Why smarter AI models could drive up compute prices 10x

Patel says Anthropic’s revenue grew about 10x annually for three years, reaching $9 billion last year, and estimates $100 billion to $150 billion this year if that pace continues. Meanwhile, lab compute grows about 3x a year. He identifies three ways to bridge the gap: higher margins, higher compute prices, or a larger share of compute going to inference.

Patel says labs would rather spend compute on training and experiments than become cloud providers focused on inference. That leaves higher margins or higher compute prices as the main ways to reconcile revenue and compute growth. He cites Google’s reported $900 million monthly payment for 110,000 GPUs rented from SpaceX, at twice the spot price, while spot prices had already risen more than 40% since February.

Patel argues that smarter models could earn more from the same hardware. As an illustration, he estimates that an H100 running a human-level software engineer could be worth over $250,000 a year at current software-engineer prices, more than 15 times its current spot price. He allows that a sudden supply of AI labor could reduce its marginal value, but says innovation and specialization may keep the value of labor, and compute, high.

If leading labs can monetize compute better, Patel says they may outbid smaller competitors and command higher margins. More efficient models could justify a premium because they use less of an expensive resource to produce the same result. He also suggests that consumer applications may be priced out if labs are willing to pay more for tokens used in AI research.

Patel compares his scarcity argument with the Simon-Ehrlich commodities bet, while arguing that compute supply is less elastic than metal supply. He says annual capacity growth depends on Moore’s Law, new fabs and wafer allocation, and sees limits in each, including ASML EUV machine production and AI taking a growing share of TSMC’s leading-edge capacity. He frames higher prices as a feature of the current pre-singularity period and says compute could become cheap again if robots can build chips from raw materials.