A neutral, sourced, public record of organizations capping, cutting, or centralizing employee AI-coding spend — the 2026 shift from "use AI as much as possible" to rationing ("tokenmaxxing → tokenminimizing").
Last updated: 28 Jun 2026
Every entry in the table below is backed by a named source. Spotted one we're missing? See CONTRIBUTING.
| Company | What they did | Figure | Date | Tool(s) | Source(s) |
|---|---|---|---|---|---|
| Uber | Per-employee monthly cap on agentic AI coding spend (per tool; exceedable with approval), after exhausting its 2026 AI budget in ~4 months. | $1,500 / month per tool | Jun 2026 | Cursor, Claude Code | Bloomberg · TechCrunch |
| Microsoft | Cancelling internal Claude Code licenses in the Experiences + Devices division; engineers moved to GitHub Copilot CLI. | Reported ~$500–$2,000 / engineer / month | by 30 Jun 2026 | Claude Code → GitHub Copilot CLI | Windows Central · TechRadar |
| Meta | Capping employee AI token usage with real-time spend tracking via a new "AI Gateway" dashboard; dismantled its "Claudeonomics" usage leaderboard (73.7T tokens in ~30 days, 85k+ employees) and is nudging staff to in-house MetaCode over Claude, with internal AI cost tracking toward "billions" in 2026. | Per-employee token caps; costs "approaching billions" | Apr–Jun 2026 | Claude → in-house MetaCode ("AI Gateway") | Fortune · The Next Web |
| Walmart | Replaced unlimited token usage with a set per-employee token allowance on its in-house "Code Puppy" AI agent after demand outran budget (announced by Global CTO Suresh Kumar). | Per-employee token allowance (previously unlimited; exact figure undisclosed) | Jun 2026 | Code Puppy (in-house) | Bloomberg · Arkansas Democrat-Gazette |
| AT&T | Began limiting some employees' access to GitHub Copilot amid rising AI-tooling costs. | Access limits for certain staff (exact cap undisclosed) | Jun 2026 | GitHub Copilot | The Next Web · GuruFocus |
- Amazon — earlier restricted Claude Code, but after internal developer pushback (a ~1,500-endorsement employee thread) expanded access to external agentic tools (Claude Code + Codex, on Bedrock) company-wide in May 2026, beyond its in-house Kiro — currently loosening, not capping. Sources: Futurism · The New Stack
An organization is listed in the table only with a named, linkable source (reputable outlet or first-party statement). Record the mechanism precisely (per-tool dollar cap vs license cancellation vs centralization). Unconfirmed leads stay in "Candidate entries" marked [VERIFY]. Figures change — re-check periodically.
A spend cap is a blunt instrument: it throttles an organization's highest-leverage engineers and does nothing about proprietary code leaving for third-party models. The structural alternative is routing, not rationing — a layer that sends only what's worth it to a premium model and keeps sensitive work local. A reference implementation of that pattern is Switchboard (MIT; reproducible benchmark, DOI).
Data licensed under CC BY 4.0. Contributions welcome.