Nvidia, OpenAI, Oracle, and CoreWeave are funding 2026’s AI data center buildout through vendor equity stakes, off-balance-sheet debt vehicles, and backlogs built on prepaid or customer-supplied hardware — not primarily through free cash flow. That shifts risk toward chipmakers, private-credit investors, and leveraged compute resellers, while the hard limit on how fast any of it can scale stays the same: the electrical grid, not GPU supply.
Quick takeaways:
- Big Tech’s 2026 AI capex guidance clusters between roughly $630 billion and $800 billion depending on which companies and which “AI share” of cloud spending an analyst includes — and actual spending has beaten every forecast issued a year earlier.
- Nvidia’s planned $100 billion OpenAI investment was never a signed contract. By February 2026, reporting pointed to a considerably smaller equity stake instead, even as chip shipments continued.
- Meta moved $27 billion of AI data center debt off its own balance sheet through a special-purpose vehicle with Blue Owl Capital — a structure other hyperscalers are reportedly copying.
- Michael Burry and others argue hyperscalers are understating GPU depreciation by tens of billions of dollars a year; the hyperscalers counter that the accounting reflects real multi-year utility across training, inference, and batch workloads.
- None of this touches the actual bottleneck: US grid interconnection queues held roughly 2,600 gigawatts of pending requests in early 2026 — about double the capacity of the entire existing grid.
Oracle’s fiscal year ended in May 2026 with two numbers that shouldn’t comfortably coexist, yet do. The company’s backlog of contracted AI infrastructure work — its remaining performance obligations, or RPO — hit $638 billion, up 363% year over year. Its free cash flow for that same fiscal year was negative $23.7 billion.
Oracle isn’t an outlier. It’s the clearest single illustration of what AI infrastructure spending actually looks like in 2026: enormous, largely real, and financed in ways that would have raised eyebrows almost anywhere else in corporate America two years ago. Every hyperscaler earnings call this year repeats some version of the same three claims — demand is unconstrained, supply is the bottleneck, next year’s spending will be higher still. All three are probably true. What gets less attention is how the checks are actually being written. Who ends up holding the exposure if the compute doesn’t monetize on the schedule everyone is currently assuming?
The number nobody can quite agree on
Ask five analysts how much the largest AI infrastructure buyers will spend in 2026, and you’ll get five different totals — and most of them can be defensible at once.
Dell’Oro Group, which tracks data center hardware spending, put the global data center capex outlook for 2026 above $1 trillion, citing accelerating hyperscale AI deployment plus a separate spike in memory and storage pricing that inflates server costs independent of unit volumes. Narrow the lens to the five biggest US-listed AI infrastructure spenders — Amazon, Alphabet, Meta, Microsoft, and Oracle — and estimates compiled from Q1 2026 earnings calls cluster closer to $660–690 billion for the year. The gap between the two figures comes down to scope: one counts every dollar of global data center capex, including non-AI cloud infrastructure; the other counts just the AI-heavy capital plans of five companies.
The direction of the revisions matters more than any single total. A Goldman Sachs analysis circulated in a 2026 SEC filing shows hyperscaler AI capex growth guided at the start of 2024 landed at 19%; actual realized growth came in around 54%. The same pattern repeated in 2025: 22% guided, roughly 64% realized. Every forecast has undershot the outcome, in the same direction, for two straight years.
Microsoft’s own disclosures make the mechanism concrete. The company told investors it was carrying an $80 billion backlog of Azure orders it couldn’t fulfill — not because of chip supply or capital, but because of power constraints on orders that were already funded. That single detail previews most of what follows: the money is largely there. Getting it turned into usable compute on schedule is the harder problem.
The vendor-financing loop
Start with the highest-profile arrangement. In September 2025, Nvidia and OpenAI announced a deal to deploy at least 10 gigawatts of Nvidia systems for OpenAI, with Nvidia stating it “intends to invest up to $100 billion” as each gigawatt came online — a commitment Jensen Huang has since called the biggest AI infrastructure project in history.
