Five Years of Capex Already Signed — What the $3T Tally Actually Locks In

Builds-on: the-shadow-balance-sheet-nikkei-1-65t-and-the-spv-layer, compute-as-collateral-the-residual-value-wrap, cyclical-20-and-the-ai-capex-mask Related: the-mask-that-eats-what-it-hides, ai-circular-financing-and-banking-exposure-audit, ai-infrastructure-endgame-indicators, why-the-market-refuses-to-crash, mechanism-vs-narrative-method

The Wall Street Journal published a tally today (August 17, 2026) putting nine tech companies' off-balance-sheet AI commitments near $3 trillion. Nikkei Asia's count on July 23 was $1.65 trillion.

the-shadow-balance-sheet-nikkei-1-65t-and-the-spv-layer already worked the Nikkei number, the SPV mechanics, and the per-company breakdown, and its perimeter ladder — $120B of SPV flows inside $800B of private-credit data-center financing inside $1.65T of off-balance-sheet obligations inside a $3T total — named this figure three weeks before the WSJ counted it. So the news is not the number. The news is what closing the gap reveals about composition, and one consequence of that composition that has not been drawn anywhere.

Reconciling the two counts

They are not measuring the same thing, and the delta decomposes cleanly.

Three distinct sources of the increase, and they should not be blurred:

1. Scope. Four companies added, and the significant ones are the guarantors — Nvidia, Broadcom, AMD. Nikkei counted the buyers' hidden obligations. WSJ counted the sellers' too. Given what the previous doc established about who is standing behind whose collateral, counting only one side was always going to understate the system.
2. Category. Residual-value guarantees are in the WSJ number and were not really in Nikkei's. That is the compute-as-collateral-the-residual-value-wrap instrument entering the official tally for the first time.
3. Actual new signing. The Nvidia $500B platform MOUs (Aug 11–13) and Broadcom's $35B AI XPV first transaction both landed between the two reports. Part of this is not a measurement revision at all. It is three weeks of deal flow.

Earlier waypoints for the trajectory: Moody's counted $662B of unrecognized lease commitments in February 2026 against $969B of total undiscounted future lease commitments. Six months, roughly a fivefold move in the measured figure.

One correction before building on it

The headlines say "hidden debt." Most of this is not debt.

A take-or-pay commitment to buy chips is a contractual obligation, not borrowed money, and a large share of it will be funded from operating cash flow that these companies actually generate. Comparing $1.65T of commitments to $1.35T of disclosed debt — as the Nikkei coverage did — is apples to oranges, and $3T against "combined leases and long-term borrowings" has the same problem. Purchase obligations have always lived in the contractual-obligations footnote. They are disclosed. Nobody hid them; nobody added them up.

What is fair to say is that the people who price capital have stopped treating the distinction as decisive. Moody's and S&P have both flagged that reported leverage understates economic exposure, S&P has already applied rating pressure to Oracle, and S&P is adding Broadcom's residual-value support to adjusted debt outright. This is a measurement regime change, not a leverage regime change — and that distinction matters, because measurement regime changes are how credit cycles usually turn. The leverage is rarely a secret. What changes is that somebody adds it up and the rating agencies start counting it.

The structure the tally exposes (briefly — the vault has this)

Beignet Investor LLC — the Hyperion SPV, Blue Owl ~80% / Meta ~20%, $27.3B of A+ paper maturing 2049 against two-to-six-year silicon — is documented in the-shadow-balance-sheet-nikkei-1-65t-and-the-spv-layer and not repeated here. The one detail worth carrying forward is the instrument: Meta leases the capacity back triple-net and provides construction support, payment guarantees, and a residual-value guarantee, and under lease accounting the obligation stays off its balance sheet until rent starts being paid. Meta's total off-balance-sheet AI obligations run ~$420B against $83.7B of reported debt; its future lease commitments alone are $347B. Alphabet discloses $811B of purchase and contractual obligations. Oracle is at ~$273B, up more than thirtyfold in four years on Stargate.

The instrument doing the work is the same one from compute-as-collateral-the-residual-value-wrap, applied from the other end:

• Vendor-side: Nvidia writes an RVG on a neocloud's GPU debt. The lender underwrites Nvidia's credit through a neocloud-shaped hole.
• Tenant-side: Meta writes an RVG on its own SPV's debt. The lender underwrites Meta's credit through a Blue Owl-shaped hole.

In both cases the residual-value guarantee is precisely what converts project-level risk into investment-grade paper, and in both cases it remains a contingent liability off the guarantor's balance sheet. That is not two coincidental structures. The RVG is the load-bearing member of the entire AI capital structure — the single component that makes $3T of obligations financeable at investment-grade cost. It is also, per the previous doc, a written put on a hardware price nobody could quote until the CME lists a curve on October 5.

What this actually locks in

Here is the number that reframes the whole thread. The nine firms' traditional annual capex runs roughly $600B. The forward commitments are ~$3T.

Five years of capex is already contracted.

The cyclical-20-and-the-ai-capex-mask thesis assumed the mask persists because cash-rich firms don't need to stop. compute-as-collateral-the-residual-value-wrap revised that to: the mask persists longer because it switched to leverage. This is the third and strongest version. The mask persists because it is contractually pre-committed and cannot be turned off quickly even if every board in the cohort wanted to.

