← Blog

The Compute Watch - No. 0: The Baseline

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

New here? Read this first: what this watch is, why it exists, and the prediction it checks.

AI companies rarely own every chip they use. Many rent them by the hour from cloud companies and from specialist firms that do little else. A growing share of the money that bought these chips was borrowed, and the rent is what pays the loans back.

But chips are built to be replaced. A better one arrives every year or two, and the old one earns less. The owners' books, and many of the loans, assume the chips keep earning for years.

This watch tracks, once a month, whether they do. The core question is simple: does the newest chip keep earning much more than older ones once something better arrives? It's a good moment to start counting. Nvidia's next generation is shipping now, and the exchanges are getting ready to let people bet on chip rent.

This first note sets the baselines. Every later note compares against them.

The ladder

What it costs to rent one AI chip for an hour, from the oldest generation to the newest. Rents rise and fall with the shortage. The ratio between generations shows how fast a chip loses ground once something better arrives, and that's what the owners' lifespans and the lenders' assumptions depend on.

Chip	Rent per hour
A100 (2020)	$1.58
H100	$2.77
H200	$3.37
B200	$5.86
B300	$6.78
H100 at the big cloud companies	$7.20

Silicon Data indexes, specialist providers unless marked, as shown on 2 October 2026.

The same H100 costs more than twice as much at Amazon, Microsoft or Google as at the specialist firms that do little but rent out AI chips. Most of the numbers below use the specialists' prices, because that's where the hourly market is.

The numbers

What	Baseline (2 Oct 2026)	What it tells you
The premium: B200 rent ÷ H100 rent	About 2.1× ($5.86 vs $2.77, Silicon Data)	How much more the newest chip the new bets will track earns than an older one. The core number. If it shrinks, the newest chips are losing their edge
Loan terms: chip-backed borrowing	Nscale's $1.4bn GPU loan at 5 points over the benchmark rate (S-1, 18 Sept)	Whether lenders are getting more cautious. Changes only when a new deal is disclosed
Server lifespans in the owners' accounts	Alphabet 6 years, Oracle 6, CoreWeave 6, Meta 5.5, Amazon 5–6 (part of its fleet shortened from 6 to 5 in 2025), Microsoft 2–6	How long the owners expect their chips to stay in use. Changes only at annual reports
The thermometer: memory (GDDR6 price)	$19.31 on Silicon Data's GDDR6 index	Whether high rents are just a shortage. If memory gets cheaper but rents stay high, something other than scarcity is holding them up

The two indexes disagree. Ornn, whose prices Intercontinental Exchange is using, puts the B200 at $8.09 and the H100 at $2.71 (30 September): a premium of about 3.0×, against Silicon Data's 2.1×. One likely reason is method: Ornn builds its index only from completed deals, while Silicon Data adjusts and averages across providers. But it means the answer to "what does the newest chip earn?" depends on who you ask. Ornn also moves far more: over the past month it had the H100 down about 18% and the B200 up about 18%. I'll track both.

Read together, today's numbers are consistent with a shortage. Rents are high and list prices are still going up. Lenders are asking for more protection, which doesn't contradict that. They are pricing what happens after the shortage, and so is this watch.

What happened in the weeks before this note

Things that might later show up in the numbers. None of them is claimed as a cause.

Prices

Rents were still going up. Nebius raised its hourly prices for selected Nvidia chips by 17–21% from 1 October, its second rise in three months. When the rise was reported, on 17 September, CoreWeave said it was still signing new contracts at higher prices, including short ones of three to six months.

Lending

Nvidia started looking for insurance. The Financial Times (29 September) reported that Nvidia is working with insurers and the broker Howden Re to spread the risk on loans backed by its chips.
Nvidia backed its customer's lease. If OpenAI stops paying for its data-centre space at a new campus in Ohio, which will hold Nvidia's chips, Nvidia covers what the landlord can't recover, up to $105 billion. The site is due to open from 2028, and OpenAI has agreed to pay Nvidia back.
Lenders pushed back. Reuters (1 October) reported that banks and investors value Nvidia's chips lower than Nvidia does, and want stronger guarantees before lending against them.
A chipmaker became a lender. Reuters reported, from Anthropic's IPO prospectus, that Broadcom has agreed to lend Anthropic up to $42 billion, through a note that can turn into shares, towards a five-year, $125.2 billion lease of chip capacity.

Demand

A lot of future compute is already spoken for. Anthropic's prospectus lists at least $518 billion of infrastructure commitments over about a decade, roughly 80% payable whether used or not.
A new frontier model from the custom-chip camp. Google announced Gemini 4 Argon on 30 September, with answers up to 1 million tokens long, up from 64,000, and an introductory price set to double later. Google says its Gemini models have run on its own chips for years. It's going to a small group of testers first.
One lab paused. OpenAI paused training and tool-using work on its most capable models, its second pause in three months. Its public services weren't affected.

Supply

The next rung arrived. Nvidia delivered its first Vera Rubin chips to AWS in late August. AMD expects its rival rack systems to reach customers in volume before the end of the year.

The market itself

Betting on chip rent got closer, but not yet. From 5 October, Intercontinental Exchange shows test prices for chip-rent futures. CME Group had aimed for 5 October but hasn't launched; the CFTC is asking the industry for comments until 20 October.

What I'm watching next month

Whether the premium holds. My prediction is that once the shortage eases and Rubin is widely available, the B200 slides toward the H100's price. On Silicon Data, in the week before this note, the premium narrowed slightly, but not because the B200 got cheaper: the B200 was flat while the H100 rose almost 5%. That's consistent with tight supply for older chips, but one week of an index can't tell you why.

How I'll judge it. I'll call the shortage easing when the memory price in the table falls at least 15% below this baseline and stays there for two notes in a row. Once it has, I'll count the prediction as right if the premium is below 1.6× on Silicon Data for two notes in a row, and wrong if it stays above 1.9× for two notes in a row. In between, it's too early to say. Until the shortage eases, a falling B200 price doesn't prove me right, and a steady one doesn't prove me wrong. If it hasn't eased by No. 6, the prediction stays open, and I'll say so.

Whether the price rises reach the indexes. Nebius raised its hourly prices from 1 October. If the October index moves up with them, the shortage is still biting.

When Rubin enters the indexes. That's a change in what the indexes contain, not a market move, and I'll note it as such.

The two indexes. Whether Ornn and Silicon Data move closer together or further apart.

Contribute

Send me what you pay for compute, a correction, or a question: send a note. Your own numbers are never published, nothing goes into AI tools, and you don't need to give your name. How contributions work.

You can get each note by email on the day it comes out.

How this is made. Headline figures come from Silicon Data's public index page, recorded on one fixed day each month (the first working day, from No. 1), using specialist-provider prices unless marked. Ornn's index is the cross-check. Silicon Data is the headline because it's steadier: Ornn, built only from completed deals, can move several percent in a day. Silicon Data is also the index CME plans to use. A screenshot of both pages is kept for every note. GDDR6 is graphics memory, not the HBM inside data-centre chips; it's used here as a public reading of the wider memory shortage. Index prices are averages across providers; individual list prices, like Nebius's, are often higher. If an index changes its method or its contents, such as a new chip or provider entering the mix, I note it, so that isn't mistaken for a market move. From No. 1, every table gets a "change since last month" column, and "no new data" where nothing was disclosed. From No. 1, each note also carries the strongest counter-argument or missing piece received, with my answer, and one reader question.

Research, not investment advice.

Sources: