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

Compute is not a uniform commodity like Bitcoin: differences in memory, networking, cooling, software, and utilization make capacity hard to price consistently, while buyers navigate hundreds of providers and brokers and may have to verify that advertised supply exists. Brett Harrison says benchmarks should reflect actual purchasing patterns, including multi-year contracts for thousands of GPUs, rather than relying on listed prices or scraped data.

GPU deployment is constrained by whichever input is tightest—capital, hardware, or power—and gigawatt-scale grid interconnections can take multiple years, sometimes prompting projects to consider on-site natural-gas power. Amid limited powered capacity, Andy Andrawes says major AI companies’ threshold for taking calls fell from 100 megawatts to 30 megawatts, partly because of inference demand.

Building compute capacity has become a financing challenge: projects can use three-year offtake agreements for hardware not yet built to secure debt, but borrowers without credit histories may face exorbitant rates or no loans. Andrawes says top-tier loans backed by creditworthy customers such as Microsoft and Meta start around $1 billion to $3 billion, while AI buildout spending pushed hyperscaler free cash flow to zero or below and shifted capital-market attention toward underwriting compute as an asset.

Compute price risk remains difficult to manage because there is no standardized spot marketplace or established hedge, while prices have risen across chip and data-center-grade GPU types as demand outpaces supply that can arrive on time. Harrison argues that derivatives could help users hedge price exposure and forward costs, but proposed CFTC products are still under review for issues including index robustness, manipulation, price tracking, and settlement.

The growing number of GPU-compute providers and shrinking hyperscaler margins are presented as signs of increasing commoditization, even as Andrawes says some fast-growing inference providers have paid up to twice the usual price to secure capacity quickly. He also says compute prices have risen because demand is growing faster than timely supply, illustrating the tension between obtaining capacity and preserving margins.