Between July and August 2026, tech writer Dwarkesh Patel published a series of posts on his personal blog examining how more capable AI models could affect compute prices. His core argument: as AI models get smarter, demand for compute could drive prices up more than tenfold. The claim offers a new angle on the current debate over AI inference costs.

The core calculation: an H100 should rent for $250,000 a year

In his July 29 post, "Why compute might get 10x+ more expensive in coming years," Patel laid out a specific calculation: if an AI system could match a human software engineer and run on an H100-equivalent chip, that H100 should rent for over $250,000 a year at current software engineer market rates.

If a human-level software engineer that could run on an H100 equivalent, at current market rates for software engineers, that H100 should rent for over $250k a year. That's 15x today's spot price.

He went on to note that this is 15x the current H100 spot price. The calculation benchmarks AI capability directly against human labor costs, suggesting that current compute pricing may sit far below its potential economic value.

Follow-up post: compute prices could rise 10x or more

On August 3, Patel published "Why smarter AI models could drive up compute prices 10x," continuing the same theme and using "The end of cheap compute?" as its summary line. The headline explicitly links "smarter AI models" to "compute prices rising 10x," reinforcing his earlier argument about shifting supply-demand and pricing logic for compute.

Neither post offers more detail in its summary, but the underlying logic is consistent: the more capable the AI model, the more valuable the human labor it can replace, and the more compute — as a "tool of production" — should be repriced accordingly.

Industry impact: AI inference costs in focus

Patel's calculation ties compute prices directly to the human labor AI replaces, offering a new framework for understanding AI inference costs. If his logic holds, the cost structure of AI applications built on cheap compute could be in for a rewrite. For an AI industry that depends on large-scale inference, a shift in how compute is priced would directly affect business models and cost expectations.

That said, the calculation rests on the hypothetical scenario of a "human-level software engineer." Actual compute prices will still be shaped by supply and demand, chip generations, model efficiency and other factors. Patel's posts do not offer a specific timeline or further market conditions.

What to watch

  • Whether compute spot prices and rental rates confirm Patel's prediction;
  • The real impact of improving AI models on inference demand and cost structure;
  • How the industry responds to a shift in compute pricing logic.