Korva is a scroll of the internet’s best ideas.
Most AI value goes to users
Value goes to users
Meta, rumored at one point to be as much as 10% of Anthropic's business, makes far more from optimizing its ad algorithms and getting engagement time 5% longer than Anthropic earns from selling it the tokens. Anthropic's revenue per megawatt has gone from $20 million toward $100 million, and it still pays about $13 million per megawatt for much of the compute it buys.
Where profit has moved
The place where value lands keeps moving. A year ago the hardware supply chain held the gross margin while nearly everyone else lost money: OpenAI and Anthropic were plowing venture money in, selling tokens for less than the infrastructure cost, and many hyperscalers were building without knowing whether it would pay off. In 2023 the memory makers made no money on HBM even though the value they delivered was huge. Now the model layer has gone from negative to massive positive gross margins, and the app layer has generated very little so far.
The labs and compute
One expectation was that the labs would end up with all of it. SpaceX showed otherwise by selling compute for $25 million to $40 million a megawatt.
Why do users capture most of the value of new technology?
Economist William Nordhaus estimated in 2004 that innovators kept only about 2 percent of the total value of technological advances in the US between 1948 and 2001. The rest went to customers as lower prices and better products. Competition among sellers pushes prices toward cost, which hands most of the benefit to buyers.
How do ad algorithms turn attention into revenue?
A platform earns when people spend more time on it, because each extra minute shows more ads. Better prediction of what a user will engage with and buy raises what advertisers will pay in the auction for each slot. A gain of a few percent in time on the platform is therefore worth a large amount across billions of users.
Why did memory makers earn little on HBM in 2023?
DRAM prices fell sharply in 2023, and the three major memory makers reported losses that year. High-bandwidth memory was still a small share of their sales. Demand from AI chips grew quickly after that.
More from Dwarkesh Podcast
