Berat Celik (@beratcelik0), Eren Berke Saglam (@erenberke0), and Kaan Yigit (@kaan_yigit0) launched Stoa Markets on July 31, opening a request-for-quotation marketplace where institutions can buy and sell GPUs and AI servers through vetted counterparties.
Berat Celik on X The founders are attacking a basic mismatch in the AI infrastructure market. GPUs are bought in transactions large enough to affect balance sheets, yet the secondary market still runs through private broker networks, calls, emails, spreadsheets, and group chats. Buyers may struggle to find available hardware, while sellers can have little visibility into what comparable inventory is clearing for elsewhere.
"A $100M GPU deal doesn't close on an exchange," Celik wrote in the July 31 launch post. "It closes in a group chat, off a voice note."
Celik said Stoa generated more than $300 million in RFQs during its first month. That figure is self-reported and measures requests, rather than completed trades, revenue, or hardware delivered. Stoa has not published corresponding transaction volume.
The claim still captures the founders' bet: aggregating purchase requests may be the fastest route to aggregating the market itself. Dealers gain access to verified demand and bid without seeing competing offers. Buyers get firm quotes without calling vendors one by one. Stoa records the RFQ, hardware grade, condition, price, delivery terms, inspection process, and settlement milestones.
Three founders shaped by engineering and markets
Celik completed a Master of Engineering in computer science at Cornell University, including Cornell Tech, in 2026. His public writing around Stoa focuses on the financial risk accumulating alongside GPU ownership. In an August 2 post, Celik argued that holders pay close attention to upside when GPU prices rise while spending less time planning for a 30% decline.
That downside case explains why Stoa is interested in lenders, funds, and data center operators alongside AI labs and hardware dealers. A lender financing a GPU cluster needs a defensible collateral value. An operator rotating older accelerators needs evidence of resale demand. A buyer deciding between ownership and cloud rental needs a current acquisition price.
Saglam brings the finance perspective. He studied computer science and economics at Dartmouth College and has written that trading rates at Citi taught him the importance of transparent prices, trusted counterparties, and risk controls. Saglam described the founders' thesis in a July 20 post: capital supporting AI infrastructure became institutional before the hardware market did.
Yigit previously worked as a quantitative developer on a commodities desk at Uniper. Oil, gas, and power positions had visible prices, standardized contracts, and curves that traders could use to manage exposure, he wrote about Stoa. GPU hardware lacks comparable infrastructure.
The founders tested the problem manually before building software. Yigit said he priced the same server node through eight dealers and received quotes that were 15% apart. He attributed the spread to market access rather than underlying costs. That experiment led Stoa to use blind bidding, forcing dealers to compete for an order while shielding their quotes from one another.
Stoa is part of Y Combinator's Summer 2026 batch. The founders have not attached a funding announcement to the product launch.
From a quote PDF to a binding trade
A buyer begins on Stoa by describing the required hardware in plain language or up an existing quote PDF. Stoa uses AI to draft the RFQ, but the buyer reviews the fields before submission. Verified dealers then return firm quotes, and accepting one creates a binding transaction under Stoa's terms.
The workflow covers A100, H100, H200, B200, B300, and GB200-class hardware, including complete multi-GPU servers. Buyers can specify quantity, region, timing, condition, financing, and delivery requirements. Stoa then records confirmation, payment, shipment, delivery, inspection, and settlement events.
The prices displayed on Stoa's homepage should not be read as current market quotes. Stoa labels them illustrative and says they are not live market data. The examples include $25,860 for an H100 SXM5, $33,850 for an H200 SXM, and $43,860 for a B200. A separate eight-node H100 example shows a claimed $156,000 saving against typical private-channel pricing, but Stoa does not identify that comparison as a completed transaction.
Stoa also limits its role in the underlying hardware sale. Its terms say Stoa acts as an intermediary and does not ordinarily own, ship, inspect, or take title to the equipment. The buyer and seller retain responsibility for authenticity, condition, delivery, payment, and performance. Stoa can collect know-your-business information and facilitate payment through Stripe, while aggregate liability under the terms is capped at $100.
Those boundaries place much of Stoa's execution risk on counterparties. They also make the quality of Stoa's verification, records, and dispute workflow central to whether institutions treat the venue as safer than their existing broker relationships.
The transaction data may become the larger product
Stoa's marketplace sits in a growing field. Compute Exchange spans physical hardware, reserved cloud capacity, and forward delivery contracts, while GPU rental markets such as Vast.ai primarily sell access to compute rather than ownership of the chips.
Stoa is concentrating first on physical assets and the information produced when buyers and dealers negotiate. Its terms give Stoa the right to aggregate and commercialize deidentified pricing, bids, offers, transaction volumes, benchmarks, indices, and market patterns generated through the service.
That data clause reveals the longer play. Transaction fees can support the marketplace, while completed RFQs can produce pricing benchmarks for lenders, operators, insurers, and investors. A useful GPU price curve requires enough verified observations across models, regions, conditions, quantities, and delivery dates. Each successful trade can make the next valuation easier.
Stoa has work to do before that data product is established. Its H100 composite index page currently says the index value is unavailable, despite describing a real-time, volume-weighted benchmark. Stoa's price-discovery layer is therefore an intended capability rather than a functioning public reference price.
The founders' first task is narrower and operational: persuade buyers to submit real demand, persuade dealers to compete with firm quotes, and keep participants inside Stoa through delivery and settlement. If they can do that repeatedly, the founders will own something the fragmented GPU market currently lacks - a record of where hardware actually clears.