling surplus compute to external enterprises.

If realized, this could diversify its revenue base by turning excess infrastructure into recurring income, though it remains unproven.

The broader shift is clear: hyperscalers are no longer just building capacity they are securing scarce inputs.

Meta’s quarter did not change the AI demand thesis; it showed the cost of committing to it early by sacrificing financial flexibility.

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

For two years the AI infrastructure debate fixed on a single axis: how large the hyperscaler capex numbers could get before the market flinched.

Amazon at $200 billion, Alphabet at $180 to $190 billion, Microsoft near $190 billion.

That framing is now exhausted.

The cohort has converged on comparable spending, and backlog has answered the demand question.

What remains unresolved is quieter and more consequential: not how much these companies will spend, but how they will pay for it, and which balance sheets can absorb a multi-year build without a financing event at the wrong point in the rate cycle.

Amazon’s first quarter of 2026 makes that shift legible.

On the surface it is one of the strongest quarters the company has printed.

Underneath, it is a disclosure about capital structure: Amazon has become the only hyperscaler funding its infrastructure gap through the bond market rather than internal cash, accepting the leverage and margin consequences.

The quarter does not raise a question about demand. It raises one about financing durability.

The Signal: Free Cash Flow Went to Zero by Design

Amazon reported $181.5 billion in revenue, up 17 percent, with operating income of $23.9 billion and a record consolidated operating margin of 13.1 percent.

Diluted earnings per share of $2.78 against a $1.63 consensus is an optical beat: it carries a $16.8 billion pre-tax gain on Amazon’s Anthropic stake, a non-operating item.

Strip it out and operating income growth of 29.6 percent is the substantive result.

But the earnings headline is not where the signal lives. It lives in cash flow.

Trailing twelve-month free cash flow collapsed to $1.2 billion, a 95 percent decline from $25.9 billion a year earlier, while operating cash flow over the same period grew 30 percent to $148.5 billion.

This is not operational deterioration; the entire delta is capital expenditure.

Amazon deployed $44.2 billion in the quarter alone, up 77 percent, against trailing property and equipment purchases of $147.3 billion.

A company generating $148 billion in operating cash converted almost none of it to free cash flow because it is routing the whole of that generation, and more, into physical capacity.

That is a capital allocation decision surfaced as a cash flow statement.

Infrastructure Strategy: Vertical Silicon as Margin Control

The differentiation sits at the silicon layer. Amazon’s Graviton, Trainium, and Nitro families reached a $20 billion annualized run rate, growing at triple-digit rates.

The Trainium3 UltraServer also began shipping at rack-scale density comparable to merchant Blackwell hardware.

The logic is margin capture: every workload on proprietary silicon rather than merchant GPUs retains economics that would otherwise flow to Nvidia.

That logic is visible in the compute Amazon has locked.

Anthropic secured up to five gigawatts of Trainium capacity, while OpenAI committed to roughly two gigawatts beginning in 2027 under broader arrangements that include a reported $50 billion Amazon investment and a $100 billion contract expansion.

That investment figure appears in the quarter’s materials and should be confirmed against independent disclosure before it is treated as settled.

The two-gigawatt commitment, however, anchors the competitive read regardless.

Capital Allocation: The Balance Sheet Does the Work

Here the cohort separates. Alphabet paused buybacks entirely, dropping repurchases to zero from $15.1 billion a year earlier, to self-fund from internal cash.

Microsoft finances its program out of a 47 percent operating margin.

Amazon alone is bridging the gap with debt: it issued $54 billion in corporate paper in March 2026 and a $25 billion investment-grade offering in July, driving long-term debt to $119.1 billion from $65.6 billion at year-end 2025.

The posture is defensible on the assets being financed, since data center shells carry roughly thirty-year lives and server assets are tied to contracted backlog, and Amazon holds $143.1 billion in cash and short-term investments.

But the choice is directional: Amazon has traded buyback flexibility and free cash flow optionality for build velocity, carrying the cost as leverage rather than forgone capital returns.

That is the deciding variable the cohort now competes on, and Amazon has taken the most aggressive position within it.

Competitive Positioning: Reacceleration Against a Compressing Margin

AWS delivered $37.6 billion in revenue, up 28 percent, its fastest growth in fifteen quarters, on a run rate near $150 billion.

Remaining performance obligations reached $364 billion, up 93 percent.

That backlog is the demand answer, but it carries a concentration signal that belongs in the same sentence: the incremental is driven heavily by a small number of frontier-lab counterparties, so forward visibility and counterparty concentration are rising together.

A backlog that nearly doubles in a year is strength; a backlog whose growth depends on two or three names is strength with a specific fragility.

The margin tells the more instructive story. AWS operating margin declined to 37.7 percent from 39.2 percent a year earlier.

This is not weakness; it is the depreciation of accelerated infrastructure hitting the operating expense line before the compute is monetized. Racks deploy, the depreciation charge lands immediately, and the revenue arrives over the contract life that follows.

For investors, that is the tension to underwrite: reacceleration and margin compression are the same event viewed from two lines of the income statement.

The System Pattern: Power First, Financing Second, Everything Else After

Across the cohort a single sequencing logic has hardened.

Combined 2026 hyperscaler capex approaches $700 billion, concentrated in a tight constraint stack of advanced packaging, accelerators, liquid cooling, transformers, and grid power.

Power is the binding constraint, and it governs the order of execution: power first, then land and interconnects, then silicon allocation, then networking and cooling.

Amazon added roughly four gigawatts in 2025 with plans to double by 2027.

The geographic consequence matters for emerging markets. Buildouts cluster where grids are reliable and regulatory continuity is long-term.

