The artificial intelligence sector is rapidly evolving from experimental software applications into a capital-heavy commodity market. Speaking on the Latent Space podcast on October 20, 2026, Evan Conrad, Chief Executive Officer of SF Compute, detailed the mechanics of a novel order book for compute power. According to Conrad, his firm has already processed hundreds of millions of dollars in transaction volume through the trading platform. The initiative aims to introduce standardized trading mechanisms for graphics processing capacity to establish transparent market pricing.
A central pillar of Conrad's market thesis revolves around the structural design of financial derivatives for compute power. Conrad stressed that liquid financial contracts must be built on physical settlement rather than cash settlement. Purely cash-settled contracts invite severe market manipulation and arbitrage risks because underlying server hardware and firmware setups remain highly heterogeneous. Ensuring that the counterparty actually delivers verifiable compute cycles to a designated data center is therefore critical to maintaining market stability.
The urgency for transparent trading mechanisms is highlighted by a stark economic arbitrage spread in today's hardware ecosystem. Fully utilized B300 clusters currently generate inference revenues ranging from 25 to 40 dollars per hour, whereas raw rental contracts for the same graphics processors trade at just 5 to 6 dollars per hour. This massive gap creates fertile ground for financial speculation while complicating long-term budget planning for model builders and infrastructure providers.
For institutional lenders, including commercial banks and private credit funds, a robust forward curve for compute has become a core necessity. Neocloud providers have taken on billions in private debt that is directly secured by physical GPU hardware collateral. Without reliable forward pricing, financiers cannot hedge residual hardware values or project future cash flows, leaving lenders exposed to sharp asset-backed credit risks.
These developments unfold amid mounting capital market scrutiny over soaring infrastructure spending across the tech industry. Infrastructure reports by SemiAnalysis project global AI data center power demand to reach 50 gigawatts by 2027, with hyperscalers and tech firms deploying over 730 billion dollars in capital expenditure during 2026. Standardized order books with physical delivery could determine whether these massive hardware-backed debt loads remain sustainable or face painful restructuring.