In a Substack post, Dwarkesh Patel argues that advances in artificial intelligence (AI) are likely to push up the price of compute. His basic point is that as models become smarter they will extract more economic value from the same hardware, which will bid up rental and purchase prices for accelerators.
Patel offers a concrete illustration: if a true human-level software engineer could run on hardware equivalent to an NVIDIA H100, then, at current market rates for software engineers, that H100 should rent for more than $250,000 a year. "That’s 15x today’s spot prices," he writes. He uses this example to show that AI looks relatively cheap today partly because it still cannot perform many tasks that top human experts can.
He also expects this increase in compute price to be temporary in the long run. Patel suggests that large-scale roboticization of the compute supply chain could eventually lower costs, bringing prices closer to the raw inputs and tooling required to produce hardware. However, he notes, that point would likely arrive well into what he describes as the singularity.
Why this matters
The central implication of Patel’s post is that economic dynamics near the singularity could be unusual: components currently treated as commodities—such as computers and accelerators—might be aggressively bid up because advanced AI systems demand vast amounts of compute and can monetize its use more effectively. Patel’s argument highlights that technological progress does not guarantee uniformly lower costs; in some cases, scarce, high-value inputs like high-performance compute could become significantly more expensive before automation reduces those costs again.
