Meta's Adam Mosseri says the company will likely need to impose per-engineer caps on AI token spending within one to two years, as the cost of running AI-assisted coding rises toward the cost of employing the engineer using it. Speaking on Lenny's Podcast, the Instagram head framed token budgets as the next line item on a corporate P&L, alongside payroll, GPUs, and labeling. Meta is on track to spend billions of dollars on AI in 2026, and the company recently shut down an internal token-spend leaderboard that was encouraging engineers to burn through capacity for sport.
The comments land amid a broader corporate reckoning with runaway AI coding costs. Uber blew through its entire 2026 AI coding budget by April. Microsoft canceled Claude Code licenses across its engineering org and consolidated staff onto its own Copilot CLI tool, a move driven at least in part by soaring per-seat token bills.
Mosseri's argument is that tokens are simply the newest constrained resource, no different in kind from compute, memory, or headcount. He said he already has to decide how to deploy GPUs, CPUs, storage, and RAM across teams, how to allocate operating expenditure for labeling budgets, and how to allocate payroll for headcount. Tokens, in his framing, join that list.
Key facts
- 01Mosseri says a strong engineer's AI token burn rate could match their salary within 1 to 2 years.
- 02Meta shut down an internal AI token spend leaderboard after costs put the company on track for billions of dollars in 2026.
- 03Uber blew through its entire 2026 AI coding budget by April 2026.
- 04Microsoft canceled Claude Code licenses and consolidated engineers onto its own Copilot CLI tool to control token spend.
- 05Meta currently has no per-employee token caps, but Mosseri expects proportional caps tied to ROI performance.
The caps, when they arrive, would be proportional to how much a company trusts a given engineer to spend the budget in an ROI-positive way. That is a meaningful shift from the current posture at most large tech companies, where AI coding tools are largely uncapped as employers race to prove productivity gains. Meta itself does not currently impose token caps on any employee, Mosseri said, but he expects that to change.
The internal token-spend leaderboard at Meta is a small window into how quickly this got out of hand. Gamifying token usage inside a company that is already on track for billions in 2026 AI outlays is exactly the kind of behavior that turns a productivity tool into a cost center. Killing the leaderboard was one of the "silly things" Meta stopped doing to rein in its bill.
The underlying economics are straightforward. If a senior engineer costs the company several hundred thousand dollars a year fully loaded, and that same engineer's daily AI usage begins to run into hundreds of dollars per day across long agentic sessions, the arithmetic converges fast. Mosseri's one-to-two-year timeframe is not a distant projection.
Mosseri also offered a longer-term counterweight: he expects token prices to fall as frontier labs enter a pricing war for developer mindshare. That prediction is consistent with what has already happened at the API layer, where per-million-token prices for flagship models have dropped repeatedly over the past two years. The question is whether prices fall faster than usage grows. So far, usage is winning.
For enterprise buyers, the near-term implication is that FinOps for AI is about to become a real discipline, not a slide in a vendor deck. Companies that treated 2025 as a land-grab, handing every engineer a Claude Code or Copilot seat with no meter, are discovering in 2026 that agentic workflows can burn through six-figure budgets in a quarter. Uber's April blowout is the canonical example, and it will not be the last.
There is a skeptic's case here worth naming. Per-engineer caps could suppress exactly the kind of experimentation that produces the productivity gains AI vendors keep promising. If engineers ration their agent runs the way consultants ration billable hours, the tools stop being ambient and start being scarce, which is a very different product. Mosseri conceded the risk implicitly when he said caps would depend on trust in individual ROI.
The Meta signal matters because it reframes AI coding tools from an unlimited utility into a metered resource, and it does so from the operator side rather than the vendor side. Anthropic, OpenAI, and Microsoft have spent the past year selling seat-based unlimited-feeling access; if Meta, Uber, and Microsoft's own internal decisions are any guide, the next twelve months will be about caps, quotas, and chargebacks. That is a harder sales motion than "give everyone a seat," and it will pressure vendors to show measurable ROI per token, not just per developer.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
