Amazon Urges Its Developers to Maximize AI Usage

The Intensive Use of AI at Amazon: Internal Pressure

Every week, an overwhelming majority of developers at Amazon, over 80%, are encouraged to use artificial intelligence tools. This usage is monitored by an internal ranking that measures token consumption. For some employees, this metric has turned into a competition, pushing them to maximize their use of AI, a phenomenon they call tokenmaxxing. According to the Financial Times, several developers are artificially inflating their scores by using MeshClaw, the AI agent developed by Amazon. An internal memo touts MeshClaw as a tool that "dreams at night to consolidate its learnings, monitors deployments during your meetings, sorts your emails before you wake up." At least thirty-six people have worked on MeshClaw.

MeshClaw: An AI Agent with Extensive Capabilities

MeshClaw is not just a simple digital assistant. It is capable of performing complex tasks, such as deploying code into production or interacting on Slack. However, this autonomy raises concerns among employees, particularly due to its broad access rights to production systems. One developer expressed to the Financial Times their fear regarding MeshClaw's freedom of action, preferring not to let it operate without supervision. These actions, while sometimes unnecessary, have real consequences on production systems.

Previous Incidents with Kiro and Security Challenges

Concerns about MeshClaw are not without precedent. In early 2026, Amazon had already encountered similar issues with Kiro, another AI coding tool that had extensive rights and insufficient security protocols. These incidents raise questions about the management of AI agent security at Amazon, despite the existence of Policy, a safeguard system intended to verify each action of the agents. MeshClaw descends from OpenClaw, capable of executing agents locally on the user's machine. As early as December 2025, Amazon introduced AgentCore and its Policy safeguards, designed to verify each action of an agent before execution.

The Impact of Tokenmaxxing on Demand Projections

The phenomenon of tokenmaxxing has repercussions beyond internal usage. In 2026, tech giants like Amazon, Microsoft, Alphabet, and Meta are expected to make colossal investments, ranging from $650 to $700 billion. Some Wall Street projections for 2027 even exceed one trillion dollars. These projections include the internal usage by developers, which skews demand estimates for GPUs and other components. Jensen Huang, CEO of Nvidia, expressed his concerns about the inefficient use of tokens by engineers, stating in the All-In podcast on the sidelines of GTC 2026 that he would be "deeply alarmed" if an engineer earning $500,000 annually did not consume at least $250,000 worth of tokens per year. Angie Jones anticipates a shift towards measuring efficiency rather than volume.

Massive Token Consumption at Meta and Microsoft

Token consumption practices are not exclusive to Amazon. According to The Information and the Pragmatic Engineer newsletter, employees at Meta consumed 60.2 trillion tokens in 30 days, which would amount to $900 million at Anthropic's public rate. A few days after The Information's revelations, Meta's internal dashboard, Claudeonomics, was removed. Similarly, at Microsoft, several distinguished engineers rank in the top 5 of the internal leaderboard, along with VP level executives in the top 20. The visibility of team statistics at Amazon has been restricted, with only the concerned employee and their manager now having access to the data.

Amazon's Business Strategy with AWS

Meanwhile, Andy Jassy, CEO of Amazon, is considering marketing these AI agents to external clients. AWS, Amazon's cloud division, accounts for 60% of the group's operating profits, and the integration of these tools could strengthen its market position. According to a CNBC survey reported in mid-May by Metaintro, nearly all companies in the Fortune 500 are now measuring AI usage among their employees. Shopify publicly praises heavy AI users and admonishes others.