AI-Driven Data Center Demand: Current State

Introduction

The demand for data center capacity is not in question. The constraint is. The four largest hyperscalers, Amazon, Microsoft, Alphabet, and Meta, now guide to combined 2026 capital expenditure of roughly $720 billion to $745 billion after their second-quarter updates, close to double the roughly $410 billion they spent in 2025. Developers have answered with a record construction pipeline, and yet vacancy in the primary data center markets sits at a record-low 1.4%, because the binding constraint has shifted. The scarce resource is no longer capital or even land. It is electricity. Understanding the current state of the sector means holding both facts at once: unprecedented demand, and a supply chain that is building faster than ever and still cannot keep up.

The scale of hyperscaler spending is difficult to overstate, and each earnings season has pushed it higher. In July 2026, Alphabet raised its 2026 range to $195 billion to $205 billion, up from $180 billion to $190 billion a quarter earlier. Amazon lifted its forecast to about $220 billion from $200 billion, citing higher memory prices, and Meta narrowed its range to $130 billion to $145 billion. Microsoft's calendar-2026 figure of about $175 billion, including finance leases, sits below its April guide of $190 billion only because it lengthened the assumed useful life of its data center buildings from 15 to 25 years, which shifts future leases from finance leases (counted in capex) to operating leases (not counted); the spending plan itself stayed the same.

What makes AI capex different from the prior cloud build is intensity. Training and inference workloads pack far more computing power into each rack, pushing power density from the roughly 5 to 10 kilowatts per rack of traditional facilities toward 100 kilowatts and beyond. A given building now consumes a multiple of the electricity it would have a few years ago, which is why the demand story is measured in gigawatts rather than square feet and why the hyperscale segment, rather than retail or wholesale colocation, drives the market.

These commitments now exceed the spenders' own cash generation. Amazon's trailing-twelve-month free cash flow swung to an outflow of $7.6 billion in its second-quarter results, and Meta's quarterly free cash flow shrank to $784 million from $8.5 billion a year earlier. Even so, Andy Jassy told investors Amazon will still lack the capacity to meet all its demand in 2026, and expects the same in 2027. For a real estate banker, the takeaway is that demand is anchored by a handful of the most creditworthy tenants on earth, spending past their own free cash flow and still short of capacity, the tenant base profiled in the hyperscaler credit universe.

Because the spend routinely exceeds operating cash, the industry has moved beyond the corporate balance sheet toward investment-grade bonds, off-balance-sheet vehicles, joint ventures, and private credit. On the bond side alone, Alphabet, Amazon, Meta, and Oracle issued more than $228 billion of bonds through October 13, 2026, more than double their total for all of 2025, according to LSEG data cited by CNBC, the issuance wave analyzed in how hyperscaler bonds are reshaping IG issuance.

The landmark structured transaction is Meta's Hyperion campus in Richland Parish, Louisiana, financed in October 2025 through a roughly $27 billion private-credit deal, the largest in history. A special-purpose vehicle arranged by Morgan Stanley issued about $27 billion of A+-rated debt plus $2.5 billion of equity, anchored by PIMCO at around $18 billion and BlackRock, with Blue Owl Capital owning 80% of the vehicle and Meta only 20%. That lets Meta build a five-gigawatt campus without consolidating the full cost onto its balance sheet, protecting its credit profile while reaching a capital pool far larger than its own cash flow. It is exactly the kind of structuring a real estate banker is now asked to run.

Off-balance-sheet SPV financing

A structure in which a data center campus is built inside a separate special-purpose vehicle, funded largely by outside debt and equity so the bulk of the project cost never appears on the sponsor's own balance sheet. Meta's roughly $27 billion Hyperion vehicle, 80% owned by Blue Owl, is the defining example.

The pattern repeats across the sector. Stargate, the OpenAI, Oracle, and SoftBank joint venture announced in early 2025, targets $500 billion of data center investment, with its first site live in Abilene, Texas. Behind all of it sits the growth of private credit, a roughly $2.1 trillion market projected to reach $3.5 trillion by 2030, which has found in AI infrastructure an asset class big enough to absorb it. Data center securitization is scaling alongside it, with JPMorgan projecting $30 billion to $40 billion of annual ABS issuance in 2026 and 2027.

The financing has also reached down to the riskier players. Neoclouds, the AI-specialist providers such as CoreWeave that rent out GPU capacity, increasingly fund their chips with GPU-backed debt, borrowing against the hardware itself, a structure that holds only as long as the GPUs retain value and the contracts behind them perform. For the banker, the data center build is no longer just a real estate or technology story; it is a structured-finance story, and the firms that can assemble SPVs, syndicate private debt, and pair hyperscaler credit with institutional capital are the ones capturing the fees.

