Dylan Patel – Two labs will soon control most of the world's workforce
A detailed discussion on the rapid centralization of AI compute power within frontier labs like OpenAI and Anthropic, driven by their superior monetization of compute. The conversation explores the massive CapEx requirements for AI infrastructure, the potential for a sovereign debt crisis due to rising interest rates, and the impact of regulation on AI progress. It also delves into the strategic shift from inference to R&D within labs and the implications of exponential growth in
Dylan Patel discusses the rapid evolution of AI lab economics, focusing on the accelerating centralization of compute power and its profound global economic implications. He highlights that AI infrastructure alone constitutes a significant portion of GDP growth, with CapEx projected to reach $1 trillion this year and over $2 trillion by 2028, potentially $7-10 trillion by the end of the decade when accounting for power and data center infrastructure.
Centralization of AI Compute Power
OpenAI and Anthropic, initially venture-funded, have transitioned to profitability. Anthropic started turning a profit in Q2, and OpenAI is expected to follow. Their margins have skyrocketed, with Anthropic generating up to $50 million per megawatt, significantly exceeding the base cost of compute ($10-15 million per megawatt). This ability to monetize compute at higher rates allows them to outbid competitors for resources.
- Compute Allocation: Currently, about 30% of new marginal compute is allocated to these labs. This is projected to increase to 40-50% next year, meaning by the end of next year, half of the incremental compute in the world will go to Anthropic and OpenAI.
- Exponential Growth: The labs’ compute capacity is tripling annually, while global compute roughly doubles. If this trend continues, by the end of 2028, they could control most of the world’s usable FLOPs, even accounting for the increasing efficiency of new hardware (3-5x performance per watt).
- Supply Chain Response: While capitalism dictates that supply should expand to meet demand, the supply chain for advanced semiconductors (like ASML’s EUV tools) has a significant lag. A $6 billion fab CapEx can generate over $1 trillion in end AI revenue over five years, creating immense economic incentives. However, physical constraints and the time required to scale production mean this expansion is not immediate.
The Impact of Regulation and Market Dynamics
Regulation is a critical factor that could slow down AI progress and the labs’ ability to monetize their compute. Instances like OpenAI pausing training or Anthropic withholding advanced models due to safety concerns prevent them from fully leveraging their innovations.
- Revenue per Megawatt: If regulatory hurdles persist, the rapid growth in revenue per megawatt for labs could decelerate, reducing their capacity to outbid for compute. However, without such constraints, revenue per megawatt could exceed $100 million.
- Value Capture: Historically, value capture has shifted from hardware manufacturers to hyperscalers, and now increasingly to the model layer (OpenAI, Anthropic). End-users like Jane Street or Meta often derive far more value from using these models than the labs capture in profit, indicating significant unrealized value.
- Compute Pricing: The intense demand from labs drives up compute prices. Companies like SpaceX and Meta, with strong balance sheets, are building compute infrastructure without immediate customers, creating optionality to either use it internally or sell it to labs at premium rates ($25-50 million per megawatt).
Shift from Inference to Training and R&D
The labs are increasingly prioritizing R&D and training over inference, despite the immediate profitability of inference. The belief is that investing in AGI development offers a far greater long-term return.
- Internal Allocation: As models improve, the internal value generated from cutting-edge AI for research and development exceeds external inference revenue, leading labs to dedicate more compute to R&D. This is already evident in their monthly compute allocation trends.
- Training Dynamics: While labs possess multiple gigawatts of compute, training runs often utilize only a fraction (e.g., Mythos pre-training uses sub-200 megawatts). However, advancements in automated coding, research, and continual learning will likely lead to a higher percentage of compute dedicated directly to training.
China’s Position in the AI Compute Race
While the US (deploying 70% of new watts) has significantly outpaced China (sub-10% of incremental watts) since 2022 due to regulations and financial systems, China is expected to experience a