Meta’s AI Token Spending Could Hit Tens of Billions With No ROI Metric
What You Need to Know
- Meta discovered employees burning through AI tokens at rate costing tens of billions annually by 2026.
- Amazon, Uber, and Meta created incentive structures optimizing for token consumption rather than measurable business output.
- Only 26% of companies have comprehensive visibility into their AI costs, discovering expenses after the fact.
- Meta building centralized “AI Gateway” with spending caps and usage dashboards to control internal AI spending.
Meta is building a centralized control layer for internal AI spending after discovering its employees were burning through tokens at a rate likely to cost the company tens of billions of dollars in 2026 alone. The shift is less about reining in AI and more about the fact that no one, including Meta, has figured out how to measure what that spending actually produces.
The pattern here is consistent enough to be a trend. Amazon pulled down an internal token-use leaderboard in late May after it appeared to be driving consumption for its own sake. Uber exhausted its entire planned 2026 AI coding budget by April. Meta itself ran a program called “Claudeonomics” to encourage employees to maximize AI usage, then quietly discontinued it. What these companies built were incentive structures that optimized for token consumption rather than output, and the bills arrived before anyone had a methodology to justify them. A KPMG survey found that only 26% of companies have comprehensive visibility into their AI costs, which means the majority of enterprises are discovering their exposure after the fact.
The Uber COO’s comment is the most honest thing anyone in a senior role has said publicly about this: the link between token spending and measurable output simply does not exist yet.
Meta’s proposed fix, a centralized “AI Gateway” with usage dashboards, spending caps, and automatic alerts for unusual spikes, is essentially an enterprise IT governance layer bolted onto a technology its own internal culture was encouraged to treat as unlimited. The company also plans to steer engineers away from third-party tools like Claude toward its own MetaCode assistant, which consolidates cost control and competitive positioning in one move. For Anthropic and other API-dependent AI providers, the downstream risk is real: if large enterprise customers begin capping consumption and migrating to proprietary alternatives, the revenue trajectory that justifies current AI infrastructure valuations gets harder to defend. Goldman Sachs projects a 24-times increase in token consumption by 2030 driven by agentic AI, but that projection assumes demand is allowed to scale, not governed.
Meta’s AI Gateway is expected to be fully implemented by 2027, with initial spending controls rolling out in the coming weeks alongside the push toward MetaCode adoption.
0 Comments