Amazon’s model adds another filter: a build financed partly through investment-grade debt depends on stable rate environments and deep bond markets, conditions absent across most regions competing for this capacity.

The cycle is not broadening globally. It is concentrating, and the financing structure reinforces the concentration.

What Resolves Next

Three signals will define the next several quarters.

The first is the third-party Trainium decision. Reports suggest Amazon is in early talks to sell Trainium racks directly to external data centers, ending AWS-exclusive distribution.

While not yet a disclosed transaction, it is a key forward signal. If it proceeds, Amazon moves into direct competition with Nvidia.

The second is free cash flow recovery against the debt trajectory. Free cash flow should recover as AI workloads monetize and the build cycle normalizes.

Whether that happens in 2026 or 2027 will determine how long Amazon carries elevated leverage.

If capex stays near $200 billion and debt issuance resumes, the market will view the balance sheetnot demand as the constraint.

The third is time-to-energize. Power-secured land is the appreciating asset, while Amazon's debt-backed build concentrates AI's expansion in one credit profile and one grid pathway.

Amazon did not report a strong quarter with a cash flow problem.

It reported a decision: to finance infrastructure on the balance sheet, separating itself from peers funding the same future out of pocket.

The demand story is settled by a $364 billion backlog.

What remains is whether capital structure compresses returns faster than AI workloads monetize.

Join Global Data Center Hub

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

Join Global Data Center Hub

TL;DR

• Meta finances Hyperion, its 5GW Louisiana data center campus, through a joint venture in which Blue Owl holds 80 percent and Meta 20 percent. Meta contributes $5.8 billion of equity against Blue Owl’s $23.0 billion, and the vehicle issues $27.3 billion of SPV debt that stays off Meta’s balance sheet, converting capex into a lease obligation that begins June 1, 2029.

• Meta’s payment guaranty and Residual Value Guarantee price the non-recourse debt at roughly 225 basis points over Treasuries with an A+ rating — one notch below Meta’s own AA-. A single notch is the market’s full price for moving tens of billions off balance sheet, which makes the structure a template Microsoft, Amazon, and Alphabet are likely to replicate as their balance-sheet capacity tightens.

• PIMCO anchored the bond as lead investor with BlackRock participating, at a 2049 maturity that treats Meta’s occupancy as a 25-year annuity. The deal moves single-tenant data center debt from project finance toward corporate-guaranteed infrastructure credit, opening the asset class to fixed-income allocators it could not previously reach.

The Structure Is The Signal, Not The $50 Billion

The number that matters in Hyperion is not $50 billion. It is 80.

Blue Owl owns 80 percent of the joint venture that owns the asset Meta will occupy, and that ownership split is the mechanism that keeps tens of billions of dollars of debt off Meta’s balance sheet.

Mainstream coverage has anchored on the headline capex and the gigawatt count.

The structural event is that Meta has built a repeatable template for financing gigawatt-scale infrastructure with institutional credit rather than corporate cash, and the template will be copied before the first hall is energized.

The Off-Balance-Sheet Split Converts Capex Into Rent

The equity split does specific work. Meta contributes $5.8 billion against Blue Owl’s $23.0 billion, and the joint venture raises $27.3 billion in SPV-issued debt against that equity base.

The debt sits in the vehicle, not on Meta’s books.

Meta converts what would have been a consolidated capital liability into a lease obligation that begins June 1, 2029.

This is the difference between a hyperscaler that must fund gigawatts from operating cash flow and one that can underwrite gigawatts against institutional appetite for long-dated investment-grade paper.

CFO commentary has already flagged new ownership structures for large sites as a strategic priority.

Hyperion is not a one-time structure.

It is the first execution of a financing model Meta intends to run repeatedly, and the next 12 to 24 months will show whether Microsoft, Amazon, and Alphabet adopt the same mechanism as their own balance-sheet capacity tightens.

Non-Recourse Plus A Residual Value Guarantee Reprices The Risk

The risk transfer is engineered with precision.

Creditors have no direct recourse to the underlying data center assets, which would ordinarily push the debt toward higher yields.

Meta offsets that with a payment guaranty and a Residual Value Guarantee covering outstanding SPV debt at the end of each four-year renewal term.

The result is a 144A bond priced at roughly 225 basis points over Treasuries and rated A+, one notch below Meta’s own AA-.

The structure extracts near-corporate pricing from a non-recourse asset by attaching Meta’s credit to the tail risk without consolidating the principal.

That one-notch gap is the price the market assigns to the structural separation, and it is narrow.

When the discount for moving debt off the balance sheet is a single notch, the incentive to replicate the structure is overwhelming.

The Anchor Tenant Model Migrates From Real Estate To Compute

PIMCO anchored the debt as lead investor, with BlackRock and other institutional credit investors participating.

This is the signal underneath the syndicate: the largest fixed-income allocators are now underwriting single-tenant data center paper as a long-duration credit asset, priced against the tenant’s guarantee rather than the asset’s liquidation value.

The 2049 maturity tells you the holding thesis.

These allocators are treating Meta’s occupancy as a 25-year annuity.

That reframes data center credit from project finance toward something closer to corporate-guaranteed infrastructure debt, and it opens the asset class to a pool of capital that project-finance structures could not previously reach.

Investor Action

Infrastructure funds and private credit managers should diligence the Hyperion structure as the reference template for the next wave of hyperscaler financing, not as a Meta-specific event.

The allocation decision is whether to position as an equity partner in the Blue Owl role or as a debt investor in the PIMCO role on the deals that follow.

The cost of waiting is that the anchor positions in the first replications will be taken by the allocators who diligenced Hyperion first.

Public equity and credit investors should recalibrate how they read hyperscaler balance sheets.

Meta’s consolidated debt will understate its true infrastructure obligations, because the lease commitments sit outside the debt line.

