When does Nikesh Arora think enterprises should use frontier AI models instead of small models?

Nikesh Arora said frontier models make sense only where the business value at stake justifies the cost, such as stopping an active breach that could cost billions. For routine work his company compresses small language models, trains them to 99.9 percent accuracy on one task, and runs them on the endpoint, inspecting each device's traffic for about $20 to $25 a year.

• Arora told teams to use the best and easiest model during experimentation and optimise cost only later.
• Arora said running every employee's traffic through OpenAI tokens would hand OpenAI a year of revenue from one customer.

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