Why smarter AI models could drive up compute prices 10x

Summary

Dwarkesh Patel makes the case that frontier AI is entering a period of compute scarcity, not abundance: if labs like Anthropic can keep compounding revenue at around 10x while compute only grows about 3x, the gap must be resolved through higher margins, higher compute prices, or more inference spend. He argues all three are already happening, citing reported jumps in Anthropic inference margins, rising spot GPU prices, and a larger share of OpenAI’s compute going to inference. Patel also says frontier labs can’t rely on generic spot capacity because they need scale, flexibility, and secure infrastructure for weights and customer data. Using examples like Google’s expensive GPU deal and a hypothetical human-level software engineer on an H100, he concludes that smarter models can monetize the same hardware far more effectively. The broader implication is that compute may stay expensive for years because chip supply is constrained by fabs, EUV tooling, and limited wafer allocation, even though prices could eventually fall again in a post-robotic manufacturing world.

Notes

• The core claim is that 10x revenue growth cannot persist with only 3x compute growth unless margins, prices, or inference share change.
• Anthropic’s inference margins reportedly moved from about 40% to above 80% in roughly a year.
• Spot GPU pricing is already materially above the February trough, suggesting a tighter market.
• OpenAI’s inference share of compute rose sharply in 2024 and may now be around half or more.
• Labs want inference revenue to fund the next training run, not to become cloud providers.
• A dominant-margin strategy requires the leading model to be far ahead of substitutes.
• Frontier labs need specialized capacity, not just arbitrary spot instances.
• Google is cited as paying $900 million per month for 110,000 rented GPUs.
• That Google deal is described as roughly 2x spot, with spot itself already elevated.
• A human-level software engineer on H100-equivalent hardware could rationally justify extremely high rent.
• The comparison is explicitly framed against standard labor economics and the lump-of-labor fallacy.
• Patel thinks the AI demand shock may be too fast and too large for the standard labor analogy to hold cleanly.
• Rising compute prices favor the top labs that can monetize the resource best.
• The Alchian-Allen effect is used to explain why expensive inputs can increase margins for the most efficient models.
• Using less compute for the same task effectively creates more compute value.
• Some consumer-facing AI uses may get priced out as frontier customers outbid them for tokens.
• At the margin, research automation may outbid low-value consumer token demand.
• Compute supply is presented as less elastic and less substitutable than commodity supply.
• Patel decomposes annual compute growth into Moore’s Law, new fabs, and reallocated wafer capacity.
• He argues the fab-expansion path is constrained through 2030 or later by ASML EUV machine production.
• AI’s share of TSMC leading-edge capacity may be approaching a ceiling.
• Patel says compute only becomes cheap again when robots can manufacture chips from raw materials.
• Anthropic’s revenue growth versus compute growth is treated as evidence of strong model-business economies of scale.
• Training is a one-time cost, while the learned capabilities are reused across all users.