Amazon's tokenmaxxing memo is a warning to every buyer without a per-engineer AI cap

In May, Amazon shut down KiroRank, an employee-built leaderboard, after people worked out that the fastest way up it was to burn tokens. On Tuesday the company published a memo, circulated internally since September, calling token usage and adoption rates deeply flawed ways to judge a team. It went further and told engineers that reporting where AI does not help their work counts as a useful result, and that being able to explain it shows the judgement Amazon is asking for. I expected corporate throat-clearing. I read it as a confession from one of the largest buyers of AI tooling on earth that once you pay for usage and rank people on it, usage is what you get.

Then comes the part I disagree with. An Amazon spokesperson told CNBC the company has not capped employees' token consumption. Amazon can afford to fix this with culture and memos. Your company probably cannot, and I think every buyer should do the thing Amazon declined to do: put a hard ceiling on each engineer's AI spend, written into the vendor contract, before the next renewal.

The case rests on the tail of the distribution, where the money goes. Uber engineers reportedly ran through the company's entire annual AI budget on Claude Code by April. Databricks chief executive Ali Ghodsi publicly pointed to one engineer who spent more than $7,000 on tokens. One engineer at OpenAI logged 210 billion tokens in a single week. Each of them had a tool with no meter on the desk and a manager who had been told adoption was the goal.

The sensible objection comes from the averages. Anthropic's own cost documentation puts typical enterprise spend at about $13 per developer per active day, roughly $150 to $250 a month. Amazon's memo says its most active users, the teams it calls frontier teams, show a correlation between AI use and higher delivery, and that people who refuse to adapt are falling behind. Jensen Huang went to GTC and proposed paying engineers about half their base salary again in compute tokens, which for his top people could mean $250,000 a year. On those numbers the typical developer looks cheap, and a cap looks like fussing over small change.

I think the averages argue for the cap. A ceiling set well above the typical spend costs the typical developer nothing and stops the runaway. Atlassian already runs this. Engineers get an AI wallet of $500 to $2,000 a month for research and development staff, sized by role, with a warning as they approach the limit, a hard stop when it runs out, and a route to ask for more. That ceiling sits several times above Anthropic's average figure. A frontier team that ships more can make the case for a bigger wallet, which is the return-on-investment test AWS's Marc Brooker described to CNBC when he talked about counting launches and successful production changes. As for Huang, he sells the chips those tokens run on. I would expect him to want engineers paid in his product.

The contract matters more than the internal policy because vendors are deciding the billing model right now. Cursor lists Teams Premium at $120 per user a month with five times the Standard agent limits, Enterprise on custom terms with pooled usage, and a token rate of $0.25 per million tokens on third-party models for Teams and Enterprise accounts that applies even when you bring your own key. Pooled usage sounds generous at signature. In practice it means one enthusiastic engineer can drain the allowance for the whole team, and the overage invoice lands with finance months later, which is how you get an Uber-style surprise in April. The alternative is the blunt response Microsoft reportedly chose: cancel Claude Code licences and herd everyone on to Copilot. I think ripping out a tool your engineers like is an expensive way to discover you never had spending controls.

So negotiate for the boring clauses. A per-seat monthly ceiling that the vendor enforces, a warning threshold the engineer sees, no automatic overage without an approver, and a usage export broken down by named user every month. I expect vendors to give ground on the export fairly easily and to fight hardest over the enforced ceiling, because uncapped seats are where their margin grows.

Before your next AI tooling renewal, ask your vendor for last month's spend by individual user and sort it from the top. If the export does not exist, or if a handful of names account for most of the bill, you have found the clause to add, and you have found it before the vendor priced your engineers' habits into next year's minimum commitment.

Prompted by Amazon's anti-tokenmaxxing memo says engineers should be pickier about using AI, CNBC.