← Blog

The Compute Watch: Tracking the Rent on AI Chips

Lida Liberopoulou ·6 October 2026 · CC BY-SA 4.0

Raw .md ↗ Share on LinkedIn ↗ Share on X ↗

Independent research on AI economics and infrastructure · Commissioned research

Companies that build AI rarely own every chip they use. Many rent them, from the big cloud companies, like Amazon, Microsoft and Google, or from smaller firms that do almost nothing but rent out AI chips. They charge by the hour, per chip, and they even offer discounts if you commit for months or years.

Now Wall Street wants to let people bet on that rent, on whether it will go up or down. These bets are called futures, and they already exist for oil, wheat and orange juice. Two of the biggest exchanges in the world, CME Group in Chicago and Intercontinental Exchange (ICE), which owns the New York Stock Exchange, want to sell these bets. Neither is allowed to yet.

Intercontinental Exchange started a test run on 5 October. It shows banks and trading firms what the prices would be. The Commodity Futures Trading Commission (CFTC), the US agency that oversees futures markets, hasn't said yes. Instead it has asked the industry for comments, and says it understands the market for computing power to be fragmented, with prices set mostly in private deals.

But despite not having the OK, they're going ahead, and the reason is a race. Nobody can place these bets yet. So the two exchanges are racing for something else: who will ultimately set the standard price for chip rent. CME's bets would follow one company's prices, ICE's first ones another's, and the two don't even agree on what the newest chip rents for. ICE has a third index lined up too, which prices compute per unit of energy. Whoever wins sets the price the whole industry will quote, lend against and bet on.

And that price isn't only for betting. It's what pays for the chips. It's a good reason to look at what the rent depends on, and who is relying on it.

What rests on the rent

The computers that run AI, and the buildings that house them, are being bought with borrowed money. A lot of it. Goldman Sachs Research counts close to $500 billion borrowed for AI this year, by early August alone. Morgan Stanley reckoned last year that through 2028, about half of all the money spent on data centres, most of it for AI, would have to come from somewhere other than the tech giants' own pockets. Much of it will be lent by funds that invest money for pensions and insurers, who expect steady returns over many years.

The plan behind all this is simple: the chips earn rent, and the rent pays back the loans. But a chip isn't oil. A barrel of oil from 2020 is as good as one pumped today. A chip is built to be replaced. A better one arrives every year or two, and the old one earns less. Today, a chip from 2020 rents for about a quarter of what the newest one does. Any PC gamer, or anyone who has ever had to run hardware, knows this.

Yet the owners' books treat chips as something closer to oil. Most of the big owners expect a server to stay in use for five or six years. CoreWeave, whose whole business is renting out these chips, spreads their cost over six years, and its own annual report lists, as a risk, being unable to put the chips back to work after their first contracts end. Some loans are paid off within the contracts behind them, and leave the risk with the customer. Others run longer, or rely on renting by the hour. Those assume that when today's contracts end, these chips will still rent for enough to pay back what's owed. Someone will still want them but the question is how much they'll pay. And the answer decides whether the funds that lent the money, and the pensions and insurers behind them, get the steady returns they expect.

So it all comes down to one number: what an older chip will rent for once something better is available. And that rent isn't set the way most people imagine.

What moves the rent

Much of the newest AI capacity can't be rented by the hour. The biggest labs own their chips, or lock them up in contracts that run for years. Nobody publishes how much of the total that is, but the examples all point one way. Even the CFTC says it believes most of the money sits in private deals, and has asked the industry what share happens at public prices. At CoreWeave, the biggest of the firms that rent out AI chips, almost all revenue, 98% last quarter, comes from contracts customers have committed to, mostly running over the next four years. Those chips are rented too, but to one customer at a time, for years, at prices nobody else sees. A smaller European provider, Scaleway, keeps about 70% of its chips on similar long contracts and offers the rest short-term.

The rent the indexes measure is the price of what's left over. How big a slice that is, nobody can say. And that leftover market is changing.

The biggest buyers are moving elsewhere. Google trains its own models on its own chips, called TPUs. Amazon, Microsoft and Meta are all building their own too. And the big labs are signing with AMD as well as Nvidia. But these same giants also rent huge numbers of chips from the specialist firms, on those multi-year contracts. Nobody breaks a contract like that to switch. They just may not renew it. And when a large one ends, a lot of locked-up capacity comes back onto the rental market at once.

More chips are coming, from both ends. At the top, each new Nvidia generation pushes the last one down a rung, and the rungs are getting closer. Nvidia used to bring out a new design every two years or so; now it promises one every year, and the next one is shipping now. AMD expects its rival rack systems to reach customers in volume before the end of the year. At the bottom, chips built only for running models, from Qualcomm and Intel, are due over the next year or two. Both companies say they will be cheaper to run than Nvidia's.

Only some old chips come back. When a newer chip replaces an Nvidia or AMD one, the old one can be resold or rented out cheaply, and it adds to the supply. It goes for a fraction of its original price, but it can still be traded. A replaced custom chip can't go anywhere. It's written off quietly inside the company that owns it.

