Reflections on Dwarkesh and Dylan’s podcast
The question underneath all of this
Does frontier intelligence create enough extra value for OpenAI and Anthropic to keep earning much more from compute than everyone else?
If yes, their loop might work. If cheap models become good enough for most valuable tasks, it might break.
Their argument
OpenAI and Anthropic can make more money from the same amount of compute than everyone else. This means they can pay more for compute, use more of it for research, build better models and repeat.
Where it clashes with what I currently think
1. The catch-up window
What I have been thinking
The shorter the catch-up window gets, the less time a lab has to earn back what it spent.
Their argument
The public catch-up window might not be the real gap. If the labs keep their best models inside and use them for research, the internal gap could grow while public models still look close.
2. How much demand needs the frontier
What I have been thinking
Most tasks do not need frontier intelligence. Once a cheaper model is good enough, the extra intelligence might not be worth paying for.
Their argument
The labs do not need every task. They only need enough tasks where the frontier creates much more value. Dylan estimates that Anthropic can already turn $10–15 million of compute per megawatt into as much as $50 million of revenue.
3. The cost of pushing the frontier
What I have been thinking
Training keeps getting more expensive while competitors catch up faster. That makes the economics of pushing the frontier look fragile.
Their argument
The high cost could make the market more centralized, not less. The lab that makes the most money from compute can outbid everyone else, put more compute into research and make the next model harder to catch.
Source: Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028, Dwarkesh Podcast, August 2026. The financial and compute figures are estimates stated in the podcast, not audited public numbers.