
Menlo Ventures partner Deedy Das is using the viral launch of Kimi K3 to argue a "compute bull" thesis. Within two days of release, K3 cracked OpenRouter's top ten by usage, processing roughly 140 billion tokens per day. Demand has already pushed up against Moonshot AI's capacity limits, forcing the company to pause new subscriptions.
Das figures that in the next few years, the people who lock in GPU, cloud, and power capacity early will be sitting on gold. As long as the best models keep getting smarter, users keep chasing top-tier capability, and inference prices don't crash, compute will remain chronically undersupplied.
There are obvious caveats. If model capabilities plateau, routine tasks shift to cheaper models, or inference costs suddenly drop, the compute crunch eases. And model makers might generate huge revenue but still bleed money on training, talent, and customer acquisition.
https://twitter.com/deedydas/status/2078874115910615544?s=46