Amazon employees inflated AI token consumption to meet internal usage targets, the Financial Times reports. The practice — dubbed "tokenmaxxing" — follows similar gaming of metrics at Meta and Microsoft.
Key facts
- Amazon set target requiring >80% of developers to use AI tools weekly.
- Employees used MeshClaw platform to maximize token consumption.
- Combined 2026 capex for four hyperscalers is $650B-$700B.
- Meta's internal leaderboard lasted days after public exposure.
- Jensen Huang expects $250K in token consumption per $500K engineer.
Amazon is the latest hyperscaler where employees have been caught inflating AI token consumption to hit internal usage targets, following similar behavior documented at Meta and Microsoft last month, the Financial Times reports. The company set targets requiring more than 80% of its developers to use AI tools each week and tracked consumption on internal leaderboards. Some employees told FT they had been using MeshClaw, an in-house agent platform that can initiate code deployments, triage emails, and interact with Slack to maximize their token numbers. Amazon said usage statistics would not factor into performance evaluations, but multiple employees said they believed managers were monitoring the data. One said there was "so much pressure to use these tools," another described how tracking created "perverse incentives."
The practice — dubbed "tokenmaxxing" — has become widespread enough to generate its own vocabulary and leaderboards, but beyond workplace culture, if a meaningful share of AI consumption is performative, how reliable are the demand figures that hundreds of billions in AI infrastructure procurement are being allocated against? Combined 2026 capex from Amazon, Microsoft, Alphabet, and Meta is tracking between $650 billion and $700 billion, with some Wall Street projections exceeding $1 trillion for 2027, and every hyperscaler has told investors that inference capacity is being absorbed as fast as it can be deployed. Internal developer consumption is obviously part of that absorption, and it sits alongside paying external customers in the usage data that informs the likes of capacity planning, GPU orders, HBM procurement, and power infrastructure.
Tokenmaxxing doesn’t mean the demand is fabricated — enterprise AI adoption is broadening, and inference workloads are scaling into production — but there’s a distinction between adoption and consumption intensity. The former is a durable driver of demand, whereas the latter is gameable, and it’s currently being amplified by the incentive structures that these companies built. The water is further muddied by reports that AI is more expensive than actual workers. Meta's internal leaderboard lasted days after public exposure, and Amazon recently restricted visibility of team-wide usage statistics. And when measurement shifts, the consumption intensity they incentivized will shift with them.
Nvidia CEO Jensen Huang has highlighted per-engineer token consumption as a key metric, stating he’d be "deeply alarmed" if a $500,000-a-year engineer was not consuming at least $250,000 in tokens. Nvidia's inference growth obviously depends on that consumption being a productive workload that persists and compounds because every inflated token is real GPU time.
What to watch
Watch for Amazon's Q3 2026 earnings call: if management adjusts internal usage metrics or acknowledges inflated consumption, it would signal a shift in how the company reports AI demand to investors. Also monitor whether other hyperscalers follow Meta and Amazon in restricting visibility of usage leaderboards.
