Why smarter AI models could drive up compute prices 10x
The author analyzes the looming imbalance between rapid AI lab revenue growth and slower compute capacity expansion. He explores how this dynamic will likely drive up compute costs and favor highly efficient, frontier models, creating significant barriers to entry and reshaping the AI landscape in the coming years.
The analysis focuses on the current trajectory of compute resource allocation and pricing for leading AI labs, projecting a significant shift in the coming years due to a growing disparity between revenue growth and compute capacity expansion.
The Widening Gap Between AI Lab Revenue and Compute Growth
Leading AI labs like Anthropic have experienced phenomenal revenue growth, with Anthropic’s revenue reportedly 10x’ing year-over-year for the past three years. This trend, if it continues, would necessitate a $1 trillion revenue for Anthropic by next year, a scenario tied directly to the exponential increase in AI capabilities. In contrast, the overall compute capacity available to labs has only increased by approximately 3x year-over-year.
To bridge this substantial gap between 10x revenue growth and 3x compute growth, one or a combination of three factors must occur:
- Increase in Lab Margins: Labs would need to generate more revenue per unit of compute.
- Increase in Compute Price: The cost of compute resources would need to rise.
- Shift to Inference Compute: A higher percentage of compute must be allocated to inference rather than training.
Evidence suggests all three trends are already in motion:
- Anthropic’s inference margins reportedly surged from 40% to over 80% for Fable models.
- Spot prices for compute have increased by more than 40% since the February trough.
- OpenAI’s compute allocation for inference rose from 25% in 2024 to potentially 50% or higher currently, according to Epoch data.
Labs’ Strategic Priorities and Economic Implications
AI labs generally prefer to minimize the shift of compute towards inference, as they view inference revenue primarily as a means to secure funding for further, more ambitious model training. A heavy focus on inference suggests a stagnation in AI progress, effectively reducing a lab to a “cloud provider” rather than an AGI builder. Labs believe they are on the cusp of developing models that will vastly outperform current ones, necessitating substantial compute investment in training and experimentation.
This leaves two primary mechanisms to reconcile the revenue-compute disparity:
- Dominance of Lab Margins: If this path prevails, leading models would need to achieve margins exceeding 90%. The author expresses skepticism that such high margins for intelligence services could persist without being eroded by competition, given the nature of market economies.
- Dominance of Compute Price Increases: This appears to be the more likely and already observed trend. Frontier labs require specialized, high-scale, and secure compute, for which they pay a premium far exceeding standard spot prices. For instance, Google is reportedly paying $900 million per month to SpaceX for 110,000 GB200s and GB300s, a price 2x the spot rate for those specific GPUs.
The Monetary Value of Advanced AI and Compute
A critical conclusion is that as AI models become smarter, they will inherently be able to monetize the same amount of compute more effectively. The author posits a thought experiment: if an H100 could run a “true human-level software engineer,” its annual rental value would exceed $250,000, which is more than 15 times the current spot price for an H100, not accounting for 24/7 operation.
Addressing concerns about a potential decrease in the marginal value of compute due to increased AI supply, the author invokes the “lump of labor fallacy.” Standard economic theory suggests that high-skill labor supply shocks, like high-skill immigration, do not necessarily depress wages long-term due to innovation and specialization. While acknowledging the potential scale and speed of an AI-driven labor shock, the argument suggests that the marginal value of both labor and compute could remain “astonishingly high.”
Consequences of Escalating Compute Costs
This paradigm shift will lead to several significant changes:
- Increased Barrier to Entry: Well-resourced, top-tier labs will have a decisive advantage, as they can outbid and more effectively utilize increasingly expensive compute resources. This makes it harder for new entrants or smaller players to compete.
- The Alchian-Allen Effect: Higher compute costs will strongly favor the most efficient and powerful models. It becomes economically “stupid” to use weaker, less efficient models that consume more tokens and compute for the same outcome. Labs capable of training models that economize scarce compute will command a substantial premium. Basically, if you have a model that can get the same result by using less compute, then you’ve, in some sense, created more compute, and the value of compute is gonna increase.
- Pricing Out Current AI Applications: Many existing, “cheap” AI applications will likely become unviable. As AI capabilities advance, leading labs will prioritize compute for high-value tasks, such as automating AI research itself, willing to pay far more for tokens than for less critical applications.
Inelastic Compute Supply and Limits to Growth
The author cautions against drawing parallels to historical examples of commodity scarcity being overcome by market forces (like the Simon-Ehrlich bet), arguing that the supply of advanced compute is far less elastic. It is less capable of absorbing large demand shocks or being easily substituted.
The 3x year-over-year compute capacity growth rate faces severe constraints from its constituent elements:
- Moore’s Law (1.4x contribution): Maintaining this rate for even a few more years is seen as a “miracle,” far from accelerating.
- New Fabs (1.2x contribution): The construction of new fabrication plants is bottlenecked by the limited production of ASML EUV machines, a constraint expected to last until at least 2030.
- Wafer Allocation (1.8x contribution): AI has already absorbed a massive share of leading-edge wafer capacity from sectors like smartphones and PCs. For instance, AI’s share of TSMC’s N3 nodes has grown from 60% to 86%. This reallocation has a ceiling, meaning this contribution to compute growth will “hit a wall” soon. Given these bottlenecks, sustaining even the current 3x year-over-year compute scaling is presented as a formidable challenge.
Long-Term Perspective and Economies of Scale
While acknowledging that compute will eventually become cheap again in a “post-singularity regime”—where robots could automate the production of chips from raw materials—the discussion primarily concerns the current “pre-singularity regime.”
The significant disparity between Anthropic’s 10x revenue growth and 3x compute growth strongly illustrates the powerful economies of scale inherent in the model business. A model’s one-time training cost allows its learned skills to be shared across all users, a stark contrast to human labor, where each instance requires individual training. This phenomenon, while economically efficient, raises concerns about power concentration.