Korva is a scroll of the internet’s best ideas.
Demand doubles weekly, compute takes months
Buying ahead of growth
Buying double what exists today covers one week of growth. Five times covers less than three weeks. Ten times means paying for capacity far ahead of the user base, and that supply is also used up within a couple of weeks.
The price of waiting
Waiting costs more. Compute bought on short notice carries a premium of three or four times, and a wrong bet on how far ahead to buy is wrong by the same multiple.
A slower pace
Even a slower pace adds up fast. Growth of 5 to 8 percent a day, held for the three or four months it takes to bring compute online, reaches 100 million users. That raises the question of whether to buy compute for 100 million users before they exist.
Why does adding compute take months?
A data center needs land, a power connection, cooling, and the chips themselves, and each has its own lead time. Utilities in some regions have reported waits of several years to connect large new loads to the grid. Advanced AI chips depend on a small number of fabrication plants and packaging steps, which limits how fast more can be made. Construction of a large facility alone commonly takes more than a year.
Why does last-minute compute cost more?
Cloud providers price capacity by how firmly a customer commits to it. Amazon Web Services has advertised discounts of up to 72 percent for committed capacity compared with on-demand prices. Providers can offer those discounts because a customer who commits guarantees them revenue, while on-demand buyers take whatever is left over. When capacity is scarce, the on-demand price is the one that rises first.
How do companies cope when demand outruns chips?
Companies ration scarce capacity with waitlists, usage caps, and queues. OpenAI paused new ChatGPT Plus signups in November 2023 because demand exceeded its capacity. Rate limits on AI model APIs work the same way, capping how many requests each customer can send per minute.
More from Invest Like the Best with Patrick O'Shaughnessy