Underwrite the off-balance-sheet lease obligations as fixed claims when benchmarking hyperscaler leverage against reported capex.

Data center developers and operators should evaluate whether the anchor-tenant-plus-SPV model is accessible to them or reserved for AA-rated counterparties.

The Residual Value Guarantee is what makes the pricing work, and only a handful of tenants carry the credit to offer it.

Operators without an investment-grade anchor will find the same structure priced materially wider.

Position now to partner with or lease to the credits that can access this financing, because the operators tied to those tenants will build at a lower cost of capital than those who cannot.

Join Global Data Center Hub

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

Join Global Data Center Hub

TL;DR

• Meta Platforms confirmed on July 13, 2026 that its Hyperion campus in Richland Parish, Louisiana will scale to 5 gigawatts across 3,200 acres, with total investment exceeding $50 billion. The project began as a $10 billion, roughly 4-million-square-foot facility at groundbreaking in December 2024 and is now Meta’s largest data center, with the first 2 GW expected online by 2030, the full 5 GW around 2032, and completion by 2036.

• Hyperion is financed through a joint venture with Blue Owl Capital — the largest private capital transaction on record for a single data center — structured to hold the development off Meta’s balance sheet. Blue Owl’s special purpose vehicle owns 80 percent and Meta 20 percent, on a roughly $28.8 billion committed capital base including $27.3 billion of SPV-issued debt arranged by Morgan Stanley as sole bookrunner, priced at roughly 225 basis points over Treasuries, rated A+ by S&P, and anchored by Pimco with BlackRock participating; Meta’s rent begins June 1, 2029, backstopped by a payment guaranty and a Residual Value Guarantee.

• Entergy Louisiana is building 10 natural gas-fired power plants delivering more than 7 gigawatts for Hyperion — an increase of more than 30 percent to Louisiana’s entire grid capacity. Meta is separately funding up to 2.5 GW of renewable generation, roughly 240 miles of transmission, grid-scale battery storage, and nuclear plant uprates, while approval from the Louisiana Public Service Commission for the gas plants remains pending and will determine delivery speed.

Meta Platforms confirmed on July 13, 2026 that its Hyperion data center campus in Richland Parish, Louisiana will scale to 5 gigawatts across 3,200 acres, with total investment exceeding $50 billion.

The expansion is financed through a joint venture with Blue Owl Capital structured to hold the development off Meta’s balance sheet.

The capital funds a supercluster in the Holly Ridge community that began as a $10 billion project at groundbreaking in December 2024.

Meta expects the first 2 GW online by 2030, the full 5 GW around 2032, and total project completion by 2036.

The Escalation From $10 Billion

Meta broke ground on Hyperion in December 2024 with an initial $10 billion budget for a roughly 4-million-square-foot facility.

In October 2025, Meta formed the Blue Owl joint venture targeting up to $27 to $28.8 billion in total development costs for a 2.06 GW facility.

By February 2026, Meta had acquired roughly 1,400 additional adjacent acres, signaling a phase-two expansion.

The July 2026 confirmation brought the campus to 5 GW and more than $50 billion, making it Meta’s largest data center.

President Trump previewed the $50 billion figure as early as August 2025.

The Blue Owl Joint Venture Structure

The transaction is the largest private capital transaction on record for a single data center.

Blue Owl’s special purpose vehicle owns an 80 percent interest in the joint venture, and Meta retains 20 percent.

The joint venture owns the landlord entity, and Meta operates as tenant.

Meta contributed $5.7 billion in equity against Blue Owl’s $23.0 billion, funded through $27.3 billion in SPV-issued debt, $2.5 billion in equity, and $1.2 billion in accrued interest income, a total committed capital base of roughly $28.8 billion.

Morgan Stanley arranged the debt as sole bookrunner, pricing 144A bonds maturing in 2049 at roughly 225 basis points over Treasuries.

S&P assigned an A+ rating, one notch below Meta’s AA- corporate rating.

Pimco anchored the debt as lead investor, with BlackRock and other institutional credit investors participating.

The Lease And Guarantee Terms

Meta’s rent obligations begin June 1, 2029 and must be paid in full even amid construction delays, with narrow abatements for force majeure or landlord default.

Creditors have no direct recourse to the underlying data center assets.

Meta backstops the arrangement with a payment guaranty and a Residual Value Guarantee designed to cover outstanding SPV debt at the end of each four-year renewal term.

The Power Buildout

Entergy Louisiana is building 10 natural gas-fired power plants delivering more than 7 gigawatts of capacity, a buildout representing more than a 30 percent increase to Louisiana’s entire grid capacity.

Meta is funding up to 2.5 GW of renewable generation, roughly 240 miles of new transmission lines, grid-scale battery storage, and nuclear plant uprates.

Meta states it is paying the full costs of the energy, water, and related infrastructure, and Entergy projects more than $2 billion in customer savings over 20 years.

Regulatory approval from the Louisiana Public Service Commission for the additional gas plants remains pending and will determine how quickly Entergy delivers the capacity Hyperion requires.

Join Global Data Center Hub

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

Join Global Data Center Hub

In this week’s issue:

First, the builders started selling. Bankers are marketing majority stakes in DataBank, EdgeCore, Netrality, and Edged to private equity, with the DataBank process alone reaching up to ~$25 billion, reportedly.

When the operators closest to build economics are the ones cashing out, what does their exit price assume that your entry price does not?

Then, the public market opened its bid. Switch tapped Goldman Sachs and JPMorgan for an IPO valuing it near $80 billion including debt, seven times its 2022 take-private mark, while Brookfield-backed Csquare debuted on the NYSE at $3.2 billion.

