Amazon is the latest hyperscaler, following similar practices documented at Meta and Microsoft last month when employees were caught inflating the consumption of AI tokens to meet internal usage targets. financial times I will report it.
The company set a goal to require more than 80% of developers to use AI tools weekly and tracked usage with internal leaderboards. Some employees said F.T.They were using MeshClaw, an in-house agent platform that can interact with Slack to initiate code deployments, prioritize emails, and maximize token counts. Amazon said usage statistics are not factored into performance reviews, but multiple employees said they believed managers were monitoring the data. One said there was “tremendous pressure to use these tools”, while another explained how tracking created “perverse incentives”.
This practice, known as “tokenmaxxing”, has become widespread enough to generate its own vocabulary and leaderboards, but if a significant proportion of AI consumption is performative across workplace cultures, how reliable are the demand figures with hundreds of billions of dollars earmarked for AI infrastructure procurement?
from the latest videos
With combined 2026 capital spending for Amazon, Microsoft, Alphabet, and Meta hovering between $650 billion and $700 billion, with some Wall Street forecasts topping $1 trillion in 2027, all hyperscalers are telling investors that inference power is being absorbed as quickly as it can be deployed. Internal developer consumption is clearly part of that absorption, sitting alongside external customer payments in the usage data that informs capacity planning, GPU ordering, HBM procurement, power infrastructure, and more.
Tokenmaxxing does not mean demand is fabricated. Enterprise AI adoption is growing and inference workloads are being scaled up into production environments. However, there is a difference between adoption and consumption intensity. While the former is a persistent demand driver, the latter is gamified and is now being amplified by the incentive structures these companies have built. The picture is further muddied by reports that AI costs more than real workers.
Meta’s internal leaderboards last for several days after the public launch, and Amazon recently restricted the visibility of usage statistics across teams. And as the measurements change, the consumption intensity they encouraged also changes.
Nvidia CEO Jensen Huang highlighted token consumption per engineer as a key metric, saying he would be “very concerned” if an engineer making $500,000 a year wasn’t spending at least $250,000 in tokens. Nvidia’s inference growth clearly relies on its consumption being sustained and compounding productive workloads, as every inflated token is actual GPU time.
Get the best Tom’s Hardware news and in-depth reviews delivered straight to your inbox.
Angie Jones, former VP of AI Tools Engineering at Block, said: lead development She hoped the industry would pivot toward measuring efficient token usage rather than celebrating volume. In a cycle where GPU orders and power commitments are made years in advance, the quality of the demand forecasting behind them is critical. Hyperscalers are building for a world where every knowledge worker spends hundreds of thousands of dollars in compute per year. Whether that consumption is productive or performative will determine how much of this year’s $700 billion will generate sustainable revenue.
to follow Tom’s Hardware on Google Newsor Add us as a preferred sourceget the latest news, analysis, and reviews in your feed.
