More AI usage, more money lost? The US tokenmaxxing controversy
What happens when AI adoption becomes a token-consumption contest? Lessons from reports that Amazon shut down an internal leaderboard
Key takeaway
Major US technology companies have reportedly been reconsidering systems that measure AI adoption by token consumption. Rising AI costs are cited as a concern. The lesson is that with usage-based AI, making high usage itself the goal can cause costs to grow in unexpected ways.
When usage becomes the target, usage-based pricing bites
No discount for wasteUnnecessary or inflated consumption can still be billed at the normal rate
On May 29, 2026, the Financial Times reported that Amazon had shut down its internal Kirorank AI usage leaderboard, with concerns about rising costs cited as part of the background.
What was reported: Amazon closed an internal leaderboard
On May 29, 2026 (UK time), the Financial Times reported that Amazon.com had shut down an internal AI usage leaderboard called Kirorank.
According to the report, Kirorank ranked employees by token consumption when using the company's Kiro coding agent. It reportedly began as an unofficial dashboard created voluntarily by employees who wanted to encourage AI adoption. That enthusiasm for introducing a new tool across an organization is, in itself, valuable.
However, the FT reported that some employees became focused on their ranking and assigned agents work that was not necessarily needed, increasing token consumption. This behavior reportedly became known internally as tokenmaxxing.
As article 1 explained, tokens measure AI usage, and consumption feeds into billing. The FT cited concerns about these rising costs as background to the leaderboard's closure.
Why did this happen? Volume targets and metered pricing can be a poor match
The point is not to blame individual employees. People pursuing a high ranking may simply be responding conscientiously to the metric they were given. The problem lies in the design of that metric.
- Cloud AI is billed according to usage (article 2)
- When usage volume becomes the goal, increasing that volume can be easy
- The additional tokens are still billable; unnecessary usage does not earn a discount
It is rather like measuring sales performance by the amount of fuel a sales car consumes. Distance driven and revenue generated are different things.
The lesson: measure outcomes, not consumption
According to the FT, Amazon engineering senior vice president Dave Treadwell urged employees to use AI to solve customer problems rather than using it simply for the sake of using AI.
The report also said Amazon adopted a measure closer to outcomes: how often AI-written code actually reached production. That shifts attention from how much was consumed to how much value resulted. Identifying the problem and quickly changing the metric can itself be a useful example for other organizations.
An alternative: change the cost structure behind the fear of overuse
There is a deeper issue. With usage-based AI, promoting adoption and containing costs pull in opposite directions. More use can produce more value, but it also increases the bill. Organizations end up asking people to use AI, but not too much. This is less a failure by any one company than a tension built into the pricing model itself.
An upfront-purchase, fixed-cost on-premises AI system, such as Sovereign GaiXer, changes that tension. Local token consumption does not create an additional token bill; the hardware is purchased upfront, while electricity and applicable operating or maintenance costs still need budgeting. The allocated hardware cost per task falls as usage increases, so encouraging useful adoption need not increase metered AI charges. See article 11 for the cloud cost illustration.
Imagine a staff cafeteria that charges the company by weight instead of a fixed all-you-can-eat fee, while displaying a leaderboard of the biggest eaters. Employees seeking recognition pile food onto their plates even when they are not hungry. The company pays for every gram, so only the bill keeps growing. The issue is not the existence of a leaderboard, but the poor fit between treating consumption as effort and paying by consumption.
Frequently asked questions
What does tokenmaxxing mean?
It was reported as a term for increasing token consumption by assigning even low-value tasks to AI when using more AI brings recognition. Because that consumption is billable, the company can spend more money while merely increasing an activity count.
We do not have a leaderboard. Does this still matter?
Yes. The same incentive can arise from volume targets such as AI usage rates or requests per employee. Pair any volume target with a cost ceiling.
Is encouraging AI adoption itself a mistake?
No. The reported change concerned measuring adoption through consumption, not abandoning AI. Amazon reportedly switched to a metric closer to results, such as code deployed to production.
Can we prevent cost incidents while keeping usage-based pricing?
Yes: set monthly budget alerts, allocate budgets by department, and match more or less expensive models to the task. These controls still require ongoing management effort, which is another cost discussed in article 20.
Can we really let employees use an upfront-purchase system freely?
Local inference does not add per-token charges, though electricity and other operating costs remain. Each machine also has finite capacity. If concurrent use exceeds it, you may need more machines: like adding seats in a restaurant.
If large technology companies struggle with costs, is AI safe for smaller businesses to adopt?
The scale of company-wide metered usage is what makes the multiplication so significant in large organizations. Smaller businesses can choose their cost structure, usage-based or fixed, at the outset and design their rollout around it.
Summary
- The FT reported on May 29, 2026 that Amazon shut down Kirorank, an internal leaderboard comparing AI token consumption.
- Tokenmaxxing behavior and concerns about rising AI costs were cited as part of the background.
- The issue is the fit between a usage-volume metric and metered pricing, rather than individual employees.
- Two lessons: measure outcomes instead of consumption, and consider the structure that makes every extra use increase the bill.
- With upfront-purchase AI, encouraging useful usage does not add local per-token charges; electricity and operating costs remain.
Related articles
Does using more AI make a better employee? The Claudeonomics controversy at Meta
An internal consumption leaderboard became a sensation. What reporting suggests about evaluation pressure, anxiety, and AI bills
How is AI priced?
A simple breakdown of pay-as-you-go pricing
What would one Sovereign GaiXer-style system cost in the cloud?
Recreating an all-in-one GPU system on three clouds: a Tokyo-region monthly cost estimate
This explanatory article is based on public reporting, without independent interviews with the parties involved. Events, statements, and figures reflect the reporting dates and may have changed. Please contact us with factual corrections. The media outlets, research firms, and companies discussed do not endorse Sovereign GaiXer.
Company, product, and service names mentioned are trademarks or registered trademarks of their respective owners.