Uber Capped AI Spending. Most Companies Haven't Even Started Thinking About This.

TL;DR

• 1

  Uber's cap makes AI spend feel very real: After burning through its 2026 AI budget in about four months, Uber set a $1,500 monthly cap per tool and says 10% of its code is now written by AI.

• 2

  Even the biggest buyers think frontier models cost too much: Microsoft's Mustafa Suleyman said Anthropic is "extremely expensive," and Microsoft responded by announcing seven in-house models, including one it claims rivals Claude Opus 4.6 on a coding benchmark for less money.

• 3

  Token usage is exploding faster than companies expected: Sam Altman said AI cost concerns appeared "quite suddenly" in early 2026, while examples cited in the segment included 603 billion tokens used in 30 days and 210 billion in a single week by OpenAI employees.

• 4

  Model routing is becoming the practical fix: Factory says its new router can pick the cheapest capable model per task and cut costs by 25%, and Coinbase says it has kept costs roughly flat even as token usage grows exponentially by routing prompts to cheaper models when possible.

• 5

  Outcome-based pricing sounds appealing, but only works in narrow cases: Sierra argues AI should be priced on outcomes rather than usage, but the hosts stress that autonomy and attribution determine whether outcome, usage, seat, or hybrid pricing actually makes sense.

• 6

  AI literacy may matter as much as model quality: The hosts argue most employees do not understand when to use Opus versus Sonnet versus Haiku, which means companies are burning premium tokens on simple tasks like emails that cheaper models could handle just fine.