Who Signed for AI’s Bill

ISSUE 31 COMPANION · FRIDAY 2 OCTOBER 2026

Who Signed for AI’s Bill

A map of the promises behind the AI build-out: what is owed regardless, who guarantees whom, and how a usage shortfall would travel.

Eleven words in a confidential prospectus say more about AI economics than most earnings calls. Anthropic’s $31.4bn commitment to Microsoft, Reuters reported, “is non-cancelable except in the event of Microsoft’s uncured material breach.”

Translated: Microsoft is owed the money unless Microsoft breaks the contract. Not if demand slows. Not if a cheaper model arrives. Not if Anthropic’s revenue disappoints. Whether Microsoft actually collects is a separate question, and it depends on Anthropic having the cash.

That gap, between being owed and being paid, is the AI build-out in miniature. Every large project now rests on someone’s promise to pay whatever happens. The useful question is who wrote each promise, and what stands behind it. This companion maps those promises, works through the two that matter most, and sketches how a usage shortfall would reach each party.

THE ANNOUNCED STARTING POINT

According to Reuters, Anthropic’s confidential draft prospectus lists at least $518bn of computing commitments with six partners over about a decade, roughly 80% of it non-cancellable or payable regardless of usage: at least $111.1bn with Google, $110bn with Amazon and $31.4bn with Microsoft, plus about $161.2bn of Broadcom-related equipment leases that neither side can cancel except on default. If Anthropic spends less than its Google commitment, it pays the difference.

Reuters via KSL · $518bn, mostly non-cancellable · 29 Sep | PYMNTS · Google shortfall clause · 29 Sep

The exceptions are telling. Agreements with xAI worth up to $84.5bn of Nvidia-based capacity through 2029 can largely be cancelled on 90 days’ notice. And AMD has agreed to buy up to $5bn of Anthropic stock while supplying computing capacity expected to exceed $20bn.

Express Tribune / Reuters · xAI and AMD terms · 29 Sep

On the supplier side, Nvidia’s July 10-Q lists $36bn of AI cloud agreements, typically six years long: capacity Nvidia commits to buy from cloud operators that buy its hardware. The commitments shrink as third parties or Nvidia’s own researchers use that capacity, and if certain criteria are met Nvidia shares in the revenue. It also has $3.5bn of guarantees on partners’ leases, reduced as the partners pay down. In August it capped residual value guaranties on OpenAI’s Ohio leases at $105bn; they apply only once the sites meet ready-for-service conditions, expected from 2028, so they’re a contingent obligation, not a loss estimate. Nvidia has also said it may support up to 25% of individual financing deals, and this week the FT reported talks with insurers about covering lenders to neoclouds.

Nvidia 10-Q · Commitments and guarantees · 26 Jul | Nvidia 8-K · $105bn residual value guaranties · 17 Aug | The Decoder · Up to 25% support per deal · 12 Aug | Investing.com · FT: talks with insurers · 29 Sep

None of this is unusual any more. Meta’s Hyperion campus in Louisiana sits in a joint venture 80% owned by Blue Owl funds and financed with about $27bn of debt, with Meta guaranteeing part of the campus’s residual value for its first 16 years. S&P rated that debt A+, and the guarantee was central to the rating.

Yahoo Finance · Hyperion venture and guarantee · Aug | Business Model Analyst · 16-year guarantee · Jul | Global Data Center Hub · A+ rating and the guarantee

Everything in the two worked examples below is invented. They illustrate mechanics. They are not estimates of any company’s contracts, costs or exposures.

FIVE KINDS OF PROMISE

Simplified; real contracts mix several. Owed is not collected: every row depends on the payer staying solvent.

My starting point: a promise moves risk; it doesn’t remove it. The useful habit is to follow each promise to the balance sheet that ends up holding it, then ask one question. Does that balance sheet weaken at the same moment the promise gets called?

WORKED EXAMPLE ONE: THE TAKE-OR-PAY BUYER

All figures hypothetical. A model maker commits to $50bn of computing a year. $40bn is a minimum it pays regardless; above that, it pays for what it uses. Every $1 of computing it actually uses earns $1.60 of revenue. Utilisation is the share of committed capacity it uses.

Per year. Illustrative only; ignores staff, training runs and price changes.

Read the columns separately. From 100% to 40% utilisation, what the cloud is owed falls by a fifth, from $50bn to $40bn, and then stops falling, provided the buyer can pay. The buyer’s gross profit falls from +$30bn to −$8bn. Same event, very different outcome, depending on which side of the minimum you sit.

