Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028

Overview

This episode explores the rapid centralization of global compute power by AI labs like Anthropic and OpenAI, their economic implications, and the potential for a sovereign debt crisis driven by the massive capital expenditures required for AI infrastructure. The discussion also touches on the societal and economic consequences of AI-driven centralization, the role of regulation, and the challenges of maintaining a decentralized future in the face of accelerating AI progress.

Notable Quotes

- By the end of next year, half of the world's incremental compute will already be going to Anthropic and OpenAI. – Dylan Patel, on the centralization of compute power.

- Why would I let Jane Street make all this money off of these tokens when I can allocate that compute internally and generate even more value? – Dylan Patel, on the labs prioritizing internal AI research over external deployment.

- We could make a trillion dollars right now, but we're just bottlenecked on the mirrors that go into the ASML machines. – Dwarkesh Patel, on the supply chain constraints limiting AI infrastructure growth.

🖥️ The Centralization of Compute Power

- Dylan Patel explains how AI labs like Anthropic and OpenAI are consuming an increasing share of global compute, with projections showing they could control the majority of the world's usable compute by 2028.

- The labs' ability to monetize compute more effectively than other industries allows them to outbid competitors, accelerating centralization.

- SpaceX and Meta are emerging as significant players by building compute infrastructure and leasing it to labs, further consolidating power.

📈 The Economics of AI Labs

- AI labs are transitioning from venture-funded losses to profitability, with margins skyrocketing due to advancements in compute efficiency and revenue per megawatt.

- Dylan Patel highlights how Anthropic's revenue per megawatt has reached $50 million, with potential to exceed $70 million by 2027.

- Labs are reallocating more compute from inference (serving external users) to training and internal R&D, prioritizing long-term AGI development over short-term profits.

💸 Sovereign Debt Crisis and AI CapEx

- The podcast explores how the $10+ trillion in AI capital expenditures by 2030 could crowd out other sectors, driving up interest rates and potentially triggering a sovereign debt crisis.

- Developing countries with high debt and low tax revenues, like Pakistan and Nigeria, are particularly vulnerable.

- Rising interest rates could also devalue non-AI equities, as higher discount rates reduce the present value of long-term cash flows.

🌍 The Global Compute Divide

- The U.S. dominates AI compute growth, with 70% of new compute deployed domestically, while China's share has fallen below 10% due to export controls and supply chain constraints.

- By 2028, China is projected to have only 30 gigawatts of AI compute, much of it using less advanced chips, compared to the U.S. labs' far superior infrastructure.

- This disparity could delay China's ability to compete in AI-driven economic and technological advancements.

🏢 Centralization vs. Decentralization in an AI Future

- The conversation delves into the inevitability of centralization due to economies of scale in AI training, continual learning, and recursive self-improvement (RSI).

- Dwarkesh Patel raises concerns about a future where a single lab could have more effective AI laborers than the entire human population.

- The hosts discuss whether regulation or alternative frameworks could prevent extreme centralization, but acknowledge the difficulty of countering these economic forces.