Read the phrasing carefully, because Nvidia’s own lawyers clearly did. Two months later, Nvidia’s quarterly filing warned there was “no assurance” the arrangement would be finalized. By December 2025, CFO Colette Kress told an investor conference the deal was still at the letter-of-intent stage, with no definitive agreement signed and no money transferred. By February 2026, reporting pointed to a considerably smaller commitment — an equity stake closer to $30 billion, folded into an OpenAI funding round reportedly valuing the company near $830 billion — rather than the original $100 billion infrastructure-investment structure.
The chips kept shipping regardless. That’s the part critics of the arrangement focus on: a chip supplier taking equity in the company that is simultaneously its largest customer, where the investment helps that customer buy more of the supplier’s chips, which then reads as revenue and demand validation for the supplier. Nvidia didn’t invent this structure. A few weeks before the OpenAI announcement, AMD signed OpenAI to a six-gigawatt, multi-generation GPU agreement and issued OpenAI a warrant for up to 160 million AMD shares — potentially close to 10% of the company — vesting only as OpenAI deploys hardware, and only if AMD’s stock reaches specific price targets, including $600 a share for the final tranche.
Neither structure is unusual by venture-dealmaking standards. What’s unusual is the scale, and the fact that the “customer” in each case is a company still losing money on an operating basis, whose capacity to pay depends partly on the value of the stock it’s being handed as an incentive. OpenAI’s own growth has been real — reported annualized revenue moved from around $12 billion in mid-2025 toward the low-$20-billion range by early 2026 — which is why the arrangement currently reads as strategic alignment rather than a warning sign. If that growth curve bends, the same structure means Nvidia and AMD absorb part of the disappointment directly, as shareholders, not merely as vendors watching a customer struggle.
Taking the debt off the books
If vendor equity is one way to fund a buildout without writing a check today, off-balance-sheet debt is another — and it’s now the more common one.
In October 2025, Meta and private-credit firm Blue Owl Capital closed the largest deal of this kind to date: a joint venture to build Meta’s Hyperion data center campus in Richland Parish, Louisiana, with Blue Owl owning roughly 80% and Meta 20%. Blue Owl funded its share largely through $27 billion in bonds — anchored by PIMCO and partly bought by BlackRock — structured to amortize fully by 2049. Because Meta doesn’t consolidate the joint venture’s debt, none of that $27 billion shows up on Meta’s own balance sheet, even though Meta is the facility’s sole tenant and provided a residual value guarantee that effectively backstops the asset. S&P rated the debt A+ despite pricing it to yield 6.58%—a spread that looks more like high-yield than investment-grade, and the market is quietly pricing in more risk than the letter grade implies.
Analysts who track the deal expect Microsoft, Amazon, and Google to run comparable structures — Microsoft has reportedly discussed one with BlackRock’s infrastructure arm, and Oracle’s Stargate build-out already involves Blue Owl on similar terms. The logic is straightforward: keep the debt off the parent company’s credit metrics while keeping full operational control of the asset it’s paying to lease.
Oracle took a related but distinct approach to its OpenAI contract. When Oracle reported record fiscal 2026 results, remaining performance obligations had reached $638 billion, and Bank of America estimated more than half of that figure traces back to OpenAI alone. Oracle’s own disclosure contained the more interesting detail: roughly $75 billion of that backlog is structured so the customer prepays for GPUs, or supplies the GPUs itself, before Oracle has to spend its own capital. That reduces Oracle’s funding burden — but it also means the number that moves Oracle’s stock depends on a customer that runs at a loss continuing to prepay for hardware years before it needs to use it.
