Why does Arvind Krishna think most AI tokens will move to open-weight models?

Arvind Krishna predicted that 75 to 90 percent of AI token use will go to open-weight models for three reasons: cost, intellectual property and geopolitics. Self-hosting a Chinese open-weight model such as Kimi K3 could cut cost from about $20 to $1–2 per million tokens; companies will not want their chain of thought logged by another firm; and Western open models should catch up within a year or two.

• Krishna said Chinese open-weight models carry little danger when run in a private sandbox rather than from a cloud.
• Krishna put the gap between frontier and open models at six to 12 months at most.
• Nikesh Arora said his company would likely train an open-weight model on its own data for auditability and explainability.

▸In their words

“The gap in these things are six to 12 months at the most.” Arvind Krishna · at 14:19 — open in the transcript →

Source: our own transcription of the episode. The full conversation is on All-In with Chamath, Jason, Sacks & Friedberg (2026-10-21). Read the whole episode summary →

▸More from this episode