TechnicalOpinion

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

Dwarkesh Podcast1h 16m

Dylan Patel analyzes the exponential growth of AI compute infrastructure, projecting that OpenAI and Anthropic will control most of the world's usable computing power by 2028-2029. He discusses how this creates massive economic centralization, potential sovereign debt crises, and the near-inevitability of AI power concentration despite regulatory headwinds.

Summary

Dylan Patel and the host examine the trajectory of AI infrastructure investment and compute allocation. Patel notes that roughly one-third of incremental compute added in 2024 is going to OpenAI and Anthropic, with this percentage expected to rise to 40-50% by 2025. While total global compute is projected to grow from 30 gigawatts in 2024 to 70+ gigawatts by 2028, the labs' share grows exponentially—from ~2-5 exaflops at year start to potentially 50+ by end of 2028. This is driven by dramatically improving revenue-per-megawatt economics: labs generating $10-15 million per megawatt in costs can now generate $50-100+ million per megawatt in revenue, enabling them to fund exponentially more training compute from inference profits. The conversation explores how this compute centralization will require roughly $5 trillion in debt financing alongside $6 trillion in cash flows through 2029, with total CapEx (including data centers and power infrastructure) reaching $10+ trillion annually by late decade. This creates a profound economic distortion where interest rates rise due to competition for capital, equity valuations compress due to higher discount rates, and the opportunity cost of capital outside AI becomes prohibitively expensive. Patel argues that regulatory measures—preventing model releases, pausing training, banning data centers—are among the few constraints that could slow AI progress, but even these face pressure from market economics. The host raises concerns about developing nations defaulting due to rising interest rates (echoing the Volcker shock of the 1980s) and discusses how in an RSI scenario, interest rates could reach tens or even hundreds of percent, making non-AI investments essentially worthless. Both acknowledge that centralization appears nearly inevitable given economies of scale in AI training, deployment, and data, with value capture concentrating among 1-2 labs despite current value leakage to end users and intermediaries like Jane Street. The transcript ends with discussion of how, within a decade, a single lab could have computational capacity equivalent to billions of AI workers, exceeding Earth's human population, fundamentally concentrating economic and decision-making power.

About this episode

<p>Had a lot of fun chatting again with my twin brother Dylan Patel.</p><p>We went through lab economics over the next few years - the shift from inference to training as RSI draws near; and how Anthropic and OpenAI are on track to control most of the world’s usable FLOPs within the next few years (because they can monetize compute better and thus outbid everyone).</p><p>And then we discuss whether the >$10T of total AI capex we’ll see by the end of the decade will cause a sovereign debt crisis, where hyperscaler debt raises interest rates, drives non-AI exposed countries into bankruptcy, and crashes non-AI equities.</p><p>One question we weren’t able to resolve is whether there’s anything that can counter all the forces barrelling towards centralization in this industry - the economies of scale in training, the scarcity of compute, and eventually continual learning and RSI.</p><p>Watch on <a href="https://youtu.be/aV26V1UvkJw" target="_blank">YouTube</a>; read the <a href="https://www.dwarkesh.com/p/dylan-patel-3" target="_blank">transcript</a>.</p><p>Sponsors</p><p>* <a href="https://x.ai/bot" target="_blank">Grok Bot</a> has been quite helpful with my search for a new editor. I created a recruiter bot and described the type of editor I was looking for. That bot then spun up a handful of subagents that combed through my emails and X DMs, read the end credits of various documentaries I like, and figured out who edits for some of my favorite YouTubers. It took all of those results, and then delivered me a shortlist of candidates that matched my criteria. Try Grok Bot for yourself at<a href="https://x.ai/bot" target="_blank"> </a><a href="http://x.ai/bot" target="_blank">x.ai/bot</a></p><p>* <a href="https://antithesis.com/dwarkesh" target="_blank">Antithesis</a> lets you add time travel to your software testing toolkit. Since the Antithesis platform is fully deterministic, everything that happens inside of it is perfectly reproducible. So if your software crashes, you can rewind to the exact right moment, freeze time, and investigate. Or you can test different hypotheses by perturbing the system: kill a node or disable a feature, see what happens, then reset the trajectory and try something else. Learn more at<a href="https://antithesis.com/dwarkesh" target="_blank"> </a><a href="http://antithesis.com/dwarkesh" target="_blank">antithesis.com/dwarkesh</a></p><p>* <a href="https://janestreet.com/dwarkesh" target="_blank">Jane Street</a> is hiring for two separate ML internships right now, one focused primarily on research and one focused on engineering. In both cases, interns are expected to contribute to real work, not contrived exercises: one common project is adapting a frontier LLM paper to financial markets, which tend to come with a ton of different gnarly challenges. Importantly, you don’t need any finance background to apply. 2027 applications are open now at<a href="https://janestreet.com/dwarkesh" target="_blank"> janestreet.com/dwarkesh</a></p><p>Timestamps</p><p>(00:00:00) – Two labs will soon control most of the world’s compute</p><p>(00:07:01) – $6 billion in fab capex enables $1t+ of end revenue</p><p>(00:13:08) – Compute prices will rise if the labs outbid everyone</p><p>(00:18:22) – Which layer will capture most of the surplus?</p><p>(00:25:40) – What could slow down progress?</p><p>(00:29:43) – Labs are shifting compute from inference to R&amp;D</p><p>(00:33:27) – China gets less than 10% of new compute, but its labs need less</p><p>(00:48:48) – Will AI cause a sovereign debt crisis?</p><p>(01:07:52) – Will the world’s future workforce belong to a few companies?</p> <br /><br />Get full access to Dwarkesh Podcast at <a href="https://www.dwarkesh.com/subscribe?utm_medium=podcast&#38;utm_campaign=CTA_4">www.dwarkesh.com/subscribe</a>

