AI Token Budgets Are Coming: What Meta's Per-Engineer Caps Mean for Founders

Key takeaways

• Speaking on Lenny's Podcast, Meta's Adam Mosseri predicted per-engineer AI token caps within a year or two, sized by trust in each engineer's ability to spend in an ROI-positive way.
• Meta shut down an internal token-spend leaderboard after AI costs put the company on track for billions of dollars in 2026.
• Uber blew through its 2026 AI coding budget by April; Microsoft canceled Claude Code licenses and consolidated engineers on its own Copilot CLI.
• Mosseri's memorable line: it is not that hard to build a token incinerator, and that does not create a lot of value.
• The founder translation: what burn-vs-salary is to Meta, burn-vs-ARPU is to your SaaS. Model it before your pricing does it for you.

What did Mosseri actually say?

In an interview on Lenny's Podcast, Mosseri argued that AI token spend will need to be managed like any other constrained resource: GPUs, storage, OpEx, payroll. His forecast: within a year or two, the token burn of a strong engineer might equal their salary or total cost of employment, and at that point caps become necessary. Notably, he tied cap size to trust, meaning engineers with a track record of ROI-positive AI use would get bigger budgets.

Meta does not cap employee token spend today. But it already retired one experiment: an internal token-spend leaderboard, shut down after AI costs put the company on track for billions in 2026. Incentives shape consumption, and a leaderboard is an incentive to burn.

Why are companies capping AI spend now?

Because the bills arrived. Uber reportedly blew through its 2026 AI coding budget by April. Microsoft canceled Claude Code licenses and consolidated engineers around its own Copilot CLI as token costs soared. Meta killed its leaderboard. Three different companies, one pattern: unmetered AI usage grows until finance notices, then governance snaps into place.

Mosseri expects relief eventually, predicting model providers will enter a pricing war that pushes token costs down. But hoping for deflation is not a budget strategy, and every one of these companies acted before prices fell.

What is the equivalent risk for a SaaS founder?

Here is the observation the coverage misses: Mosseri's salary comparison has an exact analog in your P&L. At Meta the question is whether an engineer's token burn approaches their cost of employment. In your SaaS the question is whether a customer's token burn approaches their subscription price. Same ratio, different denominator.

And your version is more dangerous. Meta's burn buys internal productivity it can evaluate. Your burn is spent by customers you cannot coach, on workloads you do not control, under a price you printed on a pricing page months ago. A flat-rate tier with heavy AI usage is your token leaderboard: an open invitation to burn.

Say a power user on a flat $49/month plan starts running agentic workflows daily. If their monthly token cost quietly climbs toward that $49, you are hosting them at a loss before you notice, and the loss scales with your best users' enthusiasm. That is the illustrative version; your real numbers depend on your models, prompts, and usage distribution.

How do you set token budgets without killing usage?

Borrow Mosseri's trust principle, but map it to plans instead of people:

1. 1Meter first. You cannot cap what you cannot attribute per customer and per feature.
2. 2Tie limits to tiers, the way Mosseri ties caps to trust. Higher-paying plans earn higher burn ceilings.
3. 3Prefer soft levers before hard caps: throttles, model downgrades for background work, and usage-based top-ups preserve goodwill better than a wall.
4. 4Watch the ratio, not the absolute. Burn-to-ARPU per account is the single number that tells you when a plan stops making sense.

The takeaway: token budgets are becoming a management discipline at the biggest companies, and per-customer token economics deserve the same seriousness in yours. You can simulate exactly when a user tier flips unprofitable in Calcaas, before your invoices tell you.