AIBRIEF
Nvidia's flagship AI accelerator trades as a private quote between brokers, AI labs, and smaller AI focused clouds, and a Y Combinator S26 launch called Stoa Markets wants to build a public, real time record of AI chip trades.
The "price" of a flagship AI chip is whatever a broker can quote today. A smaller AI-focused cloud looking to offload a batch of H100s to an AI lab, or a broker clearing aged A100 stock, has no public tape to lean on. One adjacent data set on CCIR shows Nvidia's H100 SXM trading anywhere from $13,300 (90-day sold) to $25,750 (current ask) for the same stock-keeping unit (vendor-specific model number). A100 80GB shows the same shape, from $5,400 sold to $12,447 ask. Compute.exchange brackets 2026 H100 80GB prices at $15K–$40K depending on whether the unit is used, refurbished, or new.
No clearing house, no standard grade, and no settlement record sits behind those ranges. Counterparty risk is resolved deal by deal: a quoted seller may or may not actually ship.
Stoa Markets, a Y Combinator S26 launch by founders Eren, Berat, and Kaan, is trying to plug that gap. The marketplace funnels buyers through verified dealers and runs each trade from request-for-quote to settlement in recorded stages: confirmed, payment, shipped, delivered, inspected, settled. Target users include brokers, data center operators, AI labs, neoclouds, original-equipment-manufacturer channels, value-added resellers, systems integrators, liquidators, and lenders.
The pitch is a public tape for AI compute. The open question is whether large OEM allocations and hyperscalers route secondary supply through a young venue, and whether a marketplace can clear deals before it has dealers. Liquidity is a chicken-and-egg problem: the venue needs volume to attract dealers, and dealers need volume to join.