Anthropic hits $30B ARR as Meta employees compete on internal token-spend leaderboard

Summary

Anthropic hits $30B ARR as Meta's internal token spending sparks debate over incentive alignment

Anthropic has crossed $30 billion in annualized run rate revenue, marking one of the steepest revenue growth trajectories in software history.

The milestone arrived alongside reporting that Meta employees consumed 60.2 trillion tokens over 30 days—a figure that initially sparked claims Meta was spending roughly one-third of Anthropic's annual revenue on AI inference. That top-line estimate collapsed under scrutiny.

The math doesn't hold

If Meta's token consumption actually cost one-third of Anthropic's $30B ARR—roughly $10 billion annually—each of Meta's ~30,000 engineers would be spending $333,000 per year on tokens. Back-of-envelope math shows that's implausible.

Assuming Meta employees primarily use Anthropic's Opus 4.6 model, the actual cost structure depends on token type. Input tokens cost $5 per million, cached input costs $0.50 per million, and output tokens cost $25 per million. The vast majority of tokens consumed in code generation workflows are inputs—around 98.9% on platforms like OpenRouter—because developers feed enormous context windows (entire codebases) into models that change only small portions of code.

Recalculating with realistic input-output ratios, 60.2 trillion tokens would cost roughly $136 million per month, or $1.6 billion annualized. That works out to $4,500 per engineer per month. If the usage skews more heavily toward cached tokens—likely given Meta's internal coding agents—the cost drops to roughly $55 million monthly, or $669 million annualized, or $1,800 per engineer. That figure aligns with industry norms Jensen Huang cited at GTC, where engineers making $500,000 might command $250,000 annually in token budgets.

The incentive problem

Meta's public token leaderboard has prompted internal discussion about whether ranking engineers by token consumption distorts behavior. One Meta employee reportedly told colleagues that people are "building bots that just run in a loop burning tokens as fast as they can" to game the ranking—similar to how lines-of-code metrics once incentivized cargo-cult programming. The comparison to Goodhart's Law appeared repeatedly: when a measure becomes a target, it ceases to be a good measure.

The risk is real but the scale matters. At $1,800–$4,500 per engineer monthly, the incentive to waste tokens would need to be extraordinarily strong to materially distort Meta's overall token bill. The leaderboard may reflect cultural enthusiasm for AI adoption rather than economically damaging gaming.

The vertical integration story

The token spending reveals a deeper strategic bet. Meta is running frontier inference at scale for internal code generation and ad targeting—tasks that require hundreds of millions of dollars in annual token consumption. If Meta were paying Anthropic or another lab for this, the cost would be material. Building its own frontier model through Meta Super Intelligence Lab lets the company amortize training costs against this internal demand alone, before accounting for any external product.

That shifts the math on whether Meta needs to launch a standalone AI product. Vertical integration of inference—generating trillions of tokens from internally-trained models—can justify the training investment without requiring a blockbuster consumer app. The token volume is the business case itself.