Finally, power became equity. Blackstone, Apollo, and KKR committed $5.3 billion for 49 percent of five behind-the-meter plants Williams is building for data centers, the clearest sign yet that secured megawatts are bought, not contracted.

Let us get into it.

THIS WEEK BY REGION

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

North America

This week, the market tested data center valuations across both private and public markets. Bankers launched sales of DataBank, EdgeCore, Netrality, and Edged, with the DataBank process reportedly reaching ~$25 billion, while Switch pursued an IPO valuing it at ~$80 billion, including debt.

On the power side, Blackstone, Apollo, and KKR invested $5.3 billion in behind-the-meter generation, Meta expanded its 5 GW Hyperion campus, and New York became the first U.S. state to impose a statewide data center moratorium.

For infrastructure funds and public-markets investors, the entry multiple is now the whole trade, and the operator selling into strength has better information than the buyer stepping in.

Asia-Pacific

Capital is strengthening the infrastructure underpinning AI growth. Symphony Communication and Nokia are upgrading a key Southeast Asian subsea cable network, while India's AI infrastructure buildout is gaining momentum.

Across both markets, the opportunity lies in funding cables, land, and power ahead of demand.

For tech and infrastructure investors, the seabed and the interconnect are where position is won in Asia this year, well before the marquee campus announcements land.

Europe

Europe is increasingly being underwritten by hyperscaler demand before projects are built. Pure DC secured Microsoft as the anchor tenant for a new AI data center in Finland, making the project financeable before construction.

Meanwhile, Brookfield-backed Csquare debuted on the NYSE at a $3.2 billion valuation, reflecting investor demand for future capacity in the power-rich, low-cost Nordic market.

For real estate and infrastructure investors, the signed hyperscaler offtake is the instrument that turns a European development into a financeable asset, and the projects without one wait.

NOTABLE TRANSACTIONS

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

Blackstone, Apollo, and KKR: $5.3 billion for 49% of Williams’ behind-the-meter power portfolio

The consortium invested $5.3 billion for a 49% stake in five 2.6 GW behind-the-meter gas plants, treating onsite generation as an equity asset rather than a power supply contract.

For the financing landscape, the mechanism is power-as-equity, giving credit and insurance capital ownership of scarce generation while returning long-term control to the developer.

Switch: IPO valuing the operator near $80 billion including debt

Switch has engaged Goldman Sachs and JPMorgan for an IPO that could value the company at ~$80 billion, including debt, up from $11 billion in 2022. The valuation reflects expectations for both long-term growth and leverage.

For public-markets investors, this listing sets the reference multiple for every private data center platform now weighing an exit, and the print will either validate or puncture the marks those private processes are seeking.

DataBank: majority stake process reaching up to ~$25 billion

Bankers are marketing a majority stake in DataBank, reportedly worth up to ~$25 billion, alongside sales of EdgeCore, Netrality, and Edged. Buyers are paying platform valuations for immediate scale and secured growth.

For infrastructure funds, the signal is that the sophisticated developer-owners are electing to sell control into peak demand, and the buyer inherits both the platform and the build risk the seller is choosing to hand off.

When the Builders Start Selling

The people who understand data center build economics best are selling.

That is the signal this week.

Bankers are marketing majority equity stakes in DataBank, EdgeCore, Netrality, and Edged to private-equity buyers, with the DataBank process alone reaching up to ~$25 billion, reportedly.

These are not distressed sales. They are exits at the top of the market.

Read the timing against the cost curve.

Build costs are climbing, not falling.

Shortages of turbines, transformers, electricians, and memory have pushed the cost of a new gigawatt higher every quarter.

The operators selling equity are the ones who see those costs first.

That is what makes the sell-side compelling.

Builders typically sell for two reasons: to raise capital for the next phase or because they believe the current valuation exceeds future value.

Often, both are true. The buyer’s challenge is determining which motive is driving the transaction.

A controlling stake buys more than operating capacity it includes secured sites, customer relationships, and a development pipeline.

Its value depends on whether future demand continues to outpace new supply.

The entry multiple defines the investment.

If it assumes sustained demand growth, the buyer pays a premium for future build risk. If it prices in a slowdown, the seller may be exiting too early.

Switch illustrates the point: after being taken private at $11 billion in 2022, it is now pursuing an IPO valuing the company at roughly $80 billion, including debt a re-rating that reflects expectations far beyond current cash flows.

It is a repricing of what contracted compute capacity is worth, and the public market is about to vote on whether that number holds.

So here is the discipline.

When the sellers know more about the asset than you do, the entry price is the entire argument.

Do not underwrite the stake on demand you assume.

Underwrite it on the cost to replace what you are buying, and on the multiple you would still pay if demand grew slower than the seller’s ask implies.

The seller has priced their conviction. Price yours.

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

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

Have a great week.

— Obinna

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

This article is the 6th article in the series: Ground to Grid: A Free 21-Lesson Guide to Mastering Data Center Development

Join Global Data Center Hub

The sequence of data center development has inverted.

For most of the industry’s history, a developer secured land first, completed the supporting work, obtained entitlements, signed a lease, and arranged power near the end.

That order no longer holds for projects at institutional scale.

Power access and fiber connectivity now sit at the front of the site selection process, and the power purchase agreement is often the document that decides whether a project proceeds at all.

This primer explains how that process works.

It covers the utility interconnection sequence, the main power sourcing strategies, the fiber engineering that governs performance, and the way these constraints shape where AI infrastructure is being built.

Why the order changed

The cause is a timing mismatch. A data center shell can be built in 18 to 24 months.

The high-voltage transmission lines and substations that feed it take seven to ten years to permit and construct.

United States high-voltage transmission construction fell from roughly 1,700 miles per year in the early 2010s to about 180 miles per year by 2022.

