Adam Mosseri, head of Instagram and one of Meta's senior leaders, predicts that within one to two years companies will need to impose per-engineer AI token budgets — similar to managing headcount or operating expenditure. Mosseri made the remarks on Lenny's Podcast, where he acknowledged that the AI token burn rate of a single productive engineer could soon equal their annual salary.
Key takeaways
- Adam Mosseri predicts per-engineer AI token caps at Meta within 1–2 years
- Meta shut down an internal token spend?Token spend: The total cost of tokens processed by AI models — measured in input and output tokens and billed by model providers on a pay-per-use basis. leaderboard and reined in unproductive AI usage after costs approached billions of dollars annually
- Uber exhausted its 2026 AI coding budget by April — four months into the fiscal year
- Microsoft dropped Claude Code licenses for engineers, consolidating them around its own Copilot CLI tool
- Mosseri compares token budgets to payroll: a finite resource allocated across teams based on ROI
Token spend: a new line in the operating budget
Token spend — the cost of processing AI model queries — has moved from a technical footnote to a real budget line item at major tech companies. Meta AI encountered this acutely: the company launched an internal token spend leaderboard to encourage efficiency, but quickly shut it down after costs approached billions of dollars annually. As The Information reported, engineers were literally competing to maximize token consumption, generating waste rather than efficiency.
Mosseri described the lesson plainly: "It's not that hard to build a token incinerator, and that doesn't create a lot of value." Meta addressed the problem by eliminating what he called "silly things" — unproductive AI usage that did not translate into business outcomes. The company does not yet have hard per-engineer limits in place.
The race to limits: Meta is not alone
Meta's token budget problem reflects a wider trend hitting multiple major tech companies. Uber exhausted its entire annual AI coding tools budget by April 2026 — four months into the fiscal year — forcing abrupt cuts. The company restricted access to premium models, leaving engineers with cheaper alternatives.
Microsoft chose not to renew its Claude Code licenses for engineers, consolidating them around its own Copilot CLI tool. According to The Verge, the decision was driven by cost. 1Password entered the AI cost management segment, announcing a tool for tracking and optimizing enterprise LLM spending — a signal that AI token governance is becoming a dedicated product category.
Token budgets as a new management discipline
Mosseri framed his view through an analogy to other corporate resources. Just as a company manages GPU capacity, storage, labeling budgets, and headcount allocations — it will have to manage tokens. The difference is that tokens are an elastic, invisible resource for the end user: an engineer using an AI coding agent does not see the cost of each query in real time, while the company sees it only on a bill.
AI model pricing is consumption-based — cost per million input and output tokens. For frontier models, input token pricing ranges from a few to over ten dollars per million tokens; reasoning models (which generate large numbers of internal thinking tokens) are more expensive still. A heavy agentic AI user can easily generate several million tokens per day — which at $15 per million output tokens amounts to tens of dollars daily per engineer.
At 10,000 engineers with that usage profile, the annual token bill reaches $150M — and those are the numbers Meta is beginning to observe in its systems. Mosseri acknowledged that at that scale, per-engineer limits become inevitable, with the cap proportional to the company's trust in a given engineer's ability to deploy the budget in an ROI-positive way.
Why this matters
Mosseri's remarks matter not as a Meta story, but as a signal of the whole industry maturing. For the past three years, companies have had de facto unlimited access to AI coding tools under generous corporate licenses — productivity was measured by "does it work", not "what does it cost". That phase is ending. Token spend is becoming an OpEx?OpEx: Operating Expenditure — recurring operating costs such as software licenses, cloud computing and salaries, as opposed to CapEx (one-time capital investments). category like any other: measurable, reportable, and managed. For AI model providers (Anthropic, OpenAI, Google), this means growing pressure to cut prices or introduce enterprise pricing with more predictable cost structures. Mosseri explicitly predicted a pricing war between model providers in the coming years. For tech companies, it means building new ROI metrics for AI — today few can precisely show how much value they have generated from the tokens they have spent.
What's next
- Meta does not yet have hard per-engineer token limits — Mosseri is talking about a 1–2 year horizon, contingent on whether model prices fall fast enough
- The AI cost management tools segment is growing — 1Password is likely one of the first of many new players offering LLM spend monitoring and optimization for enterprise
- Pricing pressure on frontier model providers will increase — Mosseri predicted a pricing war, which aligns with the trend (DeepSeek cut prices 75% in July 2026)
