Dwarkesh Podcast

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

1h 17min

Dylan Patel discusses the rapid centralization of AI compute, predicting that by 2028 Anthropic and OpenAI will control the majority of the world's computing power. He explains how these labs are transitioning from venture-funded losses to profitability, with revenue per megawatt soaring from $10-15 million to $50 million or more. The conversation explores the implications for global economics, including potential sovereign debt crises and the concentration of AI labor.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

Anthropic and OpenAI are expected to control the majority of the world's compute by 2028, with their share of incremental compute reaching 40-50% as early as next year.
2

Revenue per megawatt for frontier labs has skyrocketed from negative gross margins to as high as $50 million, enabling them to reinvest profits into training.
3

The concentration of compute and AI labor could lead to a scenario where a single lab has more effective 'AI workers' than the entire human population by the end of the decade.
4

Rising interest rates driven by massive AI infrastructure spending could trigger a sovereign debt crisis, especially in developing countries.
5

Government regulation and safety concerns may slow external deployment of AI models, but internal use could accelerate progress even faster.

💬 Notable Quotes

"If you spend $10 on inference capacity, you generate $50 of revenue, and then you can turn around and spend all of that profit on training."
"By the time you're towards the end of 2028, if this trend continues, you've got them controlling most of the usable flops in the world on their own."
"The limiter on AGI is not how fast the research engineers can crank the gears—it's actually just how much the rest of the world lets that happen."