Why smarter AI models could drive up compute prices 10x
In a Nutshell
Frontier labs are closing the gap between 10× revenue growth and 3× compute growth by raising inference margins and pushing compute prices higher, with spot prices already up >40 % and deals like Google’s $900 M/mo GB200/300 contract showing 2× premiums over spot. As models become human-level engineers, the same H100-equivalent hardware will generate 15×+ more revenue, letting frontier labs outbid everyone else for scarce tokens and pricing out weaker users and applications. With Moore’s Law, fab capacity, and wafer re-allocation all maxed out, compute supply cannot scale fast enough, so the pre-singularity regime will see sustained, sharp price increases until robotics eventually makes chips abundant again.
These notes were generated by AI and may contain inaccuracies.
Anthropic's revenue has 10x'd year over year for three consecutive years and is likely to do so again. They ended last year with nine billion in revenue and are projected to reach one hundred billion to one hundred and fifty billion dollars this year. For this trend to continue, Anthropic would need to reach one trillion dollars in revenue by the end of next year. Whether this occurs depends on AI capabilities and whether models become sufficiently useful.
Lab compute only 3x's year over year. To maintain 10x revenue growth with only 3x compute growth, one of three things must happen: lab margins increase, compute prices increase, or the percentage of compute spent on inference versus training increases. All three are already occurring. Anthropic's inference margins went from forty percent in the middle of last year to upwards of eighty percent now. Spot prices for compute are more than forty percent higher than the February trough. OpenAI spent just a quarter of its compute on inference in 2024 according to Epoch, with that number likely closer to fifty percent now.
Labs prefer not to increase the share of compute spent on inference. Inference revenue is viewed as a means to secure investor funding for training the next model. Spending most compute on inference would signal that AI progress has stalled and that the lab is now a cloud provider, which is less compelling than building AGI. Labs believe they will build models within a year that make current ones look extremely poor and need to allocate the majority of compute to training and experiments for the next model.
This leaves two options: either lab margins increase so the lab captures the surplus, or compute prices increase so entities below the lab capture the surplus. It is unclear which scenario will occur. Margins reaching greater than ninety percent would require the leading model to be significantly ahead of competition, as margins exist because the offering is substantially better than alternatives. It is considered unlikely that intelligence margins would exceed ninety percent without being competed away.
Compute prices are already rising. The effect is stronger for the specific compute tranche frontier labs need, as they require scale for efficiency and flexibility, plus compute with security features for weights and customer information. Google is paying nine hundred million dollars a month for one hundred and ten thousand GPUs consisting of GB200s and GB300s from SpaceX. This price is 2x the spot price per hour, and that spot price itself is more than forty percent higher than February levels.
As AI models become smarter, they monetize the same amount of compute more effectively. A true human-level software engineer running on an H100 equivalent would justify renting that H100 for over 250K a year at current software engineer prices, representing over 15x the current spot price, without accounting for the AI working nights and weekends.
If ten million extra software engineers appeared, the marginal value of software engineers might decrease, reducing the revenue an H100 could generate. However, applying this argument to humans would constitute the classic lump of labor fallacy. Economists generally believe high-skill immigration does not decrease wages long-term due to innovation and specialization increasing labor value. If this labor supply shock is too large and rapid, the general heuristic may not apply, but standard economics suggests the marginal value of labor and compute should remain high.
As top labs better monetize compute and costs increase, it becomes harder for others to compete, as they must bid against entities that make better use of the resource. If a lab trains the best, most efficient model, it can charge much higher margins. This is the Alchian-Allen effect: if H100 rental costs twenty dollars an hour, using a weaker model becomes extremely inefficient because it burns more tokens on expensive compute. Labs will charge larger premiums for models that better economize this scarce input, effectively creating more compute by achieving the same result with less.
Many current popular AI applications will likely get priced out. AI is relatively cheap now because it cannot perform many tasks top humans can do. When this changes, frontier labs will pay more for tokens to automate AI research than individuals will pay to generate AI content.
This analysis may pattern match onto past incorrect predictions about scarcity, such as the Simon-Ehrlich bet where Paul Ehrlich predicted a basket of commodities would increase in price rather than decrease before 1990. Ehrlich's Malthusian worldview was shown to be wrong, as market signals and human ingenuity found ways to economize scarce inputs. The analogy may be incorrect, as other analysis shows Ehrlich might have won the bet in a different decade. Compute supply is much less elastic, less capable of absorbing large demand shocks, and less capable of being accommodated by substitutes than metal extraction.
The 3x compute capacity year over year is difficult to accelerate or sustain. Of the 3x, 1.4x comes from Moore's Law, which will be a miracle to maintain for a few more years. 1.2x comes from building new fabs, which will be bottlenecked through 2030 and beyond by ASML EUV machine production. 1.8x comes from AI absorbing wafer allocation previously used for smartphones and PCs, which will hit a wall by the end of next year when AI will have gone from sixty percent to eighty-six percent of leading-edge N3 nodes at TSMC. At some point, all leading-edge wafer capacity for AI will be absorbed.
Compute will eventually become cheap again when robots can convert silica sand and copper mines into new computer chips, making compute price equal to raw inputs and processing tools. This analysis concerns the current pre-singularity regime where AI compute only 3x's year over year, which is insufficient to offset AI's increasing value.
Anthropic's revenue 10x'ing while compute only 3x's illustrates strong economies of scale in the model business. When training a model, a one-time cost is incurred to learn skills that are then shared across all users. This differs from human labor, where each instance must be retrained from scratch. Strong economies of scale for intelligence raise concerns about power concentration.
Keep Dwarkesh Patel in your library
Save the videos and channels worth coming back to, and find them again in one place.