That demand has produced the tightest conditions in the property type's history. In CBRE's H1 2026 North American data center report, primary-market vacancy held at a record-low 1.4%, unchanged from year-end 2025, even as total supply rose 33.7% to a record of almost 10.9 gigawatts. Net absorption reached 1,456 MW in the first half, up nearly 12%, after a record 2,497.6 MW for full-year 2025. Pricing has followed, with average asking rents up for every deployment size in the first half.

Preleasing

Preleasing is the practice of signing tenants to a data center before construction is complete, often before it has even broken ground. In the current market, hyperscale facilities are routinely fully preleased years ahead of delivery, which is why headline vacancy stays near zero: the space is spoken for long before it exists.

Northern Virginia remains the gravitational center of the market. Its inventory reached about 4.5 GW in the first half of 2026, it led all primary markets with 467.6 MW of first-half absorption, and its vacancy fell to 0.2%, leaving barely 10.8 MW available. The growth frontier is widening, though: Atlanta now has more capacity under construction than Northern Virginia, about 2,882 MW against 2,420 MW, and Dallas-Fort Worth has more than 765 MW underway, roughly 95% preleased. The largest hubs are where the constraint bites hardest, which is why the sector sits in the supply-constrained expansion phase described in cycle positioning across sectors.

The most important number in the sector is not a capex figure or a vacancy rate. It is four years, the average wait for a grid connection in the United States. That single fact reframes the entire investment thesis: a developer can have capital, a signed hyperscale tenant, and entitled land, and still be unable to deliver because the local utility cannot supply power for years.

Grid Interconnection

Grid interconnection is the physical and contractual connection of a facility to the electric transmission system, governed by the local utility and grid operator. For data centers it has become the gating item: securing an interconnection agreement, and the multi-year queue position that comes with it, now happens before site selection rather than after.

The queue no longer stops developers from breaking ground so much as it decides when their buildings can switch on, a distinction that is easy to miss in a year of record construction.

The industry's response has been to stop waiting for the grid. Bring-your-own-power arrangements and on-site generation, including batteries and natural gas turbines, are becoming routine, and bridging power is gaining ground wherever utility dates miss tenant schedules. Power procurement has moved from an afterthought to the first item in the development pro forma, the dynamic explored in power as the binding constraint and increasingly handled alongside the energy and power coverage teams that understand generation and interconnection.

The Power Deals Behind the Headlines

The scramble for electricity has produced a wave of landmark procurement deals. Constellation Energy agreed to restart the shuttered Three Mile Island reactor under a 20-year agreement dedicating its full 835 megawatts to Microsoft, investing roughly $1.6 billion to bring it back online. Amazon contracted with Talen Energy for up to 1,920 megawatts from the Susquehanna plant, and Meta signed a 20-year deal with Constellation for the 1.1-gigawatt Clinton plant. On the faster side, on-site natural gas is becoming standard: Hyperion is planned with roughly ten gas turbines, and about 30% of planned new capacity is designed to run independently of the grid. The common thread is that hyperscalers, not utilities, are now the marginal buyers setting the price of clean power, and a platform's energy contracts have become as valuable as its land.

The grid queue gets the headlines, but the physical supply chain is just as binding. The scarcest component is the large power transformer, where lead times have stretched to roughly 128 weeks, with generator step-up units near 144 weeks and prices up on the order of 77% to 95% since 2019. A campus cannot energize without this gear, so the transformer order book gates delivery as tightly as the interconnection queue, and the two delays compound. Industry analysts estimate that 30% to 50% of the capacity planned for 2026 will slip to 2028 as a result. Cooling is next: AI racks drawing 100 kilowatts and beyond have made air cooling insufficient, forcing a shift to liquid and direct-to-chip cooling, and converting an air-cooled hall for liquid-cooled hardware can add more than six months. Developers now run procurement as a strategic function, ordering transformers and turbines years ahead and treating a secured equipment slot almost the way they treat a secured megawatt.

Because power is the constraint, new capacity follows electricity rather than population. The established giants, Northern Virginia foremost among them, are running into local opposition and grid limits, so developers are pushing into secondary markets chosen for power availability: Phoenix, Atlanta, Columbus, central Texas, and emerging sites near low-cost generation.

The same logic plays out internationally. The Gulf states are courting hyperscale capacity with cheap power and sovereign capital, while the Nordics offer cool climates and clean grids. Europe's core markets, Frankfurt, London, Amsterdam, Paris, and Dublin, face the same power and permitting bottlenecks as Northern Virginia, and Dublin has effectively paused new connections. Why this favors the operators who control power is laid out in data center real estate and the AI demand surge.