The compute deployment cycle and the grid buildout cycle now run on entirely different clocks.

The AI load profile widened the gap.

A training cluster draws near-continuous full-load power.

That flat, predictable consumption curve looks nothing like the variable residential and commercial demand utilities were designed to serve.

A single campus requesting 100 MW or more triggers a multi-stage study sequence before any new infrastructure is built.

Goldman Sachs Research estimates roughly $720 billion in grid spending may be needed through 2030 to support data center expansion.

The utility interconnection sequence

Connecting a large load, generally 20 MW or more, moves through four stages.

The first is the feasibility study.

The developer submits an interconnection request, and the utility conducts a preliminary, non-binding assessment of the upgrades required for the connection.

The second is the system impact study, the substantive engineering phase.

The utility models load flows, short-circuit behavior, and voltage and dynamic stability to quantify the new load’s effect on the transmission system.

For large campuses, this stage frequently surfaces a need for transmission infrastructure the developer did not anticipate.

The third is the facilities study, which scopes, designs, and prices the required upgrades: new substations, transformer additions, line extensions, or reactive power compensation.

The developer is generally responsible for the costs it directly causes.

The fourth is the interconnection agreement and construction. This is where the longest delays occur.

PJM data shows projects spending more than three years to reach an agreement and roughly four more waiting to come online, with total time from application to operation averaging over seven years for AI projects entering service in 2025.

The backlog is immense. Estimates place 2,060–2,600 GW of capacity in U.S. interconnection queues roughly equal to or exceeding the installed capacity of the entire grid.

Conversion rates remain low: only about 40% of ERCOT projects and 24% of comparable PJM projects have reached agreement or become operational.

FERC Order 2023 replaced the old serial queue with an annual cluster study process, grouping requests that arrive within a 45-day window, and raised deposits and withdrawal penalties to clear speculative projects.

A subsequent FERC proceeding aims to build the first dedicated large-load framework for AI data centers, including an expedited track for loads that agree to curtail under grid stress.

Power sourcing strategies

Four approaches now define how projects secure electricity.

Developers now rely on four primary power procurement strategies, each with distinct tradeoffs.

Direct utility interconnection remains the most common, but agreements increasingly include curtailability, network upgrade costs, and minimum-billing commitments.

Northern Virginia illustrates the challenge: transmission constraints not generation have driven connection delays, with Dominion Energy reporting pending data center requests far exceeding its historical peak load.

Power purchase agreements (PPAs) have shifted from an operational decision to a financing priority.

Physical PPAs secure direct supply, while virtual PPAs hedge price risk but introduce basis risk.

Nuclear PPAs have expanded rapidly, highlighted by Microsoft’s Three Mile Island agreement and Amazon’s restructured long-term Susquehanna deal.

Behind-the-meter generation bypasses grid queues through on-site power, typically gas turbines or fuel cells.

While it reduces transmission risk, it creates fuel supply and pipeline dependency, as seen in xAI’s 422 MW off-grid deployment in Memphis.

Hybrid configurations combining grid power, dedicated generation, and storage are emerging as the preferred model for large AI campuses, with gas providing near-term baseload and nuclear paired with storage representing the longer-term direction.

Fiber route engineering

Power governs whether a facility can operate. Fiber governs how well it performs.

Inside the data center, AI workloads require an ultra-low-latency fabric connecting GPUs during training, with bandwidth demands rising each generation from roughly 800 Gbps for the H200 to over 1,200 Gbps for the B200.

As a result, building design, including ceiling heights and conduit routes, is increasingly driven by network requirements.

Outside the facility, resilience depends on true carrier diversity: independent providers using physically separate routes, entry points, and equipment.

Shared conduits do not provide redundancy, making dark fiber ownership and long-term route agreements strategic assets.

Latency also shapes deployment.

Training clusters can locate where power is cheapest, while inference workloads require sub-10-millisecond proximity to users, anchoring them to metro markets.

This rural training–urban inference split is becoming the defining architecture of AI infrastructure.

How the constraints shape geography

The same forces are reshaping every major market differently.

Northern Virginia remains the largest hub, supported by decades of internet exchange and subsea cable infrastructure despite power constraints.

Dallas-Fort Worth continues to attract investors, but Oncor’s 255 GW interconnection queue against a 31 GW system peak signals mounting congestion.

Johor has emerged as Southeast Asia’s fastest-growing hub, driven by low-latency connectivity to Singapore and strong subsea cable access.

Abu Dhabi has accelerated development through a sovereign speed-to-power model that aligns utility control, grid investment, and major cable landings.

The underlying pattern is the same: generation is seldom the constraint power delivery is.

Markets that secure power and fiber connectivity early command higher valuations and faster lease-up.

Developers that lock in grid access and fiber routes capture the greatest upside, while those that delay face rising costs and stronger competition in a market where tenants have global options.

Part 7 examines the permitting gauntlet before groundbreaking: Permitting, Zoning, and Environmental Approvals: Where Data Center Projects Stall

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

Ask ChatGPT a question and the reply appears on your screen in about a second. It feels like the answer comes from your phone, or from somewhere in the air. It does not.

It comes from a specific building, most likely hundreds of miles away, a long windowless box behind a fence with no sign out front.

Your bank runs on one of these buildings. So does Netflix, so does your email, so does the app that holds your photos.

For about twenty years, almost nobody outside the technology business gave these buildings a second thought.

They sat quietly at the edges of towns and did their job, and the rest of us used the internet without ever asking where it physically lived.

That changed almost overnight.

In the last two years, these buildings have shown up in election campaigns, in fights over electricity bills, in warnings from the people who run national power grids, and in the plans of the biggest companies on earth.

To understand why, you first need to know what a data center actually is. It is simpler than it sounds.

