Stoa put more than $300 million of quote requests through its system in its first month. The number worth staring at is not the volume. It is that nobody outside the trade can tell you what a GPU is worth this week, and every dollar of that $300 million was haggled over in the dark.

Three people are trying to change that, and none of them builds hardware.

What they built

Stoa is a marketplace where buyers ask for quotes on AI hardware. A buyer posts what they need, checked dealers come back with firm prices within 48 hours, and the buyer picks one. Its launch announcement was about as plain as these things get: "A Marketplace for GPUs and AI Servers".

What makes it a business rather than a noticeboard is everything wrapped around the quote. Stoa runs the identity checks on both sides, writes the contract, arranges the shipping and handles the money. Buying several million dollars of chips from a dealer you found in a group chat brings a set of problems that have nothing to do with hardware: who is this company, who pays first, who owns the boxes while they are on a plane, and who is liable when the delivery arrives short. That admin work is why a trade closes on the platform instead of slipping off it.

The company is in San Francisco and trades at stoaexchange.com.

The point: the marketplace is the means, the price is the product

GPUs became a major asset in about four years. They sit on company balance sheets in the billions, they are bought with borrowed money, they lose value on schedules nobody has much history for, and they change hands constantly. What they do not have is a price.

Not a price in the sense of a number somebody will tell you if you ask. A price in the sense that oil has Brent, gas has Henry Hub and copper has the London Metal Exchange: a public figure, updated constantly, produced by completed trades, that everyone in the industry quotes against. AI hardware has nothing like it. It changes hands through brokers, resellers, private introductions and word of mouth, and each deal is priced by how much the buyer happens to know, which is usually not much.

The effects pile up. A buyer has nothing to compare against, so cannot tell a good deal from a bad one. A seller cannot value what is in the warehouse, so cannot tell whether holding on is costing money. Most importantly, a lender cannot value the security. Lending against hardware depends on a valuation: somebody independent has to be able to say what it would fetch if it had to be sold on a bad day. Without that number, the asset either cannot be borrowed against at all, or gets borrowed against at a rate that assumes the worst. That is a tax on the whole industry.

Stoa's stated aim is to create transparent pricing through settled trades and establish reference prices for an asset class that currently has none. That is not a feature of the marketplace. It is the reason for the marketplace.

A reference price is the quietest and most lasting position in any commodity market. The agencies and exchanges that publish them did not get there by being clever about pricing. They got there by being the place where enough trades completed that their number became the number. Once an industry starts writing contracts that say "the Stoa price on the settlement date", the position is close to unassailable. Lenders quote it. Insurers quote it. Auditors quote it. And when financial contracts on compute eventually arrive, they settle against it.

Taking a fee on trades is a business. Owning the price of a whole asset class is a position, and a far bigger one.

Which brings us to the founders, because the line-up says plainly what this company thinks it is. Berat Celik, the CEO, is a repeat founder who was previously at xAI and studied computer science and economics at UBC before a Cornell M.Eng. Eren Berke Saglam, the CFO, traded interest rate derivatives at Citi and was salutatorian at Dartmouth in computer science and economics. Kaan Yigit, the CTO, was a quantitative developer at Uniper building trading systems for energy commodities, with computer science from UIUC.

That is two traders and a quant developer. It is not the team you put together to move boxes. The energy background is especially telling: in power and gas, the published benchmark is the foundation everything else is built on, and everyone in the business understands that whoever hosts the completed trades holds the valuable end of the deal. People who have worked inside a market with reference prices know exactly what it is worth to be the one producing them.

No inventory, and the only thing both sides actually want

How it works deserves attention on its own, because it is the cleverest part of the company and it goes far beyond GPUs.

Stoa steps between buyers and sellers holding absolutely nothing. It does not own a single card. It does not store, finance, refurbish or ship anything of its own. It carries no stock and takes no position, so it carries none of the risk that has killed most businesses built around physical hardware, which is that the thing in the warehouse is worth less this quarter than last.

What it has instead is the one asset neither side can obtain for itself.

A buyer wants a seller. A seller wants a buyer. Neither one can create the other, and in an informal market neither can reliably find the other. Owning that introduction takes no money at all.

It is tempting to call this arbitrage, and it is worth being exact, because the truth is better than that. An arbitrageur needs money, takes a position, and carries the risk on the gap. Stoa does none of those things. By standing in the middle of enough deals it ends up holding the thing that makes arbitrage possible for everybody else, which is knowing where the price actually is, without ever putting a dollar at risk. That is a stronger position than trading the gap, and a much safer one.

What you need to repeat it is specific: something valuable, buyers and sellers who genuinely cannot find each other, deals currently done by introduction and private haggling, and no published price. Where all four are true, someone can step in with nothing but a phone and a good name, and within a year know more about that market's prices than anyone trading in it.

The list of markets that still tick all four boxes is longer than it should be in 2026. Used industrial and construction equipment. Aircraft parts. Data centre space and power contracts. Second-hand chip manufacturing equipment. Shipping space on quieter routes. Scrap and recycled metal. Carbon credits. All valuable, all traded through relationships, none with a price you can look up.

Who else is in this field

Selling access to GPUs has been one of the most crowded businesses of the last three years. CoreWeave went from crypto mining to a public company on the back of it. Lambda, Crusoe, Nebius, Fluidstack and Voltage Park all sell clusters and cloud. Vast.ai and RunPod built spot markets for rented capacity. Modal and Replicate abstracted the machine away entirely. Together AI bundles compute with inference, and SF Compute went furthest toward treating it as a traded commodity with a clearing market.

Almost all of them rent you time on a machine. Stoa is doing something different: broking the sale of the hardware itself, which is the deal that has never had a public price attached to it.

Why it is bigger than it sounds

The headline from Demo Day was $300 million in requests for quote in the first month. The asset being built underneath it is worth more than the number.

  • A price benchmark built from settled trades compounds. It is the kind of thing lenders and auditors eventually cite, and once it is the reference nobody needs a second one.
  • Zero inventory removes the risk that killed the businesses it resembles. Nothing is bought, nothing is held, and nothing depreciates while it waits for a buyer.
  • What it holds is buyers and sellers who each want the other. That is the highest technique of arbitrage: no stock, no position, and both sides arriving because the other side is already there.

What to watch

One number settles this: the gap between quotes requested and deals completed. A quote is only an intention. Only a completed deal produces a real price, and only real prices can become reference prices. If $300 million of requests turns into a thin sliver of closed business, Stoa has built a very busy noticeboard, and the data underneath will never be thick enough to publish anything from.

The second thing to watch is who quotes the number first. The day a lender, an insurer or an auditor writes a Stoa price into a document, the company has stopped being a marketplace and become part of the market's plumbing, and the two are worth nothing like the same.