In early 2026, American technology reporting described engineers competing on internal leaderboards ranked by AI tokens consumed, with The New York Times reporting in March that one OpenAI engineer had processed 210 billion tokens in a single week, a practice named ‘tokenmaxxing’. A token is the basic unit of text that a large language model reads and writes, and tokenmaxxing treats volume as evidence of productivity, though a token count records computational spend, not the quality of work produced. Google processed 9.7 trillion tokens a month in May 2024; a year later, that figure was nearly 480 trillion; by May 2026, 3.2 quadrillion. Entelligence Research's May 2026 analysis of one million-plus pull requests across 2,444 organisations found that for every dollar an enterprise spends on AI coding tools, 18 cents becomes shipped product, while the remaining 82 cents goes towards fixing, reworking, and reviewing the code generated by those tools. Amazon employees ran agents on meaningless tasks to protect their usage statistics, Meta withdrew its internal token leaderboard, and Uber spent its full-year 2026 token budget in the first four months, while Salesforce's chief executive put its annual model bill near $300 million. Open-weight models, whose trained parameters are publicly released, lower the price per token, with DeepSeek's January 2025 release of its R1 reasoning model making near-frontier reasoning publicly accessible.