The Full Pyramid: What Big Tech's $3 Trillion in Hidden AI Commitments Means for Healthcare Program Planning

A Wall Street Journal analysis published August 16 put a number on something that has been hiding in plain sight across the quarterly filings of the four largest AI spenders: Alphabet, Amazon, Meta, and Microsoft have accumulated approximately $2.4 trillion in off-balance-sheet AI commitments that do not appear in the debt figures most investors and analysts are watching. Add Oracle, Nvidia, and a handful of other major players, and the WSJ puts the total across nine companies at roughly $3 trillion.

The diagram the WSJ published to illustrate the finding is worth studying carefully. It shows a pyramid divided by a dotted line — on-balance-sheet obligations above the line, off-balance-sheet below. The above-the-line portion, what investors see in standard financial reporting, is a fraction of the total structure. The below-the-line portion — the part that doesn't appear in headline debt figures — dwarfs it.

For healthcare organizations building AI programs on top of hyperscaler infrastructure, this is not a financial story to hand off to the CFO and move on from. It is a vendor stability and program continuity story with direct operational implications.

What the Pyramid Actually Shows

The WSJ diagram breaks the commitments into two off-balance-sheet categories. The first is leases not yet started — data center space under contract that hasn't come online — totaling $904 billion across the four companies. Amazon leads at $137.2 billion, Microsoft at $329.1 billion, Meta at $347.0 billion, and Alphabet at $91.0 billion.

The second category is purchase commitments — contracts for chips, power, networking equipment, and other infrastructure components — totaling $1.52 trillion. Here the breakdown is striking: Alphabet at $811 billion, Meta at $349.3 billion, Microsoft at $228.6 billion, and Amazon at $130.1 billion. Alphabet's purchase commitments alone exceed the combined on-balance-sheet long-term debt of all four companies.

For comparison, the on-balance-sheet figures — what appears in standard financial reporting — show lease liabilities of $248 billion and long-term debt of $356 billion. The off-balance-sheet obligations are more than three times the visible debt. The WSJ analysis of the broader nine-company universe, including Oracle and Nvidia, puts the total off-balance-sheet commitment at approximately $3 trillion, with $1.65 trillion of that representing 122% of the companies' actual on-balance-sheet debt.

These commitments are disclosed in regulatory filings — they are not hidden in the sense of being fraudulent or concealed. They are disclosed in footnotes and supplemental tables in quarterly reports rather than in the headline debt figures that most financial coverage focuses on. The WSJ's contribution is making the aggregate visible.

Why This Matters Beyond the Financial Pages

Michael Burry, who shorted the mortgage market before the 2008 crisis, shared the WSJ graphic on August 17 and claimed vindication for warnings he issued in 2025 about AI capex. He added a forward-looking note about what he called "compression" risk — the possibility that the gap between AI infrastructure spending and AI revenue narrows violently rather than gradually.

The compression thesis is straightforward: the four hyperscalers are spending at a rate that requires AI services to generate returns that current enterprise adoption does not yet support. They are on pace to spend $730 billion on AI infrastructure in 2026 alone. JPMorgan raised its estimate for global AI-related capital expenditures through 2030 to $5.5 trillion, with AI-related debt financing projected at $4.1 trillion. The current build cost runs approximately $50 billion per gigawatt of compute capacity, and companies are now borrowing rather than funding purely from operating cash flow.

The 30-day correlation between major AI spenders and semiconductor stocks has collapsed from +0.78 to near zero, the lowest reading in 4.5 years — a signal that investors are beginning to price the spending and the returns separately rather than treating them as the same trade.

None of this means the AI infrastructure buildout collapses. The hyperscalers have enormous balance sheets, strong operating cash flows, and long track records of investing ahead of demand. But the off-balance-sheet commitment structure means that if a correction comes, it arrives with significant pre-committed obligations that constrain the companies' ability to pull back quickly. They cannot simply stop spending — they have signed contracts.

What This Means for Healthcare

Healthcare AI programs are not exposed to hyperscaler balance sheets directly. But they are exposed to the downstream consequences of a hyperscaler capex cycle correction in ways that are worth thinking through now rather than after the fact.

AI service pricing is downstream of this spending cycle

The current economics of cloud AI services — inference pricing, API costs, model access fees — are partly a product of hyperscalers competing aggressively for enterprise AI workloads during a buildout phase. If the spending cycle turns and capital becomes more constrained, the pricing environment for AI services can change. Healthcare organizations that have built budget models around current AI service costs should be tracking this dynamic. A 20% increase in inference costs is not a catastrophic event, but it is a budget planning variable that healthcare finance teams are not currently modeling.

Vendor stability assumptions in healthcare AI procurement need to account for this

Healthcare vendor risk assessments typically evaluate financial stability through balance sheet metrics — debt ratios, cash on hand, operating margins. The off-balance-sheet commitment structure the WSJ identified means that standard balance sheet analysis understates the obligation load for the hyperscalers and the enterprise AI vendors that depend on them. Healthcare organizations procuring multi-year AI contracts with vendors whose infrastructure runs on hyperscaler commitments are inheriting some exposure to this dynamic indirectly. Vendor financial due diligence should be accounting for disclosed off-balance-sheet AI commitments, not just headline debt figures.

The AI program sustainability question is now a governance question for healthcare boards

The Bank for International Settlements flagged AI bust risk alongside inflation and fiscal stress as top global threats in its 2026 annual report. The WSJ's $3 trillion off-balance-sheet finding is the mechanism behind that risk made concrete. Healthcare boards and audit committees that are evaluating AI program investments should be asking whether their AI program plans are resilient to a scenario where hyperscaler AI services become more expensive or less available due to a capex correction. That is not a prediction — it is a scenario planning question that prudent governance requires.

The energy infrastructure constraint is the longest-duration risk for healthcare AI

The WSJ pyramid includes significant lease commitments for data center space that hasn't come online yet. The constraint on that construction is power — global AI compute demand is projected to require 55 gigawatts of new power capacity by 2030. Healthcare AI workloads that depend on specific regional cloud infrastructure could face availability constraints not from market dynamics but from physical infrastructure limitations — power grid capacity, data center construction timelines, cooling infrastructure. Healthcare organizations with regional data residency requirements for PHI should be evaluating whether their preferred regions have the power infrastructure pipeline to support the AI compute capacity they will need over a three-to-five year horizon.

The Bigger Picture

The WSJ pyramid diagram illustrates a fundamental characteristic of the current AI infrastructure cycle: the spending commitments that will determine AI service availability and pricing for the next five years have already been made. The contracts are signed. The data center leases are in place. The chip purchase agreements are locked. What has not yet been determined is whether the AI revenue those commitments are supposed to generate materializes at the scale and speed the spending assumes.

Healthcare organizations are not positioned to influence that outcome. They are positioned to plan around it — building AI programs that are robust to pricing changes, vendor instability, and infrastructure availability variation rather than assuming that the current hyperscaler spending cycle produces stable, predictable, and inexpensive AI services indefinitely.

The visible tip of the pyramid is what most AI program planning assumes. The full pyramid is what healthcare security and governance programs need to account for.

For related coverage, see The Flight Recorder for AI Agents: What the SAFE Framework Means for Healthcare Security Programs and OpenAI Opens the Door for Defenders: Daybreak Red, GPT-5.6-Cyber, and What the 95% Completion Rate Means for Healthcare Security.