Enterprises are blowing through their 2026 AI budgets months ahead of schedule, with the FinOps Foundation reporting that some companies are running 3x over their full-year token spend as of April. Uber burned through its entire 2026 AI coding budget by that same month. Microsoft revoked Claude Code licenses from its developers months after enabling them. And one unnamed company landed a $500 million bill from Anthropic after forgetting to set employee usage limits.
The shock has rewired enterprise sales conversations. Per-token prices have actually fallen, but agentic features released in November — including Claude Opus 4.5, GPT-5.1, and Gemini 3 Pro — have driven consumption sharply higher. Autonomous agents chew through tokens at a rate that flat per-token discounts can't offset.
Priceline saw a routine Cursor contract renewal come back 4-5x more expensive, according to Chris Reed, the company's senior director of IT finance. Priceline has now started placing token limits on specific employee groups. The pattern — heavy adoption in early 2025, sticker shock in 2026 — has spread broadly enough that the Linux Foundation unveiled the Tokenomics Foundation this week, a new standards body modeled on the FinOps Foundation's role in cloud cost discipline.
Key facts
- 01Uber exhausted its entire 2026 AI coding budget by April, and one unnamed company racked up a $500M Claude bill after failing to set usage limits.
- 02FinOps Foundation reports companies running 3x over their full-year 2026 token budgets as of April; Priceline saw a Cursor renewal come back 4-5x higher.
- 03Per-developer token consumption rose 18.6x in nine months, per Jellyfish; top users were 2x more productive but spent 10x the tokens to get there.
- 04The Linux Foundation is launching the Tokenomics Foundation in July 2026 to standardize AI cost metrics, with 180 vendors already in the FinOps Foundation.
- 05Goldman Sachs projects global token usage to multiply 24x by 2030.
J.R. Storment, who runs the FinOps Foundation, said the tone changed sharply in spring. Companies that had spent 2025 pushing engineers to use the best models regardless of cost started reporting what he called existential crises. The conversation shifted from tokenmaxxing and going fast to building guardrails.
The productivity math is genuinely murky. A March survey from Faros AI covering 20,000 developers found output rising, but bugs and rewrites rising alongside it. Jellyfish found that engineers using the most tokens were roughly twice as productive as lighter users — but spent 10x the tokens to get there. Per-developer token consumption rose 18.6x in nine months, according to Jellyfish research head Nicholas Arcolano.
Vitaly Gordon, CEO of Faros AI, said a CTO recently told him that one engineer had spent $40,000 on tokens in a single month, and the CTO genuinely didn't know whether to stop the engineer or hold him up as a model for everyone else. Arcolano framed the better answer as moving the broad middle of developers from low to moderate usage rather than pushing heavy users higher — the ROI math favors breadth over depth.
Part of the problem is scale. Cloud cost tracking is a hundreds-of-millions-of-rows-a-month data problem, Storment said. Token cost tracking is a trillions-of-rows-a-month data problem. The accounting systems most enterprises built for cloud don't survive the jump.
Reed at Priceline said he's already seeing discrepancies between vendor-reported usage and internal data — the kind of billing-audit problem that defined the early years of telecom and cloud expense management. A market is forming to address it. Pure-play startups like Pay-i and Paid track token costs and tie them to value-based billing. Jellyfish, Waydev, and Faros AI sell AI agent monitoring aimed at proving developer ROI. Ramp has moved into AI spend management. Datadog and New Relic have added token-level observability and GPU monitoring. AWS is expected to introduce new AI financial management features at the FinOps X conference next week.
Storment said most of the roughly 180 vendors inside the FinOps Foundation are leaning toward this space. Tiffany Luck, a partner at NEA, expects token efficiency to be built into the harness or application layer — she pointed to Factory, which this week launched a model router that automatically picks the cheapest adequate model for each task. Gordon expects frontier labs themselves to adopt OpenRouter-style routing internally, sending queries to Sonnet or Haiku even when a customer calls Opus, because the smaller models are often good enough.
None of this works without a common vocabulary. The Tokenomics Foundation is building open standards for AI token billing, plus new metrics like cost-per-intelligence and tokens-per-watt. Nishant Gupta, chief availability officer at Salesforce, said token economics is more abstract and opaque than anything enterprises managed at cloud scale, and requires operational muscles the industry hasn't built yet. The foundation launches formally in July.
The skeptic's case is that the standards body arrives months after the budgets have already blown. Goldman Sachs projects global token usage to multiply 24x by 2030, meaning the spend problem gets harder, not easier, before any standard takes hold. Arcolano's own data — that the heaviest spenders are only 2x as productive for 10x the cost — suggests a lot of the current spend isn't paying for itself, and most companies still can't measure whether shipped code actually drove revenue. Gordon put it more bluntly: the industry built a steam engine but hasn't figured out the assembly line.
The interesting market signal here is that AI's cost-control layer is now where AI's product layer was two years ago — a land rush with no standards, fragmented vendors, and enterprise buyers who will pay almost anything to get visibility. Whoever wins the FinOps-for-tokens category captures a structural choke point on every enterprise AI bill, which means the next wave of AI infrastructure value may accrue not to the labs but to the companies metering them. Anthropic and OpenAI should be paying attention: the same enterprises that signed blank-check contracts in 2025 are about to demand itemized invoices.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
