Meta and the 'Token Incinerators': When AI Costs as Much as the Engineer

The consumption of artificial intelligence by a single engineer could soon cost as much as their salary. This is the scenario that Adam Mosseri, head of Instagram, outlined during a conversation on the YouTube channel Lenny's Podcast, explaining why he believes that within a year or two large companies will need to introduce spending caps on the tokens of their employees.

"One can imagine, at least within a year or two, that the consumption rate of a good engineer could match their salary or the cost of their hiring. And in that context, spending caps are likely to be introduced," Mosseri said.

Tokens are the unit used to measure the processing of requests and responses from AI models, and thus their cost. Mosseri frames this expenditure as a resource to be allocated with the same criteria used to distribute other budget items.

"I consider it like any other resource," he explained. "I have to decide how to distribute the capacity among my teams because I have a limited number of GPUs, CPUs, storage, and RAM. I have to decide how to allocate operating expenses for data labeling budgets. I have to decide how to distribute the payroll budget for the staff." Limits on tokens, he added, will follow the same logic, with a cap per engineer proportional to the company's trust in their ability to spend them profitably.

No Caps Today, But the Internal Ranking Has Already Disappeared

Currently, Meta does not impose any limits on its employees, Mosseri clarified, while believing that it could prove to be a healthy measure in the future. In the immediate term, the company has contained costs by eliminating what it termed "silly things," starting with an internal ranking that effectively rewarded those who consumed the most tokens; it was shut down after the costs of AI had projected Meta towards expenditures in the range of billions of dollars by 2026.

"It's not that hard to build a token incinerator, and it doesn't create much value," he noted.

Meta is not the only one revisiting its approach to AI experimentation. Uber exhausted its budget for AI for code writing intended for 2026 back in April, while Microsoft has shifted toward DeepSeek for Copilot CLI in light of rising costs.

In the longer term, Mosseri expects costs to decrease, with model makers ready for a price war to attract users at the expense of competitors. Until then, token spending remains a line item to be managed with the same tools as payroll or operating expenses: a budget to be assigned team by team, and presumably engineer by engineer.