Uber burned through its entire 2026 AI budget by April. Microsoft revoked developers' Claude Code licenses after six months. One company reportedly racked up a $500 million Claude bill in a single month. The pattern is identical everywhere: per-token prices collapsed 98%, yet enterprise AI bills tripled.
This isn't a billing error. It's the token explosion—and every CIO needs to understand why it's happening before the next budget cycle.
The Numbers Don't Add Up (Until They Do)
GPT-4-equivalent performance now costs roughly $0.40 per million tokens, down from $20 per million in late 2022. That's a 98% reduction in unit cost. Yet according to multiple industry analyses, enterprise AI bills have risen by an estimated 320%. The average enterprise AI budget has grown from $1.2 million per year in 2024 to $7 million in 2026.
For CFOs: Your AI spend isn't increasing because vendors raised prices. It's increasing because consumption exploded faster than prices fell.
For CIOs: The same engineers who were cost-effective in 2024 are now burning 18.6 times more tokens in 2026, according to Nicholas Arcolano, head of research at engineering management platform Jellyfish.
The culprit is agentic AI—systems capable of independently completing complex tasks through orchestrated multi-step workflows. A simple linear workflow in 2023 cost about $0.04 per interaction. An orchestrated agentic system in 2026 costs roughly $1.20 per interaction. That's 30 times more expensive for what looks like the same output.
Real-World Budget Disasters
The data shows this isn't theoretical:
Uber: Blew through its entire 2026 AI coding budget by April, forcing immediate spending caps across engineering teams.
Microsoft: Revoked developers' Claude Code licenses six months after rolling them out. Individual engineers were reportedly spending between $500 and $2,000 per month on tokens before the licenses were pulled.
AT&T: Internal AI systems now consume 27 billion tokens per day, up from 1 billion eighteen months ago. That's a 27x increase in raw consumption.
Anonymous enterprise: One company reportedly ran up a $500 million Claude bill in a single month after forgetting to set usage limits on an agentic workflow.
Priceline: A routine Cursor contract renewal came back four to five times more expensive than the previous year, according to senior director of IT finance Chris Reed.
J.R. Storment, executive director of the FinOps Foundation, described the shift bluntly: "In April and May, I started hearing from companies: 'Oh my god, we are 3x over our entire 2026 token budget and it's only April.' The whole conversation shifted from tokenmaxxing and 'go fast' to 'we need guardrails, how do we control this?'"
Why Agentic AI Changed Everything
Traditional AI workflows were straightforward: user sends prompt, model generates response, user receives output. One call, predictable token count.
Agentic AI doesn't work that way. These systems chain multiple calls together:
- Initial prompt processing (tokens consumed)
- Document retrieval from vector databases (tokens consumed)
- Multi-step reasoning loops (tokens consumed at each step)
- Error correction and retry logic (tokens consumed)
- Final output generation (tokens consumed)
Each step burns tokens. When you deploy an agent that autonomously completes tasks, you're not paying for one API call—you're paying for dozens or hundreds of calls per task.
For CTOs: This is why your engineers using agentic tools are 2x more productive but spending 10x the tokens. Jellyfish's data shows the productivity gains are real, but the cost-per-output-unit exploded.
For CFOs: The ROI question becomes: are those productivity gains worth 10x the token spend? Most companies still can't answer that question because they lack the infrastructure to measure business value of shipped code.
The Tokenomics Crisis
Alexander Embiricos, OpenAI's head of enterprise, told TechCrunch the conversation with customers has completely shifted: "Six months ago, I would have a conversation with a customer and it would be all about 'What can it do? Is it good enough?' Now the conversations are about, 'We're spending so much. What visibility do you have? What token controls do you have?'"
This is the tokenomics crisis: enterprises adopted agentic AI when vendors offered all-you-can-eat subscriptions in early 2025. Then the vendors realized developers were burning thousands of dollars in compute on $200-per-month plans.
The result? Every major AI vendor is converging on metered pricing:
- Anthropic eliminated flat-rate enterprise pricing after discovering the consumption gap
- OpenAI moved Codex to per-token billing the same month
- Microsoft is steering enterprise customers toward Azure OpenAI Service with transparent token tracking
- Google launched Gemini with usage-based pricing from day one
For procurement teams: The era of predictable AI budgets is over unless you negotiate reserved capacity deals or find vendors who absorb token-volume risk.