Not every AI infrastructure company can access this kind of financial engineering. CoreWeave, the largest independent “neocloud” GPU provider, has built its roughly 1.5-gigawatt fleet mostly on secured debt — about $35 billion by the second quarter of 2026, against a few billion dollars of cash. Interest expense alone reached $640 million for the quarter. Short-sellers including Kerrisdale Capital have flagged that CoreWeave borrows against Nvidia GPUs as collateral to build capacity ahead of signing the customer contracts meant to eventually pay for it. CoreWeave’s roughly $104 billion revenue backlog is real. So is a debt load that consumes a meaningful share of every dollar of revenue before it reaches the bottom line.
| Entity | Primary financing mechanism | Disclosed scale (mid-2026) | Key risk flagged by analysts |
|---|---|---|---|
| Meta (Hyperion) | Off-balance-sheet SPV with Blue Owl; residual value guarantee | $27B debt + ~$3B equity | Debt effectively backed by Meta despite not being consolidated on its books |
| Oracle (Stargate/OpenAI) | Backlog with prepaid or customer-supplied hardware | $638B RPO; ~$75B prepaid or customer-supplied | Over 50% of RPO tied to a single, loss-making customer |
| CoreWeave | GPU-collateralized secured debt | ~$35B total debt; ~$104B backlog | High leverage, negative free cash flow, customer concentration |
| Nvidia / AMD | Vendor equity stakes and stock warrants in customers | Up to ~$30B equity (Nvidia–OpenAI); ~10% warrant (AMD–OpenAI) | Supplier’s own returns tied to a loss-making customer’s growth |
What happens if the backlog doesn’t convert on schedule
Contracts and backlogs assume the hardware behind them keeps earning for years. That assumption runs into a quieter, arguably more consequential dispute: how long a GPU is actually worth using.
Between 2020 and 2023, Amazon, Microsoft, Google, and eventually Meta each extended the depreciation schedule on servers and networking gear from a historical three-to-four-year window out to five or six years. Microsoft disclosed the mechanics directly: extending useful life from four to six years added roughly $3.7 billion to fiscal 2023 operating income, by its own estimate, simply by spreading the same hardware cost over more quarters. Google made a similar change in January 2023. Each move is a legitimate change in accounting estimate under US GAAP, not a restatement or an error — but an estimate is still an estimate, and estimates can be wrong in a direction that flatters earnings.
Michael Burry, the investor known for shorting mortgage bonds before 2008, made that argument publicly on November 11, 2025, estimating hyperscalers were collectively understating depreciation by roughly $176 billion across 2026 through 2028, with Oracle’s and Meta’s reported earnings specifically overstated by around 27% and 21% respectively by 2028. He disclosed short positions in Nvidia and Palantir the same day. Separate, independently built analysis using a slightly different useful-life counterfactual arrived at a similar order of magnitude, near $200 billion.
The hyperscalers’ counterargument has real technical substance, not just financial motive. A GPU cluster’s economic life doesn’t have to match its frontier-training life: Nvidia ships a new architecture every 18 to 24 months, but a chip no longer competitive for training the largest models can spend years afterward running inference workloads, which are far less sensitive to raw performance and far more sensitive to cost per token. That tiered reuse — frontier training, then high-value inference, then batch and analytics work — is the same logic Amazon used when it first extended server life back in 2020, and it isn’t obviously wrong on its own terms. What’s harder to defend is the near-simultaneous, near-identical timing of four competitors reaching that same conclusion within roughly the same eighteen-month window — a pattern one investment analysis diplomatically described as “safety in numbers” rather than four independent engineering studies arriving separately at six years. Notably, Amazon partially reversed course in 2025, shortening the useful life of a subset of servers even as Meta extended its own further: the first sign the group consensus might be cracking.
The constraint no term sheet can fix
None of the financing innovation above solves the one bottleneck money can’t shortcut: getting electricity to a building.
US grid interconnection queues — the list of generation and storage projects waiting for utility approval to connect — held roughly 2,600 gigawatts of pending capacity requests as of early 2026, according to tracking of PJM and other regional grid operators. That’s larger than the entire existing US grid. Former Google CEO Eric Schmidt told a congressional panel that data centers would need 29 additional gigawatts of power by 2027 and 67 more by 2030 — demand a grid built for low-single-digit annual growth was never designed to absorb on that timeline. In mid-2026, the Federal Energy Regulatory Commission issued a show-cause order to six regional grid operators, giving them roughly 60 days to fix interconnection backlogs or justify why the existing process should stay in place — a regulator signaling that the queue itself has become a national economic problem, not just an industry inconvenience.
