Compute Is A Trillion-Dollar Market Trading In Group Chats | Brett Harrison & Andrawes Bahou

GPU rentals and purchases are among the AI economy’s largest costs, with large compute contracts ranging from $100 million to $10 billion, according to Brett Harrison. The market has expanded from a handful of GPU-as-a-service providers to about 500 tracked by Compute Desk, but buying remains opaque and fragmented: customers negotiate through providers, brokers, prior suppliers, and trusted contacts over calls, email, Slack, and WhatsApp. Andrawes Bahou said there may be hundreds of brokers as well as hundreds of providers, and buyers must verify that offered capacity exists. In some cases, customers sign multiyear offtake agreements for clusters that have not been built, which providers may then use to seek hardware financing.

Compute is becoming more commodity-like as supply expands, but it is not fully fungible. Memory, networking, cooling, software, and utilization affect the value of a service, leaving no single observable market price. Prices also differ substantially between hyperscalers and neoclouds. Compute Desk publishes reference prices for H100, H200, B200, and B300 compute using negotiated transaction data from neocloud invoicing systems; listed website and on-demand rates can be poor guides because large GPU contracts are usually privately negotiated. Harrison argued that hedging requires indexes tied to real contract and spot-market exposure, not prices that fail to represent the economics of large, multiyear deals.

Architect has proposals before the CFTC for compute derivatives, and the commission is seeking public comment on index reliability, manipulation, price tracking, and cash versus physical settlement. Harrison said the goal is to transfer risk, not simply create a venue for speculation, and identified CME as a competitor. A compute operator with a fixed-term customer contract, for example, may face a lower renewal price when its capacity comes back to market. In the episode’s example, compute sold at $5 per GPU-hour could be worth $2 at renewal; futures or options might help secure a $4 or $3.50 floor. Expiring futures can match a known contract end date, while perpetual products cannot lock in cash flow for a specific expiry.

The discussion also described how options could help buyers manage uncertain demand. One AI company expected to need either four or eight H200 nodes for a year but wanted 60 days to learn more. The proposed arrangement combined a one-year commitment for four nodes with a put expiring in 60 days and a call allowing it to take eight nodes then; each node contained eight GPUs. A counterparty would accept the risk of unused or additional capacity in exchange for a premium. More generally, buyers face a timing trade-off: moving quickly can cost up to twice as much, but for a fast-growing inference provider, lost revenue from insufficient compute may outweigh worse unit economics. Sellers may seek three- or five-year commitments even when buyers cannot confidently forecast demand that far ahead.

Compute finance spans distinct borrower and deal-size tiers, from large institutional loans of roughly $1 billion to $3 billion for major operators to middle-market GPU data-center loans of about $100 million to $1 billion and asset-leasing services for smaller borrowers. Lenders assess creditworthiness alongside uncertain future GPU values, capacity requirements, and algorithms; weaker borrowers may face very high rates or no financing. Subprime-style compute loans could help smaller firms launch, but speakers also warned of the risks. Long-term offtake agreements give lenders confidence that a project will have customers, while hedging compute exposure could reduce dependence on such agreements and help projects advance faster. Harrison likened offtake to a car’s handbrake: hedging could let lenders take more risk.

Capacity depends on hardware, capital, power, and location. Memory-heavy configurations and InfiniBand can take longer to procure than memory-light setups and generic networking, while gigawatt-scale projects may wait years for grid connections; some operators are building behind-the-meter natural-gas plants. The episode estimates physical compute trading at $1 trillion, versus roughly $2.8 trillion to $3 trillion for physical crude-oil trading. Since crude derivatives are described as about 40 times physical trading, the speakers see substantial potential for compute derivatives, though no forward compute curve yet exists. They expect inference demand and more efficient algorithms, including use of older hardware, to sustain deployment despite constraints. The transcript also suggests that financial exposure could let people oppose a data center locally while still sharing in potential gains from compute futures.