The Linux Foundation's Answer
Against this backdrop, the Linux Foundation unveiled plans for the Tokenomics Foundation, a new standards body aiming to bring the same cost discipline to AI tokens that FinOps brought to cloud spending.
The Foundation plans to build:
- Canonical definition of "tokenomics" (standardized cost accounting for AI)
- Open standards for AI token usage and billing (comparable across vendors)
- New metrics including cost-per-intelligence and tokens-per-watt
A formal launch is planned for July 2026. Nishant Gupta, chief availability officer at Salesforce, said in a statement: "Token economics is fundamentally more abstract and opaque than anything we've managed at this scale before."
The challenge is enormous. J.R. Storment explained: "Tracking cloud costs is a hundreds-of-millions-of-rows-a-month data problem. Tracking token costs is a trillions-of-rows-a-month data problem."
What Smart CIOs Are Doing Now
The enterprises that will win the agentic AI era are solving cost predictability before the next budget cycle forces the conversation. Here's what's working:
1. Model Routing (Primary Cost Lever)
Factory, an enterprise AI coding startup, launched a model router that automatically picks the cheapest adequate model for each task. The logic is simple: not every task needs GPT-5.1. Many can run on Sonnet or Haiku at 10x lower cost.
Vitaly Gordon, CEO of Faros AI, says frontier labs are already doing this internally: "The financial report for how much you spend on Anthropic, even if you call the Opus model, some of the spend will be on Sonnet or Haiku, because they are smart enough to do it."
Action: Implement intelligent model routing before deploying agentic workflows at scale.
2. Reserved Capacity Deals
Several CIOs are negotiating reserved throughput agreements with cloud providers—fixed capacity commitment in exchange for predictable monthly cost. This transfers forecasting risk to the enterprise, but it also caps the downside.
Action: If your token consumption is predictable (it rarely is with agentic AI), negotiate reserved capacity at a discount.
3. Fixed-Price Infrastructure Providers
A growing category of infrastructure providers absorbs token-volume risk and converts it into predictable monthly costs. The mechanics vary—traffic shaping, intelligent caching, model routing, deep capacity planning—but the structural commitment is the same: per-token volatility stops at their layer, not yours.
Action: Add at least one fixed-price AI infrastructure provider to your supply base as a hedge against metered pricing volatility.
4. Token-Level Observability
Platforms like Datadog, New Relic, Pay-i, and Jellyfish now provide token-level spend tracking, budget alerts, and real-time cost breakdowns by model, team, and API key.
Action: Deploy token observability before your first agentic workflow hits production. Visibility won't reduce the bill, but it prevents $500 million surprises.
5. ROI Measurement Infrastructure
Jellyfish's data shows engineers using the most tokens are 2x more productive but spend 10x the tokens. The question is whether that productivity translates to business value—and most companies can't measure it yet.
Action: Build the infrastructure to measure business outcomes of shipped code, not just lines of code or velocity metrics.
The Forecast: It Gets Worse
Goldman Sachs projects global token usage will multiply 24 times by 2030. If you think budgets are unpredictable now, wait until agentic AI is embedded in every business function.
The companies already over budget need solutions now. The Tokenomics Foundation's first deliverable is still months away. In the meantime, CIOs are left negotiating with vendors who control all the variables.
As Vitaly Gordon put it: "Maybe we created a steam engine, but we still haven't figured out the assembly line."
Bottom Line
Per-token prices fell 98%. Enterprise AI budgets tripled. Agentic AI consumption exploded 18.6x. This isn't a pricing problem—it's a consumption problem. And consumption is driven by how you deploy AI, not what the vendor charges per token.
The enterprises that survive the token explosion will be the ones who solve three things before the next budget cycle:
- Visibility: Token-level observability across all models and workflows
- Optionality: At least one non-metered AI infrastructure provider as leverage
- ROI discipline: Infrastructure to measure business value, not just productivity
If you're still running pure metered relationships with every AI vendor, hoping per-token prices keep falling faster than consumption rises, you're playing a game where the other side controls all the variables.
Uber and AT&T learned this the hard way. The rest of the Fortune 500 is next.
