Twenty Years or Ninety Days

ISSUE 32 COMPANION · FRIDAY 9 OCTOBER 2026

Twenty Years or Ninety Days

How long each part of the AI build-out is promised for, how long it’s financed for, and what happens in the gap between the two.

Two of Google’s capacity contracts this year were signed four months apart, and they could hardly look less alike. In June, SpaceX’s prospectus showed Google paying $920m a month for about 110,000 Nvidia GPUs until mid-2029, on terms reported to let the deal end on 90 days’ notice once that right opens after 31 December 2026. On Tuesday, Google signed with Constellation Energy for power from 11 reactors, for 15 and 20 years, the longer deal running from 2028 towards the end of the 2040s.

Same buyer, same goal. Why the difference? One reading is scarcity: Google expects power to stay harder to find than chips. Another is replacement: Google called the SpaceX deal ‘bridge capacity’, which suggests it expects its own data centres to take over. Both can be true. For an investor, the motive matters less than the consequence. A short contract leaves the risk with whoever owns the asset; a long one moves much of it to the buyer.

TNW · Contract terms · 6 Jun | The Decoder · ‘Bridge capacity’ · Jun | The Star / Reuters · Constellation terms · 6 Oct

THE THREE-DATE TEST

For any AI asset, find three dates. First, the earliest a customer can leave: the first day notice can be served, plus the notice period. Second, when the debt behind the asset must be repaid, from the maturity or repayment schedule, not just the size of the loan. Third, how long the asset can keep earning a competitive return. That third date isn’t the accounting life. The big clouds depreciate servers over five to six years, going by their 2025 filings; whether a GPU still earns a decent rent in year five, against newer chips, is one of the market’s louder arguments.

Webull · Hyperscaler depreciation lives · Dec 2025

Here is what public reporting shows for this week’s examples, and what it doesn’t.

‘Not disclosed’ means we couldn’t find it in public reporting this week, not that it doesn’t exist. Exit dates are the earliest reported.

Bloomberg via Advisor Perspectives · June bonds due 2056 · 7 Oct | The Star / Reuters · $40bn talks · 7 Oct | Reuters via KSL · Anthropic’s leases · 29 Sep | Korea Herald · SK Hynix · 29 Jul | Humanoids Daily · Agility · 6 Oct

My starting point: look at whichever date comes first. If it’s the customer’s exit, someone is lending against revenue that may not be there, and the question is who. If it’s the debt, the risk is refinancing. If it’s the asset’s economic life, the contract outlives the thing’s usefulness, and that hurts the buyer, not the owner. The table already shows the hard part: several of the dates aren’t public.

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

WORKED EXAMPLE ONE: THE GPU LANDLORD

All figures hypothetical. A landlord borrows $10bn at a fixed 8% to buy GPUs, repaid in five equal annual payments that combine interest and principal. Its one customer pays $3.6bn a year and can leave on 90 days’ notice. Running costs, mostly power and staff, are $0.6bn a year. The table shows any year after the customer leaves, depending on what a new tenant will pay.

Per year. Illustrative only; ignores taxes, fees and empty months between tenants.

The cushion is thin even when nothing goes wrong: $0.5bn on $3.6bn of rent. The landlord breaks even if a new tenant pays about 86% of the old rent. Below that, the debt payment doesn’t shrink, so the shortfall comes out of cash or new borrowing.

Now compare alternative loan rates for the same deal. Borrowing at 6% would lower the break-even to about 83%; at 10%, it rises to about 90%. A fixed-rate loan already signed wouldn’t move, but the next loan, or a refinancing, would. That’s why a ten-year yield near 5.3% matters to anyone still raising money for chips.

Then add the notice period. A customer that can leave in 90 days is most likely to use that right when cheaper capacity is available, which is also when re-letting prices are weakest. Last week we called that wrong-way risk when it arrived through a guarantee. Here it arrives through a contract clause.

A caveat: SpaceX isn’t a typical landlord. It has $93.5bn of cash and businesses that have nothing to do with GPUs. The example shows the shape of the risk, not its size relative to SpaceX, and if chips stay short, nobody serves notice at all.

TIKR · SpaceX cash and debt · 7 Oct

WORKED EXAMPLE TWO: THE LONG CONTRACT HOLDER

Hypothetical again. A generator sells 1,000MW for 20 years at a fixed $80 per megawatt-hour. At 95% availability that’s about 8.3m megawatt-hours a year, or about $666m of revenue whatever the market does. The table compares that with selling the same output at market energy prices.

Per year. Illustrative only; real contracts often index prices and share some risks.

A long contract is insurance, and insurance cuts both ways. If energy gets cheaper, the seller is protected; if it gets dearer, the seller hands the upside to the buyer for 20 years. PJM’s capacity prices, a separate payment for being available at peak, have risen more than elevenfold since 2024. That’s a different market from the energy price in this example, but it shows how tight the grid has become, and why an undisclosed contract price matters more than the megawatts.

What a long contract does buy is a lender’s confidence: the customer can’t leave in 90 days, so the seller can borrow and build against the revenue. The risk left over is the seller’s own, delivering the uprates on time and on budget. Whether Constellation’s contract penalises a late start isn’t public.

The Star / Reuters · PJM capacity prices · 6 Oct | Stocktwits · Uprate timeline to 2032 · 6 Oct

IF SUPPLY CATCHES UP: A STRESS SCENARIO

These are different businesses with different contracts, so there’s no single order of losses. Share prices would probably move first, credit next and contracts last. Here’s how a glut of computing capacity in 2027 or 2028 could plausibly reach each.

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

Two things could carry a glut further than the table suggests. The flexible end of the market is concentrated: the cancellable capacity in recent filings runs on Nvidia chips, and Nvidia owns about 122.8m SpaceX shares and works with Apollo, SpaceX’s expected lead lender, on its computing-finance platforms. And rates are high: the Fed raised in September, most officials expect another rise by year-end, and at least $150bn of new chip debt is looking for buyers.

TIKR · Nvidia on every side · 7 Oct | Scotsman Guide · Fed minutes · 7 Oct

WHAT TO WATCH

TURN THE ANALYSIS INTO A DECISION

The bull case for the borrowers: chips stay scarce through 2027, nobody serves notice, and buying GPUs with today’s debt turns out cheap. Nothing this week rules that out.

The thesis strengthens if SpaceX’s debt prices wider or smaller than planned, SK Hynix confirms 2027 prices, and Constellation’s uprates stay on schedule. It weakens if notice periods lengthen, if SpaceX’s debt prices tightly despite the clauses, or if memory pricing for 2027 stays open. Last week’s question was whose balance sheet stands behind the payment. This week’s is for how long.

This companion supports Issue 32 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. SpaceX’s contract terms are as reported from its prospectus, and its financing talks may not lead to a deal. iPrompt Signals is not a registered investment adviser.

iPrompt Signals Deep Dive · Issue 32 · 9 October 2026 ·