
The machine is worth something for two or three years. The contracts that pay for it run for up to twenty. The power underneath them is a twenty-year commitment. And the demand that is supposed to pay for all of it is flat. Those clocks don’t agree. The gap between them is the part the market has not priced yet.
Everyone is asking the same question right now, and two big groups have formed. Is AI a bubble?
On the surface, it looks like one. The spend is running far ahead of anything we can prove.
So here is where I land. I don’t think we are in a bubble. But there is a mechanism in this buildout that makes it look exactly like one. And you can’t see it from reading the headlines. And if you can’t name that mechanism, you won’t be able to separate the real risk from the noise.
I will name it because it has a name. The amortization gap. And it’s sitting in a filed document.
In November I wrote that the bill on AI is no longer a promise. Anthropic put more than $750 billion of compute commitments in a prospectus, and most of it cannot be canceled. Later, the number climbed. In its latest amendment, filed with the SEC, Anthropic puts its aggregate committed spend at approximately $900.9 billion, substantially all of it on compute and infrastructure. It is take-or-pay. The bill is locked before the demand is proven.
The document has clocks in it. And the clocks don’t agree with each other.
1. The paper trail
Let’s understand what the filing says.
The largest single piece is an equipment lease with Broadcom, around $161 billion, to secure TPU-based capacity.
Google: at least $121 billion, running to July 2033.
SpaceX’s AI unit: up to $120.5 billion, extending into 2029.
Amazon: at least $110 billion, running to April 2036, with a clause that if Anthropic spends less, it still pays the difference.
Microsoft: between $47 and $49 billion, through June 2033.
And a further spread: Nscale at roughly $45 billion, Lambda at roughly $41 billion.
On top of the table, the filing flags another $125 to $150 billion of compute that does not even specify a fixed dollar amount.
Read that as a set of clocks, not a spending plan. About $71 billion of it comes due within a year. About $452 billion by the end of 2030. And the bulk of the rest is incurred between 2031 and 2045. And the large ones are take-or-pay; the filing says “regardless of actual usage.” You pay whether or not anyone uses it.
2. The machine clock
Now put a second clock next to those. The machine itself.
The chip that does the work has a useful life, and it is not ten years. The companies depreciate their GPUs over five or six years. But the economic life of the chip or the point at which a new generation makes the old one uneconomic, is closer to two or three years. The operators who actually buy and run them say so. CoreWeave’s chief executive has put the useful life of a GPU at two to three years, and the short-seller Michael Burry has argued it is under three. Amazon itself changed its books. It cut the useful life of a subset of its servers from six years to five, on the stated reason of an increased pace of technology development.
So the first gap is already in the books. They are spreading the cost of a three-year machine over five or six years, and then signing contracts that run for a decade or more.
3. The power clock
And under the machine there is a third clock. The power.
A data center is more than chips. It’s a power plant, a grid connection, a substation. In the filing, the infrastructure arrangements carry initial terms of twenty years, with take-or-pay compute obligations and uncapped fuel cost pass-throughs that run for the duration of the term. That is the longest-lived thing in the whole stack, and it is amortized on the longest clock of all.
So now you have three clocks. The machine is worth something for two or three years. The contracts lock you in for up to twenty. The power is a twenty-year commitment. The cost is spread across the long clocks, for an asset that decays on the short one. That is the amortization gap.
4. The demand side
And there is a fourth clock. The one the first three are supposed to be paid by. The demand.
Here is where the data stops being a theory. Ramp publishes an index of how American businesses actually spend on AI, from real transaction data, and independent reporting on that data tells the same story. The typical American company spends in the low teens per employee, per month, around $12 to $15. The top ten percent spend in the hundreds, up to about a thousand. Only the top one percent spends like a data center, close to ten thousand dollars per employee per month, near the peak it hit over the summer. These numbers had an increasingly decreasing trend (to wordplay a little) during the last weeks, maybe months. Employees who just yesterday had $100 tokens a day got $25 the next day.
The concentration is the tell. Roughly 80% of spending on OpenAI and Anthropic comes from 1% of firms. More than half of businesses already pay for something, so the growth now has to come from using it harder, not from new customers.
And on the flatness: token spend (the actual consumption) went flat for two straight months, August and September, even as usage kept rising. It re-accelerated in October. But the base is low, the growth is lumpy, and it sits in the hands of a very few buyers. And the contracts that were supposed to lock that demand in are, in the words of Ramp’s lead economist, “not particularly restrictive.” The buyer can stop. The price is falling.
So the demand is not locked in either.
The reframe
So here is what the gap actually shows. It’s not a demand problem, at least not only. It’s a timing problem. The buildout is being amortized on a twenty-year clock. The contracts lock in for up to twenty years; the bulk of the $900.9 billion is incurred between 2031 and 2045. The machine is worth something on a two- or three-year clock. And the demand is on no clock at all. It’s flat, it’s concentrated, and it can walk. For now. This is the snapshot in time we have now.
When I wrote about the bill early in November, the point was that the market finally had a document to price. The point is that the document has a timing problem in it. The cost is being spread over decades, for a machine that lasts a few years, on demand that is not growing. And that gap is the difference between the depreciation schedule in the books and the clock the asset actually runs on.
Why the market has not priced it yet
The reason is that it’s hidden. The depreciation schedule is an accounting choice. You can spread the cost of the machine over five years or over three, and the cash you pay is the same, but the profit you report is different. So the gap sits inside the numbers, in a line that most people don’t look at. It’s an assumption. And assumptions are the last thing the market prices, because there is nothing to point at.
Until there is a clock in it.
The bottom line
The amortization gap is the most expensive risk in the stack, because it is the one nobody names. You can argue about the valuation. You can argue about the model. But you can’t argue with a clock. A machine that lasts three years, or even six, doesn’t become a twenty-year asset overnight because you sign a twenty-year contract. It just means you are paying for the years after it’s no longer there.
Likely to happen. Within a quarter, one of the large data center investors shortens its GPU depreciation schedule in a filing, on the stated reason of faster technology change. That’s the tell. It is the market admitting the machine clock is shorter than the contract clock.
Less likely, but possible. The Anthropic offering prices at its target and the market funds the bill. That would mean the market is choosing to amortize the gap, to pay for three-year machines over a twenty-year clock, on demand it knows is flat. That is not a strategy. It is a bet on the clock.
When the bill is a document and the document has clocks in it, the question stops being whether you will pay. It becomes which clock you are being held to, and who is left holding the machine when the other clocks have already moved on.
The most expensive risk is the one your organization cannot name.
If your organization has committed to compute on a twenty-year clock, but your models become uneconomic in three, you are carrying an amortization gap with a countdown. The Mirror Test exists to surface it before it gets expensive.
The white paper is free, gated by email. The Forensic Read is the 14-day diagnostic.
Cristina Imre is a forensic architect of organizations and the creator of the Mirror Test Framework™ and People Debt™. She helps boards and executives read the human system behind the numbers, before the distortion reaches the balance sheet.
Sources: Anthropic S-1/A, filed with the SEC 17 Nov 2026 (aggregate committed spend, per-counterparty commitments, take-or-pay terms); Ramp AI Index, Oct 2026; TechCrunch, 6 Nov 2026 (per-employee AI spend); CNBC, 14 Nov 2025 (GPU depreciation — Michael Burry, CoreWeave); Amazon FY2025 10-K (server useful life, 6 to 5 years).
🍀
