Amazon told its own engineers this week to stop tokenmaxxing.
The memo circulated internally in September and Amazon published it on Tuesday, Oct. 27. Tokenmaxxing is the practice of using AI as heavily as possible with little regard for whether the output justifies the spend. The memo calls token usage, adoption rates, and code-change counts “deeply flawed” ways to measure a team.
It’s worth considering who said this. Amazon’s cloud arm prices Bedrock models by the million tokens. The company selling the tokens is telling its own engineers that counting them proves nothing.
Let’s be clear on this: this is not a retreat from AI. AI is still expected to be at the center of work at Amazon, and the company treats those numbers as signals to read “with care and judgment.” An engineer whose work doesn’t benefit from AI has turned up “a finding, not a failing.”
What it does retreat from is the idea that more usage alone is a good thing. Amazon says it would rather measure long-term outcomes, while conceding those are time delayed and hard to isolate, which is why the proxy metrics stay. Nobody counts tokens because they think tokens are the point. They count tokens because tokens are countable today and value is countable in six months, if ever. The memo lands the same week Amazon reports third-quarter results, after the bell on Thursday, Oct. 29.
This is bigger than Amazon
For the past two years, the working instruction at many companies was straightforward: use AI as much as possible and sort the bill out later. Makes sense, because it’s a learning curve and a baptism by fire. That’s a good strategy in many cases if you want widespread AI adoption. And the proof that it was working was in its usage. How many seats were activated? How many tokens burned? What share of engineers (or other teams) used an assistant every week?
It’s great to see the line climb, but then the bill grows with it. Finance teams know that much, and they’re now struggling to see what the AI spend actually bought.
In many cases, including at Amazon, the bill arrives followed by a correction. Often, it’s an over-correction. The question is what that looks like.
Before you can tell waste from work, you have to be able to see both. Most finance teams can’t. Deloitte’s State of AI in the Enterprise report, released today (Oct. 29, 2026) and based on 3,500 director-to-C-suite leaders across 25 countries, found that 71% of companies track their total AI inference and API spend, but only 51% track token volume and 44% track per-unit inference cost. Deloitte’s own reading is that the visibility gap makes it harder to judge whether the value AI creates justifies its expense.
CloudZero’s own findings in its survey of 252 U.S. finance leaders are similar. More than half (54%) can see AI spend only as a single line on the bill or a rough category split. Yet 61% of those finance leaders say they already spend more on AI than they can justify with measurable outcomes.
2026 State of AI Spend Report
Why Finance Caps What It Can’t See
Learn how finance leaders are managing AI spend, and how they’re approaching FY2027
The cap and the gate
When you can’t see what the money is doing, the only lever left is less of it. That is what most corrections look like in practice, and they come in two shapes.
The first is the cap. Accenture, for instance, rolled Claude out across the firm and then capped each person at $29 a month earlier this month, with anything over that billed to a project code.
The second is the gate. A full quarter (27%) of finance leaders named new approval gates as their primary AI budgeting change for FY27, compared with just 18% who tie funding to outcomes. And gates do damage: 20% of those leading with gates killed or paused AI initiatives, against just 7% who tied funding to demonstrated ROI.
Both are blunt instruments. A flat cap would have equally impacted the team gaming the system and the team shipping real work. So would a committee. Neither can tell the difference, because neither was built to.
Give Amazon credit for aiming more carefully. It didn’t cap its engineers or route them through an approvals process. It went after one specific behavior it judged wasteful, tokenmaxxing, and left the rest of the spend alone. It had done something similar in May, when it shut down an internal leaderboard employees had been gaming to climb the rankings. In both cases the target was the thing encouraging the waste, not the budget.
That is better than a cap, and it still isn’t an answer. What Amazon did was put a second question around the first: not just how much did you use, but what did it get us? Removing the incentive stops people running up the bill for nothing. It doesn’t tell you what the rest of the bill bought.
What to measure instead
Rather than capping spend, cancelling initiatives or gating them behind a committee, the work is knowing what the money is going to and what it is bringing back. Per-unit inference cost is the spend behind one customer, one feature or one transaction, rather than one month’s bill. Put that next to what the customer, feature or transaction returned, and you no longer need a cap to tell waste from work.
Amazon got halfway there by removing the reward. The other half is seeing what the spend bought, which is what CloudZero gives finance to work with. The move was never a smaller number. It was a number you could explain.
