The money for AI infrastructure is flowing faster than ever, but it's being financed in unconventional ways that shift risk away from hyperscalers and toward chip suppliers, private-credit investors, and compute resellers. The real bottleneck stopping this buildout isn't the availability of graphics processing units (GPUs) or even capital. It's something far more mundane: the electrical grid simply can't keep up with the demand.

How Much Are Tech Giants Actually Spending on AI Data Centers?

Estimates for 2026 AI infrastructure spending vary widely depending on what analysts count. Global data center capital expenditure (capex) is projected above $1 trillion when including all cloud infrastructure, but the five largest US-listed AI infrastructure spenders,Amazon, Alphabet, Meta, Microsoft, and Oracle,are expected to spend between $660 billion and $690 billion specifically on AI-focused projects.

What's striking is how consistently forecasts have underestimated actual spending. A Goldman Sachs analysis showed that in 2024, hyperscalers guided for 19% AI capex growth but delivered 54% in reality. The same pattern repeated in 2025: 22% guided, roughly 64% realized. Every forecast has missed in the same direction for two consecutive years.

Why Aren't These Companies Using Traditional Financing?

If you're spending $660 billion on infrastructure, you'd normally expect companies to fund it through free cash flow. That's not what's happening. Instead, hyperscalers are using three unconventional financing structures that would have raised red flags in corporate America just two years ago.

• Vendor Equity Stakes: Nvidia and OpenAI announced a deal in September 2025 where Nvidia stated it "intends to invest up to $100 billion" as OpenAI deployed AI systems. However, by February 2026, reporting revealed the actual commitment was considerably smaller,an equity stake closer to $30 billion folded into an OpenAI funding round. Similarly, AMD signed OpenAI to a six-gigawatt GPU agreement and issued warrants for up to 160 million AMD shares, potentially close to 10% of the company.
• Off-Balance-Sheet Debt Vehicles: In October 2025, Meta and private-credit firm Blue Owl Capital closed a $27 billion bond deal to build Meta's Hyperion data center campus in Louisiana. Because Meta doesn't consolidate the joint venture's debt on its own balance sheet, the $27 billion in bonds anchored by PIMCO and partly bought by BlackRock doesn't appear in Meta's financial statements, even though Meta is the facility's sole tenant.
• Prepaid Customer Orders and Backlogs: Oracle's fiscal year ended in May 2026 with a $638 billion backlog of contracted AI infrastructure work,up 363% year over year,while simultaneously reporting negative free cash flow of $23.7 billion. This illustrates how companies are using customer prepayments and long-term contracts to fund buildouts without immediate cash outlays.

What Happens If the Compute Doesn't Generate Revenue on Schedule?

The financing structures create a hidden risk: if AI infrastructure doesn't monetize as quickly as expected, the exposure falls on chip suppliers and private-credit investors, not the hyperscalers themselves. When Nvidia takes an equity stake in OpenAI or AMD issues warrants vesting only if stock prices hit specific targets, both companies absorb disappointment directly as shareholders if growth curves bend. This is venture-dealmaking at an unprecedented scale, but applied to companies that are still losing money on an operating basis.

OpenAI's 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. But the structure means the risk profile has shifted fundamentally.

The Grid Is the Real Constraint, Not GPU Supply

Microsoft's own disclosures reveal the actual bottleneck. 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 constraints, but because of power limitations on orders that were already funded. This single detail previews the core problem: the money is largely there. Getting it turned into usable compute on schedule is the harder problem.

The US electrical grid interconnection queues held roughly 2,600 gigawatts of pending requests in early 2026, about double the capacity of the entire existing grid. This is the hard limit on how fast any AI infrastructure buildout can scale. No amount of capital, equity arrangements, or off-balance-sheet financing can overcome the physical reality that power infrastructure takes years to build and faces community opposition, regulatory delays, and environmental reviews.

Are Hyperscalers Hiding the True Cost of Their GPU Investments?

A secondary debate is emerging around GPU depreciation accounting. Michael Burry and other critics argue that hyperscalers are understating GPU depreciation by tens of billions of dollars annually, since the chips lose value as new generations arrive. Hyperscalers counter that their accounting reflects real multi-year utility across training, inference, and batch workloads,meaning the same GPU serves multiple purposes over its lifespan, justifying slower depreciation schedules.

This accounting disagreement matters because it affects how profitable these data center investments actually are. If depreciation is understated, the true return on capital is lower than reported earnings suggest. If hyperscalers are correct, the investments are more economical than critics claim. Either way, the financing structures described above mean that chip suppliers and private-credit investors are now bearing part of this risk directly.

Steps to Understanding AI Infrastructure Financing in 2026

• Track Backlog-to-Cash-Flow Ratios: When a company reports a massive backlog but negative free cash flow, it signals reliance on customer prepayments or off-balance-sheet financing. Oracle's $638 billion backlog against negative $23.7 billion cash flow is the clearest example of this pattern.
• Monitor Equity and Warrant Deals: When chip suppliers take equity stakes or issue warrants in their largest customers, it's a sign that traditional financing isn't covering the buildout. Watch for announcements of equity stakes, warrant issuances, or joint ventures between hyperscalers and private-credit firms.
• Watch Grid Interconnection Queues: The 2,600 gigawatts of pending power requests is the real constraint on AI infrastructure growth. Follow announcements about new power plants, grid upgrades, and interconnection timelines. These will determine whether the $660–690 billion in planned spending actually materializes on schedule.

The 2026 AI infrastructure boom is real, but it's being financed in ways that distribute risk differently than traditional corporate capital spending. The bottleneck isn't capital or chips,it's power. And power infrastructure moves on a timeline measured in years, not quarters.