Two problems: power in, heat out

A data center is a building with one purpose: to keep a very large number of computers running, all the time, without ever switching off.

The computers themselves are called servers.

Picture a normal desktop computer, then strip away the screen and the keyboard, flatten it into a thin tray, and slide it into a tall metal shelf alongside dozens of others.

Those shelves are called racks, and a big data center holds thousands of them in long rows, like library stacks made of humming metal.

The servers are the part doing the actual work: storing your files, running the apps, answering the questions.

Everything else in the building exists to keep those servers alive, and it comes down to two problems.

The first problem is power.

A room full of running computers needs a constant, heavy supply of electricity, and it cannot afford to lose it even for a moment, because when these machines go dark, so do the services running on them.

So a data center pulls electricity straight from the power grid, and behind that it keeps banks of batteries and backup generators ready to take over the instant the grid hiccups. Keeping the power flowing is the whole game.

The second problem is heat.

Anyone who has felt the underside of a laptop knows that computers get hot, and heat is what destroys them. Multiply one warm laptop by thousands of servers packed into a room, and you have a furnace.

So a large part of every data center is a cooling system, essentially industrial air conditioning, and increasingly a system that runs liquid coolant right up against the chips to carry the heat away.

Add fiber-optic cables to connect the building to the wider internet, and fences and guards to keep people out, and that is the whole thing.

Power coming in, heat going out, and a small city of computers in between. For most of the internet’s history, that description was enough, and these buildings drew a modest amount of electricity and attracted no attention at all.

Then the thing running inside them changed, and the modest amount of electricity stopped being modest.

The shelf that draws a hundred homes’ worth of power

Here is the part that most people have not been told plainly.

You have probably absorbed the idea that artificial intelligence is a triumph of clever software.

That is true, but it hides something.

The kind of AI behind ChatGPT, Google’s Gemini, and the rest runs on NVIDIA chips that have helped make the company one of the world’s most valuable businesses. These chips consume enormous amounts of electricity.

A standard server rack uses about as much power as a few electric ovens. A rack filled with NVIDIA's latest AI chips can consume more than ten times that up to the equivalent of 100 homes.

The heat this produces is so intense that ordinary air conditioning cannot keep up, which is why the newest AI buildings pipe liquid coolant directly onto the chips.

Now multiply that across an entire building. A large AI data center can draw 100 megawatts or more enough electricity to power roughly 80,000 American homes for a year.

One building, drawing as much power as a small city, poured continuously into a windowless box so that a machine can answer questions and generate images for millions of people at once.

That is the reason everyone suddenly cares.

The AI story looks like a software story, and underneath it is a story about electricity, at a scale the world has never had to supply before.

The largest technology companies, Amazon, Microsoft, Google, and Meta, understood this early and began spending on a scale with no precedent.

The consulting firm McKinsey estimates the world will spend roughly $6.7 trillion building data centers by 2030.

And once you follow the electricity, you arrive at the thing that now controls the entire race.

Building it takes months. Powering it takes years.

You might assume the hard part of building one of these is the machines, or the money to buy them.

It is neither. NVIDIA can make more chips over time, and money, for these companies, is not the constraint.

The constraint is getting enough electricity to a specific patch of ground.

The surge in AI demand has broken old assumptions. The International Energy Agency estimates data centers used about 1.5% of global electricity in 2024 and expects that to nearly double by 2030 growing more than four times faster than overall electricity demand.

In the United States the pressure is sharper. Lawrence Berkeley National Laboratory, a US government research lab, found that data centers used about 4.4 percent of the country’s electricity in 2023 and could reach as much as 12 percent by 2028.

In the thirteen eastern states served by the grid operator known as PJM, data centers are expected to account for almost all of the growth in electricity demand between now and 2030.

The trouble is that a power grid cannot grow as fast as a construction crew.

Connecting a big new source of electricity to the American grid now takes a median of about five years, according to Lawrence Berkeley National Laboratory, and the line of projects waiting to connect is longer than the entire existing power supply of the country.

Building a data center, it turns out, is quick. Powering it is slow. Two real stories show what that gap does to the people building these things.

A supercomputer built in 122 days

In 2024, Elon Musk’s AI company, xAI, needed an enormous computer to train its chatbot, Grok, and it needed it immediately.

The company took over an abandoned appliance factory in Memphis, Tennessee, and set out to fill it with a hundred thousand of NVIDIA’s AI chips wired together into a single machine, which it named Colossus.

The industry estimated this build would take eighteen to twenty-four months.

xAI completed it in just 122 days, transforming an empty factory into a working supercomputer. NVIDIA's CEO called it the fastest build of its kind.

Then came the catch.

The local grid could supply only a fraction of the power Colossus needed. Rather than wait years for a grid connection, xAI installed dozens of portable natural gas turbines to generate power on-site.

That choice let the machine run, and it came at a cost to the surrounding neighborhood: the turbines drew objections from residents and environmental groups over the air pollution they added to a part of Memphis that already carried a heavy burden of it, and the dispute ended up in court.

The lesson sits underneath the whole AI boom.

You can build the building in a matter of months, but the power is so hard to come by that a company will burn its own fuel on the doorstep rather than wait in line for the grid.

Switching a dead nuclear reactor back on

Three-Mile Island is a nuclear plant in Pennsylvania, and to many Americans the name means only one thing: the site of the country’s most serious nuclear accident, in 1979. One of its two reactors was involved in that accident.

The other, Unit 1, kept running safely for decades afterward, until 2019, when its owner shut it down because it was losing money.

It sat cold and closed. That looked like the end of it.

In September 2024, Microsoft signed a twenty-year deal to buy the electricity from that plant, on the condition that its owner, Constellation Energy, bring the reactor back to life.

