Meta is preparing to cap how much computing power its own engineers can spend building with AI — and whoever sets those caps decides what the future looks like.
Instagram head Adam Mosseri told Lenny's Podcast that within a year or two, a single engineer's AI token spend could cost as much as their salary. At that point, he said, "you're going to probably need to put in some caps." Token spend refers to the cost of processing AI prompts and responses — the raw compute that makes AI work. Meta already pulled the plug on an internal leaderboard tracking which employees burned the most tokens, after costs put the company on track for billions in 2026 AI spending alone.
Here's the stake: when a handful of executives ration access to the most powerful creative and analytical tools ever built, they don't need to censor anything. They just decide who gets to build. The cap, Mosseri explained, would be proportional to the company's trust in an engineer's ability to use the budget in an "ROI-positive" way. Who defines positive return? The same people who killed a leaderboard because it was embarrassing.
"It's not that hard to build a token incinerator, and that doesn't create a lot of value," Mosseri said, defending the coming restrictions. He compared token budgets to other resources he allocates — GPU capacity, operating expenses, payroll. That comparison is the tell. Payroll is how a company decides who matters. Token budgets will be how Big Tech decides what matters.
Meta isn't alone. Uber burned through its entire 2026 AI coding budget by April. Microsoft canceled Claude Code licenses and forced engineers onto its own Copilot CLI tool. The scramble to contain costs is industry-wide — and the control is flowing upward.
Meanwhile, the AI that ordinary Americans are already trusting to shape their worldview is being fed a very specific diet. Forbes reported that ChatGPT says up to 63% of its answers are derived from traditional media sources — the same institutional press that has spent years telling you what to think. More than half of Americans using generative AI are now using it to search the internet, per Prosper Insights & Analytics. They aren't getting links to evaluate. They're getting verdicts.
"When they ask AI a question, they essentially receive a verdict," Channel V Media president Gretel Going told Forbes. "And most companies have no idea what verdict is being delivered on their behalf." Gartner forecasts PR and earned media budgets will double by 2027, driven by AI models that prioritize third-party editorial content — meaning the establishment press is about to get even more influence over what AI tells you.
Two tracks, one destination. On the inside, executives consolidate control over who gets to build with AI. On the outside, AI engines serve up narratives built from institutional media, weighted by algorithms no citizen voted for. The old question — who controls the press — just became who controls the compute and who trains the model.
Mosseri said he expects token costs to come down eventually as AI model makers enter a pricing war. But by then, the caps will already be in place. The architecture of control doesn't get dismantled when the resource gets cheaper. It just gets more efficient.