So the price comes from fewer and fewer deals. If more of the biggest buyers move to their own chips and long contracts, fewer chips are left to rent by the hour. Think of house prices in a small village where only three houses sell a year. One owner who has to sell in a hurry drags down the price for the whole village. Because so few deals set the price, one big deal can move it for everyone: a large contract ending and its chips flooding the hourly market, or a big buyer suddenly needing capacity. And the CFTC is asking the same question: how to stop a thin market from producing runaway prices, and whether the companies that supply chips could move the very prices the bets would settle on.

Yet this is the price many people lean on. Owners use it to judge what their chips are worth, even the ones on long contracts. Firms that rent out chips by the hour live on it, and so do their lenders. And it's the price the new bets will pay out on. A price set by a few deals is being used to value, lend and bet on far more than those deals.

Everything above pushes rents down. But there are two things pushing them up, and for now they are winning.

The first is demand. People use AI more every month, and every time you ask AI to do something it needs a chip running somewhere to do it.

The second is room. New data centres take years to build and need huge amounts of power, so there isn't space for all the new chips people need right now.

Together these two things make a shortage. And in a shortage, almost any chip finds a renter, even an old one.

So the real question is how long the shortage lasts. One way to watch it is memory. There isn't enough memory to go round either, for the same reasons, and some of its prices are public. Its price works like a thermometer. If memory gets cheaper but chip rents stay high, then the shortage wasn't what held them up. If memory and rents fall together, it was. It's not a perfect thermometer, since there can still be too little power even when there's enough memory, but it's a useful public one.

Here is what I think happens. When the shortage eases, renters can choose again, and the chip that has just been overtaken loses out fastest. Today the B200, the newest chip the new bets will track, rents for about twice as much as the older H100. The next Nvidia generation is shipping now. Once it's everywhere, I expect the B200 to slide toward the H100's price, well before the five or six years the owners' books assume.

If the shortage eases and the B200 keeps its premium over the H100, I'm wrong, and the watch will show it.

Why I'm keeping this watch

I track these numbers anyway, as part of my research on how the AI buildout is financed. They're too important to sit in my notes, so I publish them here, once a month, for anyone who needs them. It's easy to throw out predictions like the one above. Every platform and opinion site is full of them. But a lot is at stake here, so it's worth checking whether this one holds.

The numbers that would settle it already exist, scattered across index pages, company filings and loan documents. But I haven't found anyone who writes them down side by side, every month, in plain words, where anyone can read them. The people who'd carry the loss, if it comes, are usually the last to see it, and it takes even longer for them to make all the connections. So I've started writing the numbers down, and making the connections, here. A record kept in public, so you can see it coming rather than read about it afterwards.

What it is

One short note a month, posted on the first working day. It will have the same handful of numbers every time: what each generation of chip rents for, how much more the newest one earns than older ones, what lenders charge, how long the owners say the chips last, and memory prices as a check on the shortage. Two index providers, because they don't agree, and the disagreement is not something trivial.

It's an experiment. I'll post six issues after this baseline, then decide whether it's worth keeping. If I skip a month, I'll say so.

Why not just look it up? You can. The prices are public, the trackers are good, and I use them. An AI will find them too. But they show today's price, and most of what's said about these numbers, including what an AI will tell you, still rests on how things looked a year or two ago.

The watch keeps the record: dated, measured the same way each month, checked against filings. But it does more than record. It puts the numbers side by side (rent, lending, lifespans and memory) so you can see whether they tell the same story. It says plainly what changed, what it might mean, and where the sources disagree. It checks the prediction above against the numbers every month, in public. Alongside, it lists the month's big launches and deals from the AI labs, the tech giants and the chipmakers that might bear on what the numbers show. Whether they explain anything is for you to judge. And over time it adds what isn't on any page: what people actually pay, and what people who know better think I've got wrong.

If you write about this, you get dated numbers you can cite. If you hold anything exposed to AI, you get a steady reading, so changes stand out. If you rent compute, see the next section.

The first note, No. 0: The Baseline, sets the starting numbers. You can also get each note by email through Substack.

How to contribute

Some of what this watch needs isn't on any page, and some of what you know could prove me wrong. Each way of contributing gets you something back. Nothing you send is ever sold or shared.

Tell me what you pay, and I'll tell you what others pay. The indexes show mostly hourly rent. What contracts actually cost, by length and by type of provider, isn't on any page. Send me the chip, a rough price range, how long the contract runs, and what kind of provider it is. "More, less or about the same as the indexes" is a useful answer. Your own numbers are never published or shared. Reports are treated as unverified: I check each one against the indexes, and leave out any I can't make sense of. Once at least five people have reported on a chip, I send everyone who contributed the combined ranges privately, a month before they appear in a note. They appear only as broad ranges, in a separate, clearly labelled section, never in the main table, and any group with fewer than five reports is merged or held back. Please don't send anything your contract doesn't allow you to share.

Tell me where I'm wrong, or what I'm missing. This watch says what I think. If you think chips last longer than I do, that the premium will hold, that I've misread a number, or that something moves the rent I haven't counted, write and say why. Each month, the strongest counter-argument or missing piece I receive goes into the note, with my answer. Corrections are fixed in public, and credited if you want credit.

You can also send a question. I'll answer one per issue, from filings and notices, with sources.

Send a note. You don't need to give your name or where you work. Whatever you send comes only to me and never goes into AI tools, and nothing is published with your name or employer unless you ask.

Research, not investment advice.

Sources: