Product1 publisher3 min readPublished
Compute Exchange's new catalogue shows GPU specs and prices before it shows the seller
The startup's inventory manager pulls current, reserved and forward capacity from more than 100 GPU providers into one comparison view. By its own account it still fills only about 40% of the requests buyers bring it.
The Product Desk · Product desk
What happened
- Compute Exchange, which matches GPU buyers and sellers, has added an inventory manager meant to give enterprise buyers one view of available and upcoming AI compute across more than 100 providers.
- Coverage spans Nvidia and AMD parts including the H100, H200, B200, A100 and MI series, plus complete servers and commitments for inference tokens, with data arriving by API or spreadsheet upload.
- Providers pay only on a completed transaction, the inventory feature is free, and Li said the company has passed seven figures of revenue this year and is profitable.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision With roughly 60% of requests unfilled, a buyer cannot yet retire the vendor relationships this is meant to replace, so the choice is whether the catalogue becomes a sourcing channel or a price reference alongside the account manager calls.
- constraint Because the seller stays anonymous until shortlisting, operational track record is missing from the comparison step, and the platform's own verification carries the weight a known brand would otherwise carry.
- capability A team that over-committed on a reserved contract has a listed channel to resell it, which lowers the cost of getting a long commitment wrong.
- precedent Automated matching and spot trading are the stated destination, and that is the point at which GPU capacity would be priced like a commodity rather than merely tabulated like one.
The buyer this is built for is holding a requisition for sixty-odd H200s, a training run with a date on it, and quotes from five neoclouds that are not written in the same units. Carmen Li's description of the alternative, that you do not want to call 400 salespeople and compare specs and negotiate, is an accurate account of that afternoon [4].
What teams tell themselves in that situation is that they have a comparison problem. The company's own figure points upstream of comparison: Li told SiliconANGLE that limited inventory has kept Compute Exchange from fulfilling more than about 40% of RFQ volume [12], which leaves roughly 60% of requests landing on a marketplace that cannot find matching capacity [1]. A consolidated view does not create nodes.
The part that earns the login is the forward book. Buyers can browse capacity that is reserved or expected to come online later, not only what is racked today, and compare it on price, specification, location, SLA and commercial terms [2]. For those forward listings Compute Exchange says it reviews purchase orders and colocation arrangements; for installed systems it can check chip identifiers, configurations, network health, floating-point performance and thermal degradation [7]. That is more than a vendor spec sheet, and it is doing a job the buyer cannot do from outside.
It is doing that job because the seller is anonymous. Buyers see benchmarks, location, price and service commitments first, and the provider's name only after they shortlist [9]. So the thing that actually determines whether the training run survives, whether this operator answers at 3am in month four, is unavailable at the moment of comparison. Compute Exchange's verification stands in for reputation until the shortlist.
The onboarding numbers are worth doing by hand. More than 100 vendors over 18 months averages about 5.6 a month, near 0.18 a day, so the stated current pace of nearly one provider a day is roughly five times that average [6][2]. Either the roster is accelerating sharply or most of it arrived recently.
Li is also the one arguing against treating this as pure price shopping: she says price is one indicator, and that variation by chip type, geography and contract structure limits how much simple comparison is worth [14][15]. The commodity version of this product, with automated matching and spot trading, is the stated ambition rather than the shipping feature [16], which is a slightly awkward position for a company with "Exchange" in its name.
Two questions sort whether this belongs in your process. First, substitutability: can the job take a different region, a different accelerator, a different term length? Second, verifiability: do you have any way of your own to check a seller you have never heard of? Substitutable and unverifiable is where the platform pays for itself, and providers pay the fee only when a deal closes, with the inventory view carrying no charge [10]. Pinned requirement plus existing vendor relationships makes it a price reference, not a sourcing channel. Either way the diligence after the name is revealed stays with you, because a broker is paid for the transaction, not for your choice being right.
The useful question on the sales call is what share of requests shaped like yours got filled last quarter. A count of 100-plus providers answers a different one.
What to watch
- Whether the roughly 40% fill rate moves as the supply increase Li expects next year arrives, or demand absorbs it first.
- Whether automated matching and spot trading actually ship, since continuous pricing is what would make the commodity framing accurate.
- Whether providers keep listing once blinded side-by-side comparison starts pressing on their quoted prices.