Industry · 2026-07-17 · 7 min read

Adam Mosseri: AI Token Costs Could Soon Match Engineer Salaries

Instagram head Adam Mosseri warns that within 1–2 years, a strong engineer's token burn rate could equal their salary. Inside Meta's 73.7T-token 'Claudeonomics' leaderboard, the AI Gateway response, and what it signals for every company that hands out uncapped AI access.

TL;DR

Adam Mosseri told Lenny's Podcast that within 1–2 years, a strong engineer's token burn rate could match their salary.
Meta ran an internal leaderboard nicknamed "Claudeonomics" with tiers like *Session Immortal* and *Token Legend*. Employees burned 73.7 trillion tokens in ~30 days.
Mosseri's verdict on token leaderboards: "It's a terrible idea."
Meta's response: a centralized AI Gateway for real-time spend monitoring, plus formal token budgets by 2027.
Meta is spending $125–145B on AI capex in 2026 *and* rationing per-engineer inference. Both are true. Both are the new normal.

!Meta logo

The Quote That Should Reset Every AI Budget Meeting

Speaking on Lenny's Podcast, Instagram chief Adam Mosseri put a number on something most engineering VPs have been quietly worrying about:

> "The burn rate of a strong engineer might be the same as their salary or their cost of employment."

Not in a decade. Within a year or two. From an executive running one of Meta's largest product organizations — not a vendor selling you a FinOps dashboard.

Tokens, Mosseri said, now sit beside GPUs, storage, RAM, and payroll in the budget conversation. The bill shows up whether the work was useful or not.

Meta's "Claudeonomics" Leaderboard Was the Warning Shot

According to reporting from *The Decoder* and *MLQ News*, Meta ran an internal system nicknamed Claudeonomics — a nod to Anthropic's Claude, one of the third-party coding tools used inside the company. It ranked employees and teams by raw token consumption, with tiers reportedly called:

Session Immortal
Token Legend

After Meta tied AI-driven work results to 2026 performance reviews — with bonuses on offer for top performers — employees burned through 73.7 trillion tokens in roughly 30 days.

!Anthropic Claude logo

Asked about the leaderboard on Lenny's Podcast, Mosseri didn't hedge: "It's a terrible idea."

He's right. Rank people by token usage and some of them will optimize for token usage. It's the same failure mode as paying sales teams on calls instead of revenue, or support teams on ticket closures instead of solved problems. You get the metric you ask for. Meta just ran the experiment with expensive AI models attached.

What the 73.7 Trillion Tokens Actually Cost

To put Meta's month in perspective, here's what 73.7T tokens would run on today's frontier models — assuming a rough 70/30 input/output split:

Meta pays enterprise rates and has heavy caching, so the real invoice is a fraction of the sticker figures. But the shape is unmistakable: an uncapped incentive to consume, applied to 40,000 engineers, produces bills that show up on the CFO's desk.

Meta's Response: AI Gateway + Formal Token Budgets

According to the same reporting, Meta's fix has three parts:

1. AI Gateway — real-time spend monitoring

A centralized platform that tracks usage and spending across teams in real time and flags abnormal spikes before the invoice arrives. Same pattern as cloud FinOps circa 2016: you can't govern what you can't see.

2. Formal per-team token budgets by 2027

Department-level allocations, treated like headcount or cloud credits. Overage requires justification, not just a bigger prompt.

3. Push toward MetaCode and away from external tools

*The Information* separately reported that Meta restricted engineers in its Applied AI division from using Anthropic's Claude Code and OpenAI's Codex without approval, citing model-distillation concerns. Different issue from cost — same direction of travel: less uncontrolled external AI use, more internal visibility, fewer blank checks.

!OpenAI logo

The Contradiction That Isn't

Here's the detail that makes Mosseri's warning land hard. In the same window Meta was rationing engineer token spend, it was:

Raising 2026 capex to $125–145B (AP), driven by AI data centers.
Cutting ~8,000 jobs in May (The Verge), around 10% of the workforce.
Closing thousands of open roles.

Meta is willing to spend staggering sums on AI infrastructure and telling employees inference costs need limits. That's not a contradiction — it's the new accounting:

Capex (data centers, chips, training) is a top-of-house capital decision.
Inference is an operating expense that spreads quietly through teams before anyone with budget authority has a clean view of it.

If you run a company smaller than Meta, you don't get their margin for error. You get the bill.

What This Means for Everyone Else

If Meta — with $145B in AI capex firepower — is now metering per-engineer inference, three things are true for the rest of us:

1. Uncapped AI access is over

The 2025 memo — *"use more AI, you'll be evaluated on it"* — is being rewritten. The 2026 version reads: *"use AI where it moves the metric that matters, and here is your monthly budget."* Expect per-employee caps in the low-four-figures/month at most large enterprises by mid-2026.

2. Cost per merged PR beats tokens per developer

Usage without an attached outcome is now treated as waste. The KPIs shifting into engineering scorecards:

Cost per merged PR
Cost per resolved ticket
Cost per shipped feature
Cost per resolved customer conversation

Not tokens. Not sessions. Outcomes per dollar.

3. Model routing is a first-class architecture decision

The way to survive a per-engineer cap isn't to use AI less — it's to route the cheap 80% of tasks to sub-$1/M open-weight models (DeepSeek, GLM, Kimi) and reserve frontier calls for the tasks that genuinely need Claude Fable 5 or GPT-5.6. Coinbase's smart-routing playbook cut their AI bill roughly in half using exactly this pattern.

The Bigger Picture

The first phase of the AI productivity story was giving everyone access. The next phase is asking whether the work produced is worth the tokens burned to produce it.

If Meta has to ask that question while spending up to $145B on AI infrastructure, you should assume your company has to ask it too. Mosseri's framing — tokens allocated like headcount, GPUs, and cloud credits — is going to be normal well beyond Instagram within twelve months.

The companies that treat AI spend the way they treated cloud spend in 2015 — a finite, governed budget with real KPIs — will keep shipping. The ones still running a *Token Legend* leaderboard are the ones writing the retro next quarter.