The era of Big Tech’s "token incinerators" is drawing to a close. Adam Mosseri, Head of Instagram, has effectively signaled a paradigm shift: artificial intelligence is no longer an infinite resource but a scarce asset. In a recent episode of Lenny’s Podcast, Mosseri delivered a sobering forecast for Silicon Valley: within a year or two, the cost of AI tokens for a high-performing engineer will equal their total compensation. This means the cost of headcount effectively doubles, transforming experimentation into a luxury that requires strict rationing.
The collapse of ‘borderless R&D’ culture
Meta has already begun dismantling the culture of unchecked research. According to Mosseri, the company shut down its internal token expenditure leaderboard—not as a choice, but as a necessity. The rate of budget consumption was putting the corporation on a trajectory to spend billions of dollars by 2026. This isn't just Mark Zuckerberg’s local panic; it is a market trend. The industry has hit a ceiling where the cost of computation is outpacing productivity gains. The case of Uber, which reportedly managed to burn its entire AI coding budget through 2026 by April of this year, vividly illustrates the scale of the problem.
"I think you can imagine—at least a year or two out—the burn rate of a strong engineer being as much as their salary or their total cost to the company."
As Mosseri explained, token costs are no longer some abstract cloud fee; they are direct operating expenses (OpEx) that now sit alongside salaries and hardware procurement. For engineering departments, this necessitates a shift to a "token-plus-salary" metric, which will become the baseline for calculating the true ROI of every developer.
From innovation to rationing
This new reality dictates treating tokens as physical assets, similar to RAM or disk space. Mosseri admits that while Meta does not currently have personal limits, he views their introduction as a "healthy necessity." Under this model, an engineer's budget for model usage will be directly proportional to the company's trust in their ability to convert those costs into profit. The era when a junior developer could "hallucinate" on the company’s dime at a rate of thousands of dollars per hour is ending.
Scarcity logic is already reshaping the market. Microsoft recently cut off its engineers' access to Claude Code licenses, forcing a migration to its own Copilot CLI to rein in ballooning costs. Although Mosseri hopes that price wars between model providers will eventually temper these costs, the immediate future belongs to those who can squeeze the most out of a minimal context window.
Meta's pivot from "unlimited experiments" to "token caps" proves that the most valuable skill in the AI stack is no longer prompt engineering, but the discipline to refrain from it. The industry spent years teaching engineers to "think in AI"; now it will have to spend just as long teaching them how to pay for it. Firms without Meta-sized capital will likely have to find salvation in local inference for small models—otherwise, the budget will run out long before the first working build.
The cost of hiring an AI-equipped engineer is set to double as compute costs match salaries. Big Tech companies are shifting from R&D freedom to strict token rationing and OpEx monitoring. Local inference and small models are becoming essential for companies that cannot afford billion-dollar token bills.