A small San Francisco startup thinks it can impose order on one of tech's newest and messiest markets: the brisk but murky trade in high-end AI chips.
Stoa Markets launched its institutional GPU marketplace this summer and says it logged more than $300 million in requests for quotes during the first month. The three-person company, which came out of Y Combinator, built what amounts to a blind-bidding platform for artificial intelligence hardware. Buyers submit requests; vetted dealers respond with firm quotes. Accept one, and you've locked in a binding trade under terms Stoa sets.
It's an attempt to bring a measure of structure to what founder Berat Celik described in a launch post as an informal, high-friction process. "A $100M GPU deal doesn't close on an exchange," he wrote. "It closes in a group chat, off a voice note."
Whether that informality persists—or whether the market is ready for something resembling conventional trading infrastructure—remains an open question. Stoa has yet to disclose how many of those quote requests converted into completed trades, what its settlement success rate looks like, or how large its dealer network actually is. The $300 million figure measures interest, not execution.
Still, the ambition is clear. The platform handles Nvidia's H100, H200, B200, B300, and GB200 NVL chips, along with older A100 cards and additional models including the L40S, A6000, and V100, as well as a range of complete multi-GPU servers. Buyers can describe what they need in plain language or upload a quote PDF. AI drafts the request for quote. Dealers respond blind, meaning they can't see competing bids. The system tracks six settlement milestones from confirmation through final inspection.
Celik left xAI to build Stoa, according to the company's Y Combinator profile. He studied computer science and economics at the University of British Columbia, then completed a master's in computer science at Cornell. His co-founder and CFO, Eren Berke Saglam, traded interest-rate derivatives at Citi after graduating near the top of his class at Dartmouth with dual degrees in computer science and economics. The third co-founder, Kaan Yigit, built trading and pricing systems for oil and gas commodities at Uniper. On LinkedIn, Yigit framed the mission plainly: bring to GPUs the market infrastructure commodities have relied on for decades.
That infrastructure now includes a second product. In late summer, Stoa launched Stoa Intelligence, a pricing data service that converts trading activity into per-configuration benchmarks and index levels. The idea is to give lenders, investors, and companies with GPU-backed assets some reliable marks for valuation and collateral monitoring. Saglam announced the service on LinkedIn, saying the company would "turn pricing activity captured on Stoa into hardware marks."
At launch, though, Stoa's public H100 composite index page displayed an "index value is unavailable" message, according to an August report from RuntimeWire. The company reserves the right to aggregate and commercialize anonymized pricing data. It also runs a Data Partner Program: contribute your trade data and pay zero platform fees on eligible transactions. Individual submissions remain private, Stoa says.
The company charges fees only when trades execute and includes non-circumvention clauses in its terms. It makes a point of stating it does not act as a broker-dealer, exchange, money transmitter, or fiduciary. Those distinctions may matter as regulators begin circling a market growing faster than anyone anticipated.
Stoa is hardly alone. Compute Exchange operates an identity-verified marketplace with auctions and bid-ask trading. Nodestream runs a B2B platform for GPU servers and AI clusters. Digital Earth Exchange offers one-counterparty fulfillment contracts. RecycleVAR, a global IT hardware exchange, handles open RFQs for servers and networking gear including AI components. Clustermelt publishes GPU price ranges and quotes full server configurations on request.
Meanwhile, traditional futures exchanges are staking their own claims. ICE and a partner called Ornn announced plans in May to launch GPU compute futures. CME Group and Silicon Data followed in August with a plan to roll out two compute futures contracts in early October, pending regulatory approval. Nodal Exchange and Compute Desk announced similar contracts in September.
The rush reflects real confusion in the market. Enterprises are trying to decide whether to rent or buy while prices climb. ITPro reported in August that one former NSA technical director noted flagship GPU prices had nearly doubled over the past year. Tom's Hardware observed that consumer GPU prices kept surging, with next-generation cards not expected until sometime in late 2027 or 2028.
Stoa published research in June laying out the central problems: fragmented price discovery, GPU-backed lending without dependable benchmarks, opaque secondary-market mechanics. The company's Y Combinator launch description put it more bluntly. "A quote is an opinion. A settled trade is a receipt."
That framing assumes a market ready to graduate from group chats and voice notes to something resembling institutional trading rails. But the gap between quote requests and completed settlements—information Stoa hasn't yet disclosed—will tell the real story. For now, the startup is betting that enough buyers and sellers want a clearer view of what these chips actually cost, and that they're willing to trade on a platform that imposes structure in exchange for transparency.
Whether the hardware changes hands as smoothly as the futures contracts will remains to be seen.
