The Centralization of AI Compute and Lab Economics

The Economics of Compute Centralization

Dylan Patel argues that we are witnessing a massive shift in the global economy where compute availability is becoming the primary driver of growth. Currently, OpenAI and Anthropic are transitioning from venture-funded entities to profitable enterprises, with their revenue per megawatt significantly outpacing the cost of compute. This margin advantage allows them to outbid other market participants for hardware, creating a feedback loop where they capture an increasing share of the world's incremental compute—projected to reach 50% by the end of next year and potentially 70-80% by 2028.

The Capital Expenditure Whiplash

While the total AI-related capital expenditure is expected to exceed $2 trillion by the end of the decade, the supply chain for the necessary hardware—specifically EUV lithography tools and wafer fabrication equipment—is physically constrained. Patel notes that while capitalism will eventually expand production, the signal-to-action lag is significant. Even with massive capital injections, the supply chain for critical components like mirrors for ASML machines is unlikely to scale rapidly enough to meet the exponential demand of the labs in the immediate term.

The Sovereign Debt and Market Risk

There is a looming tension regarding whether this massive concentration of capital will trigger a broader economic crisis. If hyperscalers and labs continue to absorb the vast majority of available capital, it could drive up interest rates and potentially lead to bankruptcy for non-AI-exposed countries or sectors. Patel suggests that the current trajectory of AI investment is so aggressive that it threatens to crowd out other essential economic activities, creating a fragile dependency on the continued success and monetization of frontier models.

Regulatory and Safety Constraints

Patel highlights a paradox: the labs' own safety advocacy and internal regulatory caution are currently slowing their release cycles. By withholding their most capable models, they risk stalling their revenue-per-megawatt growth, which in turn limits their ability to outbid competitors for compute. If safety-related delays persist or intensify, it could prevent the labs from achieving the projected 100-gigawatt capacity by 2028, potentially allowing open-source or international competitors to remain relevant in the market.

Value Capture and the User Surplus

Despite the labs' dominance, much of the value generated by frontier models is currently captured by the users rather than the labs themselves. High-value industries, such as quantitative finance firms, are extracting significantly more economic value from tokens than the labs charge for them. This suggests that while the labs control the infrastructure, the downstream economic impact of AI is currently favoring the adopters, though this balance may shift as models approach AGI-level automation.