Meta ranks 85,000+ employees by AI token use with badges — but does burning more tokens actually mean better work?

This explores whether Meta's push to have employees use as many AI tokens as possible is shaping how it builds and spends on AI infrastructure. To be direct: the collection doesn't trace a link from Meta's internal token use to specific infrastructure choices. What it does document is the incentive itself. Meta built an internal leaderboard that ranks more than 85,000 employees by token consumption and awards tiered badges Does token volume measure AI productivity or enable gaming?. Employees responded the way people often respond to a vanity metric. They left agents running idle for hours and padded their prompts. The company has not shown any data connecting that token volume to better work. If token counts are feeding into infrastructure planning, the demand signal is at least partly inflated.

The cost side shows why an inflated signal is expensive. Epoch AI's model of a one-gigawatt data center puts its annualized cost at about $8.5B. Servers account for roughly $5B of that, and energy for only about $0.6B What actually drives AI data center costs?. The main expense is hardware bought up front, not electricity used along the way. So if a company builds capacity to match token demand that's partly theater, it pays for the hardware either way, whether or not the tokens produced anything useful.

The twist from adjacent research is that 'more tokens' and 'more value' aren't the same thing, and they can come apart in either direction. Anthropic found that token spending explains about 80% of performance differences in its multi-agent research systems, so volume can matter Does token spending drive multi-agent research performance?. But automatically tuning the software wrapped around an agent (its harness) cut token traffic by 45–49% with comparable task performance Can agent harnesses be automatically optimized across many environments?. A 115-day case study found that 82.9% of an agent's tokens were cheap re-reads of cached context. That suggests the useful unit of cost is a finished piece of work, not a token Do persistent agents really cost less per token?. Rewarding raw token volume measures the very thing that good engineering tries to shrink.

There's also a fair defense of Meta's approach. Brynjolfsson, Rock and Syverson's 'productivity J-curve' argues that early AI adoption looks unproductive because the real investment goes into training, new processes and new software, none of which shows up in output yet Why does AI productivity appear weak when investment is high?. On that reading, paying for wasted tokens could be the price of getting tens of thousands of people fluent with the tools. The critique is that this only holds if someone eventually checks what the tokens produced, and the leaderboard has no mechanism for that.

If you want the bigger frame, one cluster of notes argues that AI output behaves less like a product and more like a currency. In that framing, its value depends on what it does for the person who receives it Does AI actually commodify expertise or tokenize it?. A currency issued faster than anything backs it loses value What actually backs the value of AI-generated intelligence?. A token leaderboard is a literal version of that dynamic: an organization minting more of something while measuring how much exists rather than what it buys.