Interview, Fireside Chat
Dylan Patel – Two labs will soon control most of the world's workforce
- Global AI capital expenditure (CapEx) is forecast to rise from just over $1 trillion this year to more than $2 trillion in 2028, with total supply chain CapEx potentially exceeding $2 trillion annually and approaching $10 trillion by the end of the decade, necessitating over $5 trillion in new credit issuance.
- Revenue generation per megawatt for top labs like Anthropic and OpenAI has surged to as high as $50 million per megawatt, with forecasts predicting levels of $70–$80 million by the end of 2027, rising to $100 million per megawatt or more in the absence of regulatory constraints, potentially reaching hundreds of billions in annual revenue per lab.
- Compute demand is projected to triple annually for frontier labs, reaching 18 units by the end of 2027 and 54 by the end of 2028, with incremental global additions expected at 30 gigawatts this year, 50 gigawatts next year, and 70 gigawatts in 2028, peaking at an upper bound of 80 gigawatts.
- Anthropic and OpenAI are expected to consume 40% to 50% of total compute next year, eventually controlling up to half of all incremental compute by the end of next year, with a prediction that a single entity or pair of labs could own most of the world's compute within 1.5 to 2 years.
- Financial viability is expected to shift as labs transition from spending tens of billions to hundreds of billions annually, with Anthropic already profitable in Q2 and OpenAI potentially becoming profitable in Q3, though capital injections will remain necessary as cash flows are insufficient to fund the required expansion.
- Supply chain constraints are anticipated to drive significant price increases, with the cost of compute rising to $25–$50 million per megawatt if labs acquire 70% of world compute in 2028, alongside rapid price hikes for memory and slow increases for fabrication capacity.
- Manufacturing capacity constraints involve a requirement for specific wafer counts per gigawatt (55,000 N3, 6,000 N5, and 170,000 DRAM wafers), with fab CapEx of $6 billion producing one gigawatt annually, while EUV tool demand may require 100 units by decade's end to support growth.
- China is predicted to scale manufacturing rapidly, potentially adding 50 gigawatts in 2029, though current 2028 estimates suggest only around 30 gigawatts from domestic fabs like SMIC and CXMT.
- Interest rates are projected to rise significantly, potentially reaching tens of percent in the 2030s, creating a risk of a "second Volcker shock" where 40 countries default if rates exceed growth rates or if AI revenue cannot service the massive debt load.
- Economic concentration is forecast to accelerate as capitalism drives centralization, with the AI sector potentially consuming a third to a quarter of the U.S. economy and all AI labor increasing 10x year-over-year, potentially surpassing the global human population by the end of the decade.
- Regulatory interventions, such as model release delays or data center bans, could stall revenue per megawatt or reduce compute deployment, though they may also inadvertently slow Western labs more than open-source Chinese alternatives.
- Revenue models are shifting as labs allocate less compute to inference and more to training to achieve AGI, with predictions that internal revenue generation could reach $300 million to $500 million per negawatt or megawatt for firms like Jane Street and Anthropic.
- Economic divergence is expected where only AI-involved countries and stocks retain value, as non-AI equities could plummet to near zero, while AI stocks may trade at significantly higher multiples in a doubling economy.
- A scenario of rapid self-improving intelligence (RSI) is described where progress could occur at 100x to 1,000x annual rates, compressed into mere months if government release lag is introduced, though regulatory barriers remain the primary limiter on deployment speed.