That’s why nearly every hyperscaler now has a direct power deal designed to bypass the queue entirely rather than wait in it. Microsoft signed a 20-year agreement with Constellation Energy in 2024 to restart Three Mile Island’s undamaged reactor unit. Amazon has separately contracted nuclear or small-modular-reactor capacity with Talen Energy, Dominion Energy, and Energy Northwest. Meta locked in 1,100 megawatts from Constellation’s Clinton, Illinois plant. Where nuclear timelines run too slow, the fallback is gas: Oracle partnered with Volta Grid on a 2.3-gigawatt modular gas fleet, and Crusoe Energy builds data centers next to dedicated gas plants rather than waiting on utility interconnection at all. Even a 100-megawatt greenfield site with an approved application can still take four to six years to energize once transmission upgrades are triggered — a timeline that doesn’t move faster no matter how large the check attached to the request.
The chips themselves are making this worse, not better. Nvidia’s 2026 Vera Rubin NVL72 rack draws around 200 kilowatts, and the Rubin Ultra generation due in 2027 is expected to approach 600 kilowatts per rack — a roughly 60- to 120-fold increase in rack power density over the past decade. Air cooling stops working at that density, which is why direct-to-chip liquid cooling has gone from a specialized option to the default assumption in new AI data center design, and why some operators are already piloting two-phase cooling, where a dielectric coolant boils directly against the chip and condenses elsewhere in the loop. None of this is optional engineering flourish. It’s the physical price of the performance gains chipmakers are promising, and it stacks another layer of capital and lead time on top of the power problem itself.
The silicon hedge
If you’re a hyperscaler paying Nvidia’s prices and helping fund Nvidia’s own customers, building your own chip stops looking optional.
Google’s seventh-generation TPU, Ironwood, reached general availability in late 2025, delivering 4,614 FP8 teraFLOPS and 192GB of HBM3e memory per chip, scaling to pods of 9,216 chips for a combined 42.5 exaflops. Google now runs more than three-quarters of the computation behind its Gemini models on its own TPUs rather than merchant GPUs and reportedly plans to ship on the order of 4.3 million TPUs in 2026 — every one a dollar of margin Google doesn’t send to Nvidia. Amazon’s answer, Trainium3, reached general availability around the same time: its first 3-nanometer chip, delivering 2.52 petaFLOPS of FP8 compute and 144GB of memory, with a fourth generation already planned for late 2026 or 2027 that will notably add support for Nvidia’s own NVLink Fusion interconnect — a hedge on top of a hedge, keeping the door open to mix Amazon silicon with Nvidia GPUs in the same rack. Microsoft’s Maia 200 inference chip, detailed in a January 2026 datasheet, remains internal-only for now, running Copilot and Azure OpenAI traffic without an external instance type customers can rent.
| Chip | Built by | Peak FP8 compute | Memory | External availability (Sept. 2026) |
|---|---|---|---|---|
| TPU Ironwood (v7) | 4,614 TFLOPS/chip | 192GB HBM3e | Google Cloud only | |
| Trainium3 | Amazon (Annapurna Labs) | 2.52 PFLOPS/chip | 144GB HBM3e | AWS only |
| Maia 200 | Microsoft | Not publicly disclosed | 216GB HBM3e | Internal use only; no customer SKU |
| Instinct MI400 | AMD | Not yet in volume shipment | 432GB HBM3e (highest disclosed) | Merchant silicon, multi-cloud |
None of these chips is trying to beat Nvidia across the board. Each is built to capture the highest-volume, most predictable workload a hyperscaler controls directly — its own inference traffic — in-house, leaving Nvidia’s GPUs for frontier training and for external customers who need the CUDA software ecosystem, still the deepest moat in the industry regardless of how a rival chip’s spec sheet reads. Estimates of Nvidia’s share of AI accelerator revenue still run from roughly 70% to over 90%, depending on whether an analyst counts hyperscalers’ internal chips in the addressable market at all. AI labs are making the same hedge from the buying side: Anthropic trains and serves Claude across a cluster of nearly 500,000 Amazon Trainium2 chips that scaled past one million by the end of 2025, a separate deal for up to a million Google TPUs, and Nvidia GPUs — deliberately spreading dependence across more than one supplier’s roadmap and pricing power.
