The GPU market has a pricing problem that money alone cannot fix. Stoa Markets, a Y Combinator S26 startup founded by Eren, Berat and Kaan, launched on Hacker News this week with a blunt observation: the most valuable hardware on earth is still traded through phone calls, forwarded spreadsheets and long email threads. The founders, who previously traded interest rate derivatives and built pricing systems for oil and gas, are building a marketplace for new and used GPUs and AI servers, complete with know-your-business checks, firm quotes and a settlement layer. The pitch is not just convenience. It is the missing price-discovery layer for an asset class that has become the collateral backing the entire data-center buildout.
The numbers in the launch post are striking. Stoa says it collected more than $300M in requests for quotes (RFQs) during its first month. Its site lists illustrative prices for current-generation hardware: an A100 80GB at $9,490, an H100 SXM5 at $25,860, an H200 SXM at $33,850, a B200 at $43,860 and a GB200 NVL at $67,400. The company also shows a sample quote for eight H100 SXM5 nodes, estimating a total of $243,400 per node on Stoa versus $262,900 per node off-market, a claimed saving of $156k on a $1.95M order. These are illustrative levels, not live market data, but they make the central thesis concrete: the same hardware clears at meaningfully different prices depending on who is asking and who is answering.
That price dispersion is the real story. The founders describe a week where one seller quoted $200k for a server node and another quoted $240k for what looked like the same thing. Neither quote was necessarily wrong. Each seller saw only their own corner of the market. Before Stoa could compare the quotes, it had to normalize configuration, condition, warranty, location and delivery terms. The founders draw a direct analogy to Kelley Blue Book, the reference that attaches a used-car price to year, trim, mileage and condition. “An H100 server” carries about as much pricing information as “a used BMW.”
The deeper implication is financial. GPUs are the collateral in the data-center buildout, and today financing terms depend almost entirely on the offtaker, the company that has committed to use the compute. If that offtaker is a hyperscaler, the financing can look investment grade. If it is a smaller cloud or a startup, terms get expensive fast, even when the hardware backing the loan is identical. The lender’s problem is reasonable: if the borrower defaults and the servers need to be sold, what can they actually fetch? Without a functioning secondary market, the answer is guesswork, list prices and one-off appraisals. Stoa’s bet is that accumulated trade data becomes the resale evidence lenders need to underwrite GPU-backed loans at better rates for smaller players.
This is where the launch stops being a niche trading tool and becomes an infrastructure play for the AI economy. The compute market has grown so fast that its financial plumbing never caught up. Cloud providers, neoclouds, AI labs and enterprises are all buying and selling the same SKUs, but each transaction is a private negotiation. Stoa wants to turn that into a structured market where dealers return firm quotes against a standardized request, without seeing each other’s bids, and where payment is held until delivery confirms. The company does not take possession of the hardware. It runs the RFQ, the binding acceptance, and the settlement timeline: confirmed, payment, shipped, delivered, inspected, settled, with evidence required at each step.
The trust layer is worth scrutiny. The founders acknowledge that GPU trading runs on relationships and that inventory is not shown to just anyone. Dealers need to trust the people bringing them clients, and clients need to trust that quotes will actually turn into trades. Stoa built those relationships by brokering deals itself before launching the platform. The KYB checks verify the company, who owns it and who is allowed to trade for it. This is a deliberate rejection of the pure-software approach. The founders say it plainly: “We knew from the beginning that this couldn’t be a software only marketplace.” The relationships, and the history of who actually follows through, are a big part of the process.
That framing is honest about a real tension. A marketplace only works if it has liquidity, and liquidity in this market flows through brokers who guard their inventory and their client lists. Stoa’s answer is to bring those brokers onto the platform with verified demand, rather than trying to disintermediate them. The tiered fee structure, lower at higher volumes, gives volume traders a reason to consolidate their flow. The question is whether the platform can reach the scale where its clearing levels become the reference price, or whether it stays a better brokering tool in a market that remains fundamentally relationship-driven.
The timing is good. The AI hardware cycle is entering a phase where used equipment matters. The B200 and GB200 NVL systems are shipping, which pushes H100 and A100 inventory into secondary channels. Data center operators are planning fleet rotations. Lenders and lessors are sitting on GPU-backed exposure and need liquidation paths. Stoa’s site explicitly targets brokers and dealers, data center operators, AI labs and enterprises, cloud and neoclouds, OEMs and VARs, resellers and liquidators, lenders and lessors, and funds and capital partners. That is every side of the market, which is either ambitious or overreaching, depending on execution.
What would make this work is the same thing that makes any market work: enough trades to generate trustworthy price signals. Stoa’s $300M in first-month RFQs is demand, not transactions. The founders are candid that the goal is to build resale evidence over time. “As trades build up, they also leave lenders with actual resale evidence instead of list prices and one off appraisals.” That sentence is the whole thesis. The marketplace is the means; the price-discovery layer is the product.
For AI builders, the implication is direct. Compute financing is a bottleneck for anyone who is not a hyperscaler. A startup that wants to stand up a cluster faces financing terms that assume the hardware is worth less the moment it is installed, because there is no reliable way to price it on exit. If Stoa, or a competitor, builds a credible secondary market, that risk premium shrinks. Better resale data means better financing terms for smaller clouds and AI labs. It means the collateral actually functions as collateral.
There are obvious risks. The hardware market is concentrated in a handful of vendors, and NVIDIA controls the supply of the most sought-after SKUs. A marketplace can only price what trades, and if most high-end inventory moves through OEM allocation and direct cloud contracts, the secondary market may stay thin at the top end. The founders’ oil-and-gas background is relevant here: commodities markets work because there is a standardized product and a clearing mechanism. GPUs are less standardized, and the configuration variance is exactly what Stoa is trying to normalize.
The launch is a signal about where the AI economy is heading. The buildout phase is maturing into a trading phase. Hardware that was once a scarce input is becoming an asset class with a resale curve, and the people who price that curve will shape who can afford to train models. Stoa is early, and its $300M RFQ figure is a first-month number, not a track record. But the direction is clear: the next big unlock in AI compute may not be a faster chip. It may be a better way to know what the chips are worth.