Constellation is spending about $1.6 billion to restart it, has renamed it the Crane Clean Energy Center, and expects it running again by 2027.

Microsoft is buying the plant’s entire output. Every watt that reactor produces is spoken for, to help power the data centers behind Microsoft’s AI services.

Sit with what that means. A shut-down nuclear reactor, a symbol of an accident older than most people reading this, is being switched back on, and the reason is the electricity appetite of AI.

Microsoft is not alone in this. Amazon has bought a data center site next to another nuclear plant. Google and others have signed their own deals for nuclear power.

When companies this large decide that the cheapest way to feed their buildings is to restart a nuclear reactor, you are watching the constraint of this era operate in the open.

Why this lands on your electricity bill

It is easy to see this as a problem for engineers and utility planners. That would be a mistake, because it is precisely why the story moved from the technology pages to the front page.

These buildings compete with homes, hospitals, factories, and schools for electricity.

As they cluster in one region, they can consume a growing share of local power, affecting both electricity prices and grid reliability.

The quiet buildings became loud the moment they started drawing enough power to change what electricity costs for the people living near them.

That is why a governor now holds a press conference about a data center, and why the people who keep the lights on for whole regions now issue public warnings about how fast these buildings are arriving.

There is a larger point underneath the electricity.

The digital life you already live, every message, payment, photo, and stream, has always run through a physical building somewhere, powered by real electricity drawn from a real grid. AI did not create that dependence.

It made it impossible to ignore, by demanding so much, so fast, that it ran headfirst into the limits of the power system the rest of us also rely on.

These buildings are not a footnote to the AI age. They are its foundation, and the foundation is where the strain now shows.

What is the one everyday thing you rely on, without ever thinking about the building it runs through, that surprised you most to learn about here?

Join Global Data Center Hub

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

Join Global Data Center Hub

Tenant credit quality has become the primary determinant of whether a data center investment survives its own lease term.

The old paradigm underwrote concentration risk on asset quality, location, and power access, then treated the anchor tenant as a fixed input.

That paradigm no longer holds, and the cost of applying it is a facility that finances against a counterparty which cannot service the lease for its full duration.

The Intermediation Layer Nobody Priced

The AI buildout created a new intermediation layer that most underwriting models were never designed to price.

Neocloud providers lease long-dated capacity from operators, populate it with GPU clusters, and resell compute on short-duration contracts.

The capital is moving toward this layer at scale, and a widening share of new colocation demand now sits with tenants whose credit profile bears no resemblance to the hyperscalers who anchor the sector’s financing architecture.

The question is not whether these tenants are growing.

The question is whether they will be capable of paying rent for the full term of a fifteen-year lease, given every competing claim on their cash over that period.

How the Constraint Works

Here is how the constraint works.

A hyperscaler tenant carries an investment-grade rating, diversified revenue, and direct access to public debt markets.

A sub-hyperscaler tenant carries none of these.

It runs a duration mismatch at the center of its model: client compute contracts run four to five years, but the data center lease that houses them runs fifteen or more.

The tenant’s ability to service that lease depends on two things holding simultaneously. GPU hardware must retain residual value as collateral.

Compute rental rates must stay high enough to cover debt. Both are under pressure at once.

Peak GPU rental rates have compressed sharply as newer architectures cut per-token costs, and each hardware generation erodes the collateral value of the last.

A regional operator that accepted a fast-growing neocloud at premium pricing to lock in capacity ahead of a power moratorium now holds a fifteen-year obligation underwritten against a business whose margins compress faster than the lease resets.

A Lease Is Only As Long As Solvency

The industry has long assumed that a signed long-term lease is a credit-equivalent instrument.

That assumption held when the universe of tenants capable of signing a fifteen-year data center lease was small, investment-grade, and cash-generative.

It does not hold now, because the tenant pool has expanded downward into a tier of counterparties whose contractual commitment far outlasts their demonstrated ability to honor it.

A lease is only as long as the tenant’s solvency.

Concentration risk also propagates vertically.

A neocloud tenant may rely on a single customer for most of its revenue, meaning a landlord who appears diversified may still have indirect exposure to the same hyperscaler whose lease pause caused the original stress.

Default Is Not Clean

The supply side hardens the problem, and it does so through time.

Data centers are among the most capital-intensive real estate assets in existence.

A facility built to a specific tenant’s power density and cooling specification cannot be re-leased quickly when that tenant fails.

Backfilling a specialized asset commonly runs twelve to eighteen months in secondary markets, and debt service must be covered from reserves or sponsor support throughout.

The asset that looked like commodity real estate at underwriting reveals itself as a bespoke industrial facility at the moment of default. Power scarcity compounds the sequencing pressure.

Operators who might have waited for a tenant’s financials to season instead accepted the tenant early, because the capacity had to be committed before the interconnection window closed.

Where It Binds

For independent operators, the binding variable is tenant qualification.

A single-asset facility leased entirely to one below-investment-grade counterparty is not comparable to a multi-tenant building where diversification absorbs an individual default.

The operator underwrites the whole asset against one balance sheet.

When that balance sheet is speculative grade, the facility inherits the tenant’s credit as its own.

For private equity and infrastructure investors, the binding variable is the financing cost differential and its effect on returns.

A hyperscaler-anchored project finances at materially tighter spreads and higher loan-to-cost than a neocloud-anchored one.

The same asset, in the same location, with the same power contract, requires substantially more equity per megawatt when the anchor tenant is speculative grade.

That equity requirement compresses the internal rate of return directly and constrains how much the sponsor can deploy.

Poor tenant credit does not merely raise the cost of capital. It caps the velocity of the entire strategy.

For public equity investors, tenant concentration drives valuation.

