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

Summary

Dwarkesh Patel and Dylan Patel dig into the economics of frontier AI infrastructure, arguing that OpenAI and Anthropic are rapidly becoming dominant buyers of new compute and may control most usable FLOPs within a few years. Dylan says AI infrastructure already drove much of U.S. GDP growth, that a large share of new compute is being built for the labs, and that their monetization per megawatt is now high enough to recycle inference revenue into more training and R&D. They discuss supply-chain constraints across chips, EUV tools, mirrors, power, and data centers, emphasizing that buildout is still whip-lagged and capital constrained. The conversation then shifts to macro spillovers: trillions of dollars of AI capex, debt issuance, higher interest rates, crowding out of other sectors, and possible pressure on equities and sovereign borrowers. The episode closes on a stark fork between either regulatory slowdown or increasing centralization of compute, capital, and ultimately AI-enabled power.

Notes

• Dwarkesh frames the episode around lab economics, compute markets, and where AI-driven capex pushes the world economy next.
• Dylan says U.S. GDP growth was heavily influenced by AI infrastructure last year.
• About one-third of new compute this year is for OpenAI and Anthropic, even if third parties physically build and rent it.
• He projects AI capex above $1T this year and above $2T by 2028.
• He says the labs are scaling from tens of billions in annual spend to hundreds of billions and then trillions by the end of the decade.
• Anthropic is described as having turned profitable in Q2, with OpenAI potentially following in Q3.
• The monetization argument is that frontier inference now covers the marginal compute cost and can fund further expansion.
• Anthropic revenue per megawatt has reportedly reached as high as $50M.
• Dwarkesh presses on when marginal compute stops going to the broader world and starts going mostly to labs.
• Dylan says labs already take around 30% of compute added this year.
• He expects Anthropic and OpenAI to absorb 40%–50% of next year's incremental compute.
• He says the centralization trend appears to be accelerating rather than slowing.
• By end-2028, he expects the labs to control most usable FLOPs if current trends continue.
• He sketches near-term compute growth of roughly 30 GW this year, 50 GW next year, and 70 GW the year after.
• Newer frontier hardware is said to be 3x–5x more performance per watt than prior chips.
• Dwarkesh uses a fab-cost model suggesting $6B of fab capex could generate a gigawatt of annual compute.
• Dylan agrees the capex-to-revenue ratio is huge, but says opex, data centers, and power still matter.
• He notes arbitrage opportunities in bottlenecked infrastructure like turbines and EUV tools.
• Carl Zeiss is reportedly preparing to make enough mirrors for roughly 100 EUV tools by end of decade.
• He still sees the supply chain as capital constrained, with 100 ASML tools for 2030 roughly the right scale.
• Dylan says the labs cannot yet fund the full buildout from cash flows alone.
• He argues that to buy around 70% of world compute in 2028, labs may need to pay $25M–$50M per megawatt.
• He says regulation and safety review slow down model releases and reduce revenue-per-megawatt growth.
• Dylan argues the real limiter on AGI is how much the rest of the world allows the labs to scale.
• He says labs will increasingly need their own chips, data centers, and infrastructure to keep scaling.
• Dwarkesh argues frontier compute grows fast while capability cost falls, implying roughly 10x annual growth in effective AI labor at the frontier.
• He extrapolates that OpenAI could go from 10 million to 100 million to 1 billion AI laborers over successive years if the trend persists.
• Dylan says the centralization of compute is the natural outcome if labs are the most effective users of compute.
• He argues that scarcity lets the leading lab charge higher markups and capture more deployment data.
• Dwarkesh says the AI economy is already concentrated among a few firms such as Nvidia, Anthropic, OpenAI, and hyperscalers.
• Dylan says Anthropic may currently lose some outside monetization opportunities, but internal use can still justify keeping compute in-house.
• He gives a total AI capex model of about $11T from 2024 to 2029.
• Even after cash flow funding, he says more than $5T of credit issuance would still be needed.
• He says 2028 alone could require $3T–$4T of capex across compute, data centers, energy, and supply chain.
• He warns that the commonly cited $40B–$50B AI capex figure misses the data center and power plant layers.
• He treats crowding out as the key macro constraint, with AI capex competing against many debt-financed sectors.
• He illustrates how AI-linked borrowing could push corporate rates higher, using Meta as an example.
• He says higher rates would compress valuations, especially for long-duration cash-flow stocks.
• Dylan argues the world may need to accept higher rates and a more extreme growth regime if AI capex keeps scaling.
• He suggests AI firms may need to finance their own chips and infra as external capital becomes more expensive.
• The discussion ends with a fork between slowdown by regulation and extreme concentration of resources and power.