The four prices of an H100
The most-quoted number in AI infrastructure is also the least meaningful one. "What does an H100 cost?" has a single answer only if you believe the market is a single market - and it is not. Tonight the panel tracks 117 single-GPU H100 listings across 58 providers, every row with its source and fetch time attached. Read honestly, they do not describe a price. They describe four prices, serving four different buyers, and the gaps between them are where the actual market lives.
The price that markets: $1.68
The lowest number in any GPU conversation is a floor. Latitude.sh shows $1.68, Beam $1.83, GCore $1.88, GMI Cloud $2.00 - each carrying a quiet "(from)" that turns a rate card into an advertisement. A floor is the cheapest hour a provider has ever sold: a specific machine, in a specific region, at a specific moment, possibly never to be seen again. Only seven of the 117 rows are floors, but they punch far above their count, because floors are what screenshots, listicles, and sales decks are made of. We keep them in the panel - the receipts are real - but they are marked, and they stay out of every median. A market quoted in floors reads a third cheaper than the market you can actually buy.
The price that lists: $3.09
Strip the floors and 28 specialist GPU clouds publish on-demand rates that median at $3.09 an hour. This is the number analysts cite and budgets get built on, and it deserves the trust - mostly. Lambda lists, Civo lists, a dozen others list, and they will charge you exactly what they list. But a rate card is a claim about a price, not a promise of a machine. It tells you what the hour costs if capacity exists for you at the moment you ask. Between the list and the machine sits everything that makes capacity markets interesting.
The price that rents: $3.50
Then there is the price of a machine an inventory feed says is actually available this hour: median $3.50 across 45 rows - 13% above the rate card. Some of the wedge is mix: the rentable-now bucket holds different providers than the rate-card bucket. But some of it is the oldest premium in any capacity market. Immediacy costs money. A buyer paying $3.50 tonight is paying $3.09 for the silicon and forty cents for the certainty that it exists. And the wedge is not static: rate cards update quarterly, available prices update by the hour, so when demand spikes, the wedge is the first thing that moves. It is the elasticity premium rendered visible, and it only exists because someone bothers to check whether the machine is really there.
| Rate card | Rentable now | |
|---|---|---|
| Median | $3.09 | $3.50 |
| Rows on the panel | 28 | 45 |
| Refreshes | Every few weeks | Hourly |
| What it is | A claim about a price | A machine that exists |
| Premium for immediacy | - | +13% |
The rentable rows carry a second structure inside them. Single-GPU H100 PCIe machines rent at a $2.46 median; SXM modules at $3.84 - a 56% premium for the interconnect that multi-GPU training runs lean on. And at the very bottom of the rentable market sits Lium at $1.25, a marketplace price that undercuts every specialist rate card on the panel. The spread inside "available" is nearly as wide as the spread between tiers: which H100, from whom, in what form, matters as much as whether it exists.
The price that procures: $8.82
And then the number that makes newcomers assume the panel is broken: $8.82, the median across 37 H100 rows at Azure, AWS, GCP and OCI - 2.9 times the specialist rate card for silicon of the same class, topping out at $10.12 for an Azure H100 NVL in Australia East. This is not mispricing. It is a different product wearing the same SKU. The extra five and a half dollars buy availability zones a compliance team can sign off, SLAs a procurement department recognizes, egress economics that punish leaving, and the career safety of a logo nobody questions. The A100 shows the identical wedge - $4.78 hyperscaler against $1.49 specialist, a 3.2x multiple - so this is structural, not a Hopper quirk. There are two GPU markets: the one that buys machines, and the one that buys paperwork.
| GPU | Specialist rate card | Hyperscaler | Multiple |
|---|---|---|---|
| H100 | $3.09 | $8.82 | 2.9x |
| A100 | $1.49 | $4.78 | 3.2x |
Why one market became four
Segmentation this clean does not happen by accident; it persists because each tier is protected by a different moat. Floors survive because comparison sites do not check them. Rate cards hold because updating prices is work and demand signals arrive faster than pricing pages change. The rentable premium holds because inventory is genuinely scarce at the margin - a marketplace cannot list what its hosts have not switched on. And the hyperscaler premium holds because switching costs are the product: once your data, your compliance posture, and your procurement workflow live in one cloud, the GPU is the smallest part of the bill you are actually paying.
| The price that | Median | Rows | What it buys |
|---|---|---|---|
| Markets | $1.68 (from) | 7 | A screenshot - the cheapest hour anyone has ever listed |
| Lists | $3.09 | 28 | The published rate, if capacity exists |
| Rents | $3.50 | 45 | A machine that exists this hour |
| Procures | $8.82 | 37 | Compliance, SLAs, and the logo nobody questions |
How to read your own quote
The practical consequence: before comparing any two H100 prices, decide which of the four markets you are actually trading in. Building a slide for the board? The advertised floor will do, and it will be wrong by half. Setting a budget? Anchor on the rate card, then stress it with the rentable price. Need a machine tonight? The rentable price is the only one that exists - the rest are literature. Answering to compliance? The hyperscaler premium is not a ripoff; it is the price of the paperwork, and the panel makes it possible to see exactly how much paperwork costs.
The 8x spread from $1.25 to $10.12 is not the market failing to agree. It is the market segmenting, precisely, by what the buyer is really paying for - and it reprices every hour. We track all four prices continuously, with the fetch receipts attached, because the interesting question was never "what does an H100 cost." It is which H100 market you are in, and whether you know it.
Method note
All figures computed from the Compute Cafe panel snapshot of September 19, 2026: 117 single-GPU H100 listings across 58 providers, on-demand rates only, converted to USD at the day's rate. Floors (rows marked "(from)") are excluded from every median and shown only as the advertised tier. The four tiers hold different provider mixes, so the 13% rate-card-to-available wedge is structural evidence, not an exact constant; PCIe/SXM medians come from the rentable-now bucket (n=4 and n=18). A100 corroboration: 110 single-GPU rows, hyperscaler median $4.78 against specialist $1.49. Every row on the panel links the fetch that produced it.