There's a new word floating around Amazon's hallways: "tokenmaxxing." It means gaming your AI token usage numbers.
Here's why: Amazon set a hard target this year — every week, more than 80% of developers must use AI tools. The company tracks everyone's token consumption (tokens are the units AI models process) on an internal leaderboard. Amazon says the data won't affect performance reviews. Employees don't buy it. "Managers are definitely looking," one employee told the publication that broke the story. "Tracking usage creates distorted incentives. Some people treat it as a competition."
So some employees started using MeshClaw, an internal tool Amazon recently rolled out widely, to pad their stats. MeshClaw can create AI agents that take over office software — deploying code, handling email, operating Slack. Workers are asking it to do tasks that don't really need AI, just to bump up their token count.
MeshClaw was inspired by OpenClaw, an open-source project that went viral in February and is part of a wave of agent tools that have topped token usage charts. OpenClaw lets you run AI agents locally on your own machine. More than 30 Amazon employees built MeshClaw, and its internal docs sound almost sci-fi: "It dreams at night to integrate what it learned during the day, keeps an eye on deployments during meetings, and has your email sorted by priority before you wake up."
Amazon isn't alone. Meta employees are doing the same thing, burning tokens on internal leaderboards.
This is all part of a broader Silicon Valley anxiety: massive AI investments need to show returns. Amazon's capital expenditures are expected to hit $200 billion this year, mostly going to AI and data centers — the company and its peers are hunting for cheap debt to fund the AI arms race. The company needs to prove the money was well spent, so it introduced metrics and leaderboards — and, inevitably, data padding. Big Tech's combined free cash flow is expected to plummet to $4 billion in Q3 as the spending frenzy continues.
Some employees also worry about MeshClaw's security. An AI agent that can manipulate your office systems could cause real harm if it makes a mistake or does something unexpected. "The default security settings scare me," one employee said. "I'm not letting it run loose."
(Note: "tokenmaxxing" follows the internet slang pattern of "-maxxing," meaning doing something to the extreme. Here it specifically means maxing out AI token usage.)