The capital pouring into the sector has produced the largest transactions in real estate. The sale of Aligned Data Centers by Macquarie's infrastructure funds, at an enterprise value of about $40 billion, closed on July 21, 2026, nine months after it was announced. The buyers, the AI Infrastructure Partnership (whose backers span Microsoft, Nvidia, and xAI), MGX, and BlackRock's Global Infrastructure Partners, took a platform of 51 campuses and more than 6.4 GW of operating and planned capacity, and committed another $5 billion of growth capital. That deal more than doubled the prior record, Blackstone and CPP Investments' roughly $16 billion (about A$24 billion) take-private of AirTrunk in 2024. Across 2025, S&P counted 113 data center transactions worldwide worth more than $69 billion, a record by some $8 billion.

The deals share a logic. Buyers are paying for platforms that control the two genuinely scarce inputs, land with secured power and creditworthy hyperscale tenants, and they are increasingly infrastructure funds, sovereign wealth, and the chipmakers and AI labs themselves rather than traditional real estate investors. The public markets feed the same demand: Blackstone's data center REIT, BXDC, came to market as a roughly $1.75 billion blind-pool listing, and Switch, taken private by DigitalBridge and IFM in 2022, filed confidentially for a US IPO in August 2026, with a listing possible as soon as November; both are tracked in the current REIT IPO pipeline. This is where the sector's M&A and capital-markets fees concentrate, and why digital infrastructure has become a coverage priority spanning both real estate and energy and infrastructure groups.

No survey of the current state is complete without the debate over whether the build is rational, and a banker should be able to argue both sides. The arithmetic that worries skeptics is stark: the Big Four alone guide to $720 billion to $745 billion of 2026 capex, while most estimates of direct AI revenue remain in the tens of billions, an order-of-magnitude gap between investment and current return. Critics point to circular financing, in which chipmakers, model developers, and cloud providers invest in one another in ways that can inflate apparent demand, and to the cash burn at neoclouds and AI labs. The law firm Clifford Chance has flagged that something like 100 gigawatts of capacity is slated to come online between 2026 and 2030 without a clear, demand-tested picture of who will fill it.

The sharpest financial concern is a duration mismatch. The GPUs that justify these buildings depreciate fast, with useful lives plausibly as short as one to three years against the six-year schedules some operators book, while the debt financing them runs far longer. Lenders normally want an asset's life to exceed the loan's tenor; AI hardware inverts that, which is why observers draw uncomfortable parallels between today's data center securitizations and the layered structured finance that unwound in 2008.

The rebuttal matters as much as the worry. The real estate itself, the powered shell, the substation, the interconnection, and the cooling, has a useful life measured in decades, and it is leased to among the most creditworthy tenants on earth. Even if a generation of GPUs is obsolete quickly, the power and the building remain scarce and reusable. Whether the boom is a bubble therefore turns less on the real estate than on whether AI demand justifies the compute, a question that sits one layer above the property.

October 2026: The Credit Market Tests the Thesis

The duration worry moved from argument to price in mid-October 2026. On October 14, CoreWeave shares fell 15.5% as yields on its 2031 and 2032 bonds approached 15%; Oracle, which has borrowed heavily for its own build, closed down nearly 9%, and fellow neocloud Nebius dropped 10%. The next day CoreWeave's five-year credit default swap spread topped 1,140 basis points, a level S&P Global Market Intelligence treats as distressed, on more than $51 billion of debt and obligations at a B+ rating. Its backlog of over $100 billion sits about two-thirds with investment-grade customers, so investors were not doubting demand; they were doubting whether cash flow arrives before the debt comes due, the same timing risk behind the duration mismatch.

The stress landed on the most leveraged renters of compute and the builders that borrowed hardest, exactly the layer the rebuttal concedes is fragile. The honest current-state read is that the buildings are real and scarce, the financing is aggressive and now being repriced by the credit markets, and the demand underneath the chips is the variable no one can yet prove.

Powered land and shell buildings with secured interconnection command enormous premiums, because the power, not the concrete, is the scarce asset. Capital is flooding toward platforms that control sites with locked-in electricity, which is why the largest M&A and capital-raising activity in real estate now runs through this sector, from the Aligned sale to the BXDC listing.

The valuation conversation has shifted with it. A site is no longer priced mainly on rent per square foot but on its secured power capacity and the credit of the tenant behind it, so a parcel with a signed interconnection and a hyperscale lease can trade at a multiple of an identical parcel still stuck in the queue. Underwriting a data center deal now means underwriting a power-purchase agreement, a substation timeline, and an equipment order book alongside the lease, and after October the cost of capital behind platform sales, recapitalizations, and SPV financings is a live diligence item rather than an assumption.

The state of the market in 2026 is therefore a paradox that the best analysts can explain in one breath: the most capital-rich, demand-rich sector in real estate is supply-constrained not by money but by megawatts, and even a record construction pipeline cannot close the gap. That is why the build costs and time-to-power calculus covered in data center capex and time to power have become the decisive variables, and why power, more than anything on a traditional rent roll, will determine which projects actually get built over the next several years.