Dwarkesh PodcastDylan Patel – Two labs will soon control most of the world's workforce
At a glance
WHAT IT’S REALLY ABOUT
OpenAI and Anthropic may centralize global compute, reshaping economics and power
- Dylan Patel argues AI infrastructure spending is becoming a dominant driver of economic growth, with total AI-related CapEx moving from about a trillion dollars today toward multi-trillion-dollar annual levels later this decade.
- He predicts OpenAI and Anthropic will capture an accelerating share of incremental global compute—potentially around half by the end of next year—because they can monetize compute far better than other buyers.
- The episode explores how compute markets could reprice upward as scarcity bites, with balance-sheet-rich builders (e.g., Meta/SpaceX) gaining leverage by constructing compute ahead of customer commitments.
- They debate where AI’s economic surplus will ultimately accrue across the stack (end users, apps, model labs, chips/memory/fabs, datacenters/power), emphasizing that value capture has already shifted dramatically over time.
- They connect AI’s scale-up to macro-financial risks: trillions in debt funding, higher interest rates, potential sovereign debt crises in weaker countries, and political/regulatory backlash that could slow deployment or concentrate power further.
IDEAS WORTH REMEMBERING
5 ideasOpenAI and Anthropic are on a path to dominate incremental global compute—soon translating into dominance of effective FLOPs.
Patel claims labs are rapidly increasing their share of newly deployed AI datacenter capacity: ~30% of incremental compute this year, and potentially 40–50% next year, with OpenAI+Anthropic reaching ~50% of incremental compute by end of next year. Because new deployments are much more performant per watt, “controlling incremental compute” quickly becomes “controlling most usable FLOPs.”
Physical and supply-chain bottlenecks (not just money) throttle how fast compute can scale.
The conversation frames AI CapEx as already enormous (≈$1T+ this year) and potentially >$2T by 2028, with the broader buildout constrained by long-lead-time items (EUV tool supply chain, turbines, power plants, datacenter shells). Even if returns are high, the "bullwhip" delay means capacity can’t instantly expand to meet demand.
Lab unit economics have flipped: inference is becoming a major profit engine that funds more training.
They discuss a shift from negative margins (serving GPT-4-era models on Hopper at a loss) to strongly positive inference economics, with Patel citing cases like Anthropic reaching ~$50M revenue per MW and projecting $70–80M/MW blended by 2027 (subject to model leadership and ability to release). This creates a feedback loop: inference profits can be reinvested into training.
If labs outbid everyone, the price of compute likely re-rates sharply upward—benefiting “compute landlords.”
Patel argues compute prices must rise if labs try to buy 70–80% of global capacity, because third parties can also profitably serve models at today’s ~$10–15M/MW costs. He expects upward pressure toward ~$25–50M/MW in tight markets, especially when compute owners (e.g., SpaceX, Meta) have balance-sheet freedom to build first and sell later.
AI surplus is still in flux; user value capture dominates today, but pricing power is migrating toward scarce inputs and frontier labs.
A key surplus question is who captures AI value: end-users (e.g., trading firms, big tech using models internally) have often captured more value than model providers, while value capture has shifted over time among chips, memory, fabs, hyperscalers, and model labs. Patel expects rebalancing via pricing power to propagate through the supply chain, but slowly.
WORDS WORTH SAVING
5 quotesAnd so by the time you're in, like towards the end of 2028, if this trend continues, which I see nothing that's stopping it, um, you, you've got them just controlling most of the usable, you know, FLOPs in the world on their own.
— Dylan Patel
If anyone had like $400 million and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one and, and wait, wait, wait, and then sell it for north of a billion dollars, right?
— Dylan Patel
Yeah, so, so this, this, this crowding out effect is actually, like, the thing that I've, like... is the reason it's not, like, YOLO one billion gigawatts.
— Dylan Patel
Because if you believe in, you know, sort of AI researchers, RSI, AGI- Then all of this exists. All of this is the base.
— Dylan Patel
Um, yeah. I mean, I guess, I guess, like, what world do you see, Dwarkesh, where everything is not centralized? Um, because it seems to me that every force is screeching towards centralization.
— Dylan Patel
High quality AI-generated summary created from speaker-labeled transcript.