That’s why Google, Amazon and Broadcom sat on the more resilient side of this week’s Angle. Not because they’ll grow faster than their customer, or because their shares are cheap; this example says nothing about valuation. It’s because, while the buyer can pay, the minimum absorbs a shock that would otherwise land on their revenue.

Two caveats

First, a minimum is only worth what the buyer can pay. A buyer stuck at 40% for long enough runs short of cash, and the seller discovers the minimum was a credit exposure all along. Anthropic’s draft shows a 2025 operating loss above $8bn, and two customers supplied nearly a quarter of its revenue. It also warns that its own access to compute could be cut back or repriced, a risk that runs in both directions.

Yahoo Finance · Operating loss · 28 Sep | Yahoo Finance · Customer concentration · 1 Oct

Second, 80% is an average across contracts. The cancellable lines are the buyer’s escape valve, and in a downturn they go first. Note which capacity is cancellable here: the xAI agreements, for Nvidia-based computing. I wouldn’t read much into one contract. I would note that the flexible end of a big buyer’s book is where suppliers find out about a slowdown first.

WORKED EXAMPLE TWO: THE GUARANTOR

Hypothetical again. A lender advances $10bn to a neocloud to buy GPUs, secured on the chips. A supplier guarantees whatever the chip sale fails to recover, up to 25% of the original loan, so $2.5bn. The neocloud defaults and the chips are sold.

Illustrative only. Real caps, triggers and terms differ deal by deal and are mostly not public.

Two things stand out. After the collateral is sold, the guarantor pays the first slice of what’s left unrecovered, up to its cap: in mild cases it absorbs the whole shortfall, in severe ones the lender carries the rest. And the column that drives everything, resale value, falls for the same reasons the neocloud defaulted: too much capacity, too little demand, newer chips.

That’s wrong-way risk. The guarantee gets called in exactly the quarter when the guarantor’s own sales are under pressure. Putting an insurer in the middle spreads the risk rather than reversing it, and the FT’s insurance correspondent has said many large insurers are already near their limits on AI exposure.

TNW · FT on insurers’ AI limits · 30 Sep

I don’t think this breaks Nvidia. Its caps are disclosed, its obligations are mostly conditional and its cash generation is enormous. It does change how I’d think about a dollar of its revenue: a sale that needed a guarantee to close carries a contingent cost that a hyperscaler sale doesn’t. How much that matters depends on how large those sales become relative to the rest.

IF USAGE DISAPPOINTS: A STRESS SCENARIO

These are separate companies, contracts and financing structures, not layers in one capital structure, so there’s no fixed order of losses. Share prices usually reprice first, often before any payment is missed. Contractual losses come later, if at all, and their size depends on cash, collateral and terms. Here’s how a year or two of weak usage could plausibly reach each party.

A conditional scenario, not a ranking. Timing and size depend on terms that are mostly not public.

24/7 Wall St · Jupiter debt below par · 24 Sep

Two loops could make a shortfall travel further than the table suggests. The first is circularity: AMD buying shares in a customer it supplies, Nvidia committing to buy capacity from clouds that buy its chips. The Bank of England’s Financial Policy Committee named opaque and circular arrangements this week as something that could amplify losses.

The second is rates. About $450bn of AI-related debt this year, the estimate the Bank cited, is part of why the ten-year yield touched 5.34%. Higher yields raise the cost of every layer below the top. The sector’s borrowing is helping raise the price of its own borrowing.

Bank of England · FPC record · 30 Sep | Quartz · 10-year at 5.34% · 1 Oct

WHAT TO WATCH

TURN THE ANALYSIS INTO A DECISION

The bull case: utilisation stays high, minimums are never tested, resale values hold and the guarantees cost nothing. In that world Nvidia is a winner we’re underrating, and the promises were cheap marketing. Nothing this week rules that out.

The case weakens if a public filing shows usage running below commitments, if guarantee and commitment totals keep rising faster than revenue and cash flow, or if GPU loans start trading like Jupiter’s. Last week’s question was who gets paid first. This week’s is whose balance sheet stands behind the payment.

This companion supports Issue 31 of iPrompt Signals, the Friday briefing on AI and robotics investing. If someone forwarded it to you, subscribe at [SUBSCRIBE LINK] to get the next issue, with this week’s tripwires checked against the results.

For information and education only, not financial advice. Figures in both worked examples are hypothetical. Anthropic’s figures come from reporting on a confidential draft that hasn’t been made public. iPrompt Signals is not a registered investment adviser.

iPrompt Signals Deep Dive · Issue 31 · 2 October 2026 ·