The case this isn’t 2001 again
It’s worth taking the optimistic case seriously, because parts of it are simply true. Nvidia’s data center revenue isn’t speculative — it was $89 billion in a single quarter, growing 117% year over year, billed to customers paying real invoices. That’s a different animal from the mostly unlit fiber-optic capacity that sat idle for years after the dot-com telecom buildout. Sam Altman, Mark Zuckerberg, Federal Reserve chair Jerome Powell, and Apollo Global’s chief economist Torsten Slok have all separately invoked the word “bubble” about AI investment within the past year — notable candor from people with a direct financial stake in the outcome. Altman’s own framing captured the tension: he expects someone to lose an enormous amount of money in the process while acknowledging plenty of others will profit substantially along the way, and has compared the dynamic to the dot-com era without predicting an identical ending.
The useful-life argument for GPUs, similarly, isn’t purely a trick: a chip pushed out of frontier training genuinely has years of legitimate inference work ahead of it, and that reuse curve is a real change from most physical infrastructure cycles. Separately, an August 2025 MIT-affiliated study found that roughly 95% of enterprise generative-AI pilot programs weren’t yet showing measurable return on investment, despite tens of billions in enterprise spending — a useful check on hype, though it measures corporate AI adoption, not the underlying infrastructure economics covered here; foundation-model training and consumer-facing products draw on the same GPUs and are not captured by that figure. The honest read is that AI infrastructure spending contains both a real, monetizing core and a speculative, momentum-driven edge, and the financing structures described above — vendor equity, off-balance-sheet SPVs, prepaid backlogs, stretched depreciation — are precisely the tools that make it hard to see, from outside, where one ends and the other begins. That opacity, more than any single number, is the actual risk.
What to actually watch
None of this requires a prediction about whether AI infrastructure is a bubble, because that framing is a little lazy — parts of it are clearly overbuilt relative to near-term revenue, and parts clearly aren’t. What matters more for anyone tracking this is a small set of concrete, checkable signals, most of which show up in quarterly filings rather than press releases:
- Backlog-to-revenue conversion speed. Oracle’s $638 billion RPO and CoreWeave’s $104 billion backlog are only as good as the rate they convert into billed, collected revenue. Watch the realized-versus-guided gap each quarter, not the headline backlog figure.
- The prepaid or vendor-financed share of any backlog. Oracle already discloses that roughly $75 billion of its RPO is prepaid or customer-supplied hardware. A rising share of “backlog” funded by the customer rather than the vendor carries a different risk profile than the press release implies.
- Interest coverage at the neoclouds. CoreWeave’s interest expense was already running near $640 million a quarter by mid-2026. If revenue growth decelerates while debt keeps climbing to fund committed capex, interest coverage — not the backlog — is the number that moves first.
- Useful-life disclosures in the next two 10-K filing seasons. Amazon already shortened the assumed life of a subset of servers in 2025 while Meta extended its own further. A second hyperscaler moving in Amazon’s direction would be a more meaningful signal than any outside depreciation model.
- Interconnection queue timelines in a specific region, not the national average. A 2,600-gigawatt national queue number is nearly meaningless for a single project; what matters is the utility-specific study timeline and whether a developer holds a signed interconnection agreement, not just a submitted application.
If the honest answer to “is this a bubble” is “some of it, in ways that are genuinely hard to isolate from outside,” the more useful question is which of these five signals moves first. That’s the one that will show which part of the buildout was real.