Key Insights

  • By end of 2025, OpenAI and Anthropic are projected to consume 40-50% of all incremental new compute being deployed globally, up from 30% in 2024, representing a dramatic centralization trend that appears to be accelerating rather than slowing.
  • Anthropic's revenue per megawatt has reached $50 million while maintaining $10-15 million in compute costs, creating a 5x margin that enables reinvestment of all inference profits into training compute, fundamentally changing lab economics from cash-burn to self-funding.
  • Labs generating 60-70+ million dollars per megawatt by end of 2027 will rationally allocate increasing percentages of compute toward training over inference, since internal AI research ROI exceeds external token-sale revenue, contradicting the common assumption that most future compute goes to inference.
  • The $11 trillion CapEx required through 2029 for AI infrastructure will be funded by approximately $6 trillion in cash flows and $5 trillion in debt, which will measurably increase market interest rates and reduce discount rates for non-AI equities across the economy.
  • Rising interest rates driven by AI capital competition will compress equity valuations globally, causing non-AI stocks to trade at much lower multiples and potentially making mortgage lending, consumer credit, and government debt service increasingly expensive.
  • Supply chain constraints—particularly ASML's EUV tool production (100 tools per year through 2030) and wafer fab capacity—are the primary physical bottleneck preventing labs from acquiring even more compute, but these constraints can be overcome with sufficient capital allocation.
  • China's AI compute capacity will remain below 30 gigawatts through 2028 due to export controls and limited domestic production, while US labs will have 50+ gigawatts, creating a structural advantage that export controls successfully engineer but at cost of global security.
  • By year-end 2028, labs' incremental compute is so advanced (GB300s, TPUv7s providing 3-5x efficiency gains) that controlling half of new incremental compute actually means controlling 60-70% of usable flops despite smaller watt counts, multiplying centralization effects.
  • A single frontier lab could have 50+ exaflops of effective compute by 2028, equivalent to billions of AI workers with individual labor productivity exceeding human levels, concentrating economic power beyond any historical precedent in magnitude and speed.
  • Regulatory restrictions on model releases and internal AI use (preventing deployment of best models, pausing training, restricting employee access) will reduce revenue-per-megawatt growth and slow but not prevent compute concentration, since regulations only marginally impact top labs' economics.
  • Value capture is currently distributed across the supply chain (semiconductors, infrastructure, models) but concentrating toward labs as their margins widen, with end users like Jane Street capturing more value than labs today but facing long-term squeeze as labs reallocate compute internally.
  • Interest rate equilibrium under AI growth will reach tens of percent annually if takeoff occurs and AI labor grows 10x year-over-year, making government debt service, consumer borrowing, and non-AI business financing economically infeasible and triggering global sovereign debt crises.

Topics

AI compute infrastructure growth and allocationEconomic centralization in AI labsRevenue-per-megawatt economics and scalingGlobal capital requirements and debt financingInterest rate effects and sovereign debt crisesRegulatory constraints on AI progressValue capture across the AI supply chainRecursive self-improvement (RSI) scenariosSemiconductor and energy supply chain bottlenecksGovernment policy and export controls

Transcript

Okay. I'm back with Dylan Patel, founder of Semi Analysis. Our version of a family Thanksgiving dinner. It is a regular yearly podcast, but we're not actually related. We'll tell the people this. We'll destroy the myth. Basically, where the world economy is headed is more and more becoming a function of where lab economics are headed, where the compute market is headed, etc. I want to understand where the crazy future ends up within a few years. But let's start with just where we are today. So walk me through lab compute and lab revenue right now and maybe projecting out a year or two. Yeah, so when we go back to last year, even at the end of…

Full transcript available for MurmurCast members

Sign Up to Access

More from Dwarkesh Podcast

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.