Single-tenant vehicles often trade at persistent discounts because the market prices counterparty risk before fundamentals.

REITs heavily dependent on one tenant typically underperform those backed by diversified or investment-grade tenants, making valuation dispersion a lasting feature of the sector.

Emerging Markets Tighten Every Rule

In emerging markets, every one of these disciplines tightens.

The replacement tenant pool is thinner, so a default that a primary US market absorbs in months can strand an asset for far longer.

You anchor with an investment-grade cornerstone before financing, you build longer vacancy reserves, and you treat sovereign or telco offtake as a credit floor where private investment-grade tenancy does not exist.

The Direction Is Set

The rating agencies have already moved.

The direction of institutional underwriting standards is set, and it points toward progressively excluding the weakest tenants from institutional capital access.

Underwrite tenant credit for the full duration of the lease, at the portfolio level, and you convert a mispriced risk into a source of advantage.

Wait for further confirmation, and you will hold the assets the disciplined capital declined.

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 Risk Transfer Is The Signal, Not The $19 Billion

The structural signal in the Anthropic-TeraWulf lease is not the $19 billion headline.

It is the pass-through tariff filed with the Kentucky Public Service Commission, under which TeraWulf and Anthropic bear all MISO energy and capacity market risk while Big Rivers Electric and Kenergy carry none.

Mainstream coverage has read this deal as a revenue event and a crypto-to-AI pivot story.

The consequence of that read is a mispriced risk profile: the market is underwriting $950 million in average annual rent as if it were utility-grade cash flow, when the underlying power cost sits on the tenant and landlord as a floating wholesale exposure with no margin buffer.

Brownfield Interconnection Is The Scarce Asset Being Monetized

The first read is that the campus monetizes an interconnection, not a building.

The Justified Data site inherited grid access from a retired 600-job aluminum smelter a load already sized for utility-scale industrial draw.

Greenfield projects in this power category wait years for interconnection queue clearance.

TeraWulf calibrated its entire strategy around acquiring legacy high-power sites and building to an anchor tenant, and this lease validates that the interconnection is the value, not the concrete.

Over the next 12 to 24 months, expect competing developers to underwrite retired smelters, steel mills, and coal plants at a premium to greenfield, because the queue position is the moat.

The scarcity is regulatory, not physical.

The Pass-Through Structure Converts Landlord Yield Into Tenant Exposure

The second read is that the pass-through mechanism reallocates who absorbs the volatility that utility-grade cash flow is supposed to eliminate.

Power is delivered with no mark-up on wholesale cost, and the customer is responsible for all transmission and delivery costs and entitled to all MISO node revenues and credits.

This is what produces the premium unit economics approximately $2.37 million per critical IT megawatt-year against a prior industry assumption near $1.9 million but the premium is compensation for risk the tenant and landlord retain, not a free re-rate.

An investor benchmarking this yield against a triple-net utility lease is comparing two different instruments.

The MISO node exposure is the difference, and it does not disappear because the tenant is investment grade.

Rent Commencement Timing Exposes An 18-Month Funding Window

The third read is a liquidity structure hiding inside a revenue story.

Rent commences only as each phase is delivered and commissioned, which protects Anthropic from paying for unenergized capacity but leaves TeraWulf funding a $3.5 billion to $4.0 billion build across an approximately 18-month window with no offsetting lease cash flow.

The Abernathy divestiture $530 million in three installments timed to site preparation, long-lead procurement, and commissioning is not incidental.

It is the bridge across that window.

The proposed $3.5 billion debt package expected to be led by Morgan Stanley completes it.

The signal is that this deal’s execution risk is front-loaded into the pre-revenue period, and the capital stack has been engineered specifically to survive it.

Investor Action

Private Capital. Infrastructure and real estate funds evaluating brownfield conversion should re-benchmark interconnection queue position as a standalone underwritable asset, separate from the physical build.

The diligence action is to price the MISO pass-through exposure explicitly rather than accepting the headline per-megawatt yield as utility-equivalent.

The cost of waiting is that the premium sites retired smelters and coal plants with inherited transmission get bid away by developers who underwrite the queue position first.

Position now on legacy high-power industrial sites or accept greenfield timelines.

Public Markets. Equity investors in TeraWulf and comparable converted miners should separate the contracted-revenue narrative from the pre-revenue funding exposure.

The diligence action is to model the 18-month construction window against the disclosed debt raise and the Abernathy installment schedule, not against the $19 billion aggregate.

The cost of inaction is holding through a dilution or refinancing event mispriced as a pure re-rate. Verify the investment-grade credit reference against actual lease terms, because Anthropic is a private company and the language warrants confirmation.

Operators. Data center developers and hyperscalers acting as tenants should read the single-tenant direct-lease structure as the strategic move it is.

Anthropic secured dedicated, controlled 401 MW capacity outside the cloud metering model, at roughly 4% of its total planned capacity, as part of a deliberate multi-provider strategy.

The diligence action for operators competing for anchor tenants is to structure pass-through tariffs that transfer market risk cleanly, because that transfer is what makes the premium yield financeable.

The cost of failing to do so is losing anchor mandates to developers who can.

The Verdict

This lease is a template, not an outlier.

It decouples high-density compute from the public cloud and binds it directly to a utility-scale energy asset with the market risk pushed onto the counterparties best able to contract around it.

The inflection is the emergence of the brownfield-plus-pass-through structure as the repeatable unit of AI infrastructure development inherited interconnection, single-tenant anchor, wholesale risk transfer, phased rent commencement.

The open question for the next 24 months is whether the tenants signing these leases hold investment-grade credit that survives a training-economics downturn, because the entire structure rests on rent obligations that do not begin until capital has already been spent.

Join Global Data Center Hub