A take-or-pay obligation is not a plan. It is a debt that happens to be denominated in GPUs and megawatts. If AI demand disappears tomorrow, the purchases still happen. That produces a macro shape genuinely unlike 2007–09, and it is worth being precise about it:

In 2007 the investment stopped and the losses followed. In 2026 the investment is signed, so the investment does not stop — the returns stop. The capex keeps flowing into GDP while the income statements absorb depreciation on assets that don't earn (cyclical-20-and-the-ai-capex-mask's $65–110B of incremental annual D&A from the 2Q25–1Q26 cohort alone), and the credit layer absorbs the residual-value calls.

This makes the recession call later and the equity call earlier. Those usually move together. Here they separate, because the contractual structure decouples the spending from the sentiment that produced it. It also means the cleanest early evidence will not be a capex guidance cut — the thing the ai-circular-financing-and-banking-exposure-audit watchlist was waiting for. It will be an impairment, a lease-exit negotiation, or a rating action on a guarantor. Two of those three have already happened to Broadcom and Oracle.

flowchart TB
  A[3T of forward commitments<br/>vs 600B annual capex] --> B[Spending is signed<br/>through roughly 2031]
  B --> C[GDP contribution<br/>keeps printing]
  B --> D[Assets arrive whether<br/>or not demand does]
  D --> E[Depreciation lands<br/>on the income statement]
  D --> F[Residual value falls<br/>below guaranteed minimum]
  E --> G[Earnings break first]
  F --> H[Guarantee called<br/>on Nvidia, Broadcom, Meta]
  H --> I[Rating action<br/>on the guarantor]
  I --> J[Cost of capital rises<br/>for the next tranche]
  C --> K[Headline GDP still<br/>says expansion]

The path from the bottom left to the bottom right is what the thread has been trying to date. This diagram says it does not run through the capex line at all.

Revision to the thread's calibration

ai-circular-financing-and-banking-exposure-audit put "equity correction without banking crisis" at 50–60% and the severe scenario at 10–15%, with the severe path defined as hyperscaler capex pullback triggers cascading credit revaluation. The pre-commitment finding makes that specific trigger less available — a pullback is contractually hard — while making the credit path more available by a different route: the guarantee call, which does not require anyone to change their mind about anything.

Restating the transmission without changing the top-line odds much:

• Base case gets slower and longer. The equity correction can happen while capex keeps printing. Expect the discourse to be confused by this for several quarters — "the buildout is still growing, so the bears are wrong" will be technically true and analytically empty.
• The severe path's mechanism changes. Not "hyperscalers cut capex" but "residual value falls below guaranteed minimums, guarantees are called, guarantors are downgraded, and the marginal tranche stops clearing at investment grade." That chain runs on a hardware price, not on a management decision, which makes it both more predictable and less negotiable.
• The 12–18 month gap between contraction and NBER call should be widened. With five years of committed spending in the cyclical-20's equipment line, the headline print has more support than the original doc assumed.

Watchlist additions

• Whether FASB or the SEC responds. Everything here is legal under current GAAP. A disclosure rule requiring SPV consolidation or RVG notional on the face of the balance sheet would be the single largest repricing event available, and off-balance-sheet vehicles being pulled onto balance sheets is the classic late-cycle regulatory reflex.
• The Nikkei-to-WSJ gap closing in the other direction. If a subsequent tally comes in below $3T, that is scope disagreement, not deleveraging. Watch the methodology, not the headline.
• Any lease exit or renegotiation on an uncommenced facility. $1.2T of the total is leases on things not yet in service. That is the most cancellable tranche, and the first cancellation is Zitron's second Horseman arriving with a dollar figure attached.
• S&P extending the Broadcom adjusted-debt treatment to Meta's RVGs. Broadcom set the precedent as a vendor. Applying it to a tenant's SPV guarantee is the step that puts the structure on the balance sheet by rating-agency fiat rather than by accounting rule.
• Burry's position. He took a victory lap on this tally, having called the depreciation understatement first. His next specific claim is worth reading as a leading indicator of where the short interest concentrates, independent of whether he is right.

Sources

• Why big tech's AI spending is $3 trillion higher than it seems — WSJ via MSN
• Big Tech companies have $3T in hidden AI expenditures — Seeking Alpha
• Hyperscaler off-balance sheet obligations hit $3 trillion — Crypto Briefing
• Off-Balance-Sheet AI Debt Approaches $3 Trillion Across Tech Giants — Stratton Journal
• Meta's Balance Sheet Hides $420B in Off-Balance-Sheet AI Debt — TechTimes
• Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding — Nikkei Asia
• Hidden Debt at Five AI Hyperscalers Hits $1.65 Trillion, Nikkei Study Finds — Analysis.org
• AI's $1.65 Trillion Debt Story Is Mostly About One Company — Advisorpedia
• Michael Burry Takes Victory Lap as Big Tech's $3 Trillion AI Spending Tab Comes Into Focus — Benzinga
• Emerging Litigation Risks in AI Data Centers — Quinn Emanuel
• The $3 Trillion AI Data Center Build-Out Becomes All-Consuming For Debt Markets — Bloomberg