The token bill comes due.

Token prices collapsed 98% while enterprise bills tripled.

The standard narrative about AI costs goes like this: prices are falling fast, inference is getting cheaper, and the democratization of intelligence is accelerating. The data supports it. Since OpenAI launched GPT-4 in March 2023 at $60 per million output tokens, the price index for frontier AI APIs has dropped roughly 94%. GPT-4o mini runs at $0.60 per million output tokens. DeepSeek V3 charges $1.10. Google's Gemini Flash costs even less.

And yet enterprise AI spending is surging in the wrong direction. Average enterprise AI bills have roughly tripled over the past 18 months, even as the unit cost of tokens plummeted. BCG data shows companies plan to roughly double their AI spending as a share of revenue in 2026, from 0.8% to 1.7%. The math is simple but the implications are severe: consumption is growing faster than prices are falling.

The culprit is a fundamental shift in how AI is consumed. The chatbot era involved short, bounded interactions: a user asks a question, gets an answer, moves on. The agent era involves autonomous systems that chain dozens or hundreds of API calls per task, maintain persistent context windows, retry on failure, and operate continuously. A single agentic coding session on Claude Code or OpenAI Codex can consume more tokens in an hour than a developer used in a month of chat-based interactions in 2024.

So what: the industry built its adoption curve on the assumption that costs would keep falling. For frontier capability, they stopped falling in April 2026 and started climbing.

Five cost disasters broke into public view within two weeks of each other.

The abstract economics became concrete in late May 2026, when a series of corporate AI cost disasters went public almost simultaneously. Together, they painted a picture of an industry that had adopted powerful tools without building the financial controls to contain them.

The Uber case is particularly instructive because Uber is not a company experimenting with AI on the margins. Its core business runs on AI: dynamic pricing, routing, demand prediction, driver matching. If any company should be able to demonstrate AI ROI, it is Uber. That its COO publicly questioned the spend signals something deeper than a budgeting failure. It suggests the cost-value equation for frontier AI tools is broken at a structural level.

So what: the $500M Claude bill was the headline, but the systemic pattern is more important. Companies across sectors deployed AI tools without spend controls, and the bills arrived before the ROI did.

Frontier model pricing reversed in April 2026 and the mechanisms are getting less transparent.

The AI cost story is not one story. It is two stories running in opposite directions. Story one: commodity models are getting radically cheaper. GPT-4o mini, Claude Haiku, Gemini Flash, DeepSeek V3, and dozens of open-source alternatives have driven the cost of basic inference toward zero. For straightforward classification, summarization, and retrieval tasks, token costs in mid-2026 are roughly 1% of what they were in early 2024.

Story two: frontier models are getting more expensive, and the pricing mechanisms are becoming opaque. OpenAI's GPT-5.5, released in April 2026, doubled the standard list price from GPT-5.4: $5 per million input tokens (from $2.50) and $30 per million output tokens (from $15). The Pro tier jumped to $30 input and $180 output, a 12x increase. For a company running 50 million input and 10 million output tokens per month, annual costs jumped from roughly $275,000 on GPT-5.4 to $510,000 on GPT-5.5 standard, and to $1.56 million on the Pro tier.

Anthropic's approach was subtler but arguably more consequential. Claude Opus 4.7 kept the same published per-token rate as its predecessor, but shipped with a new tokenizer that inflates token counts by 32% to 45% for the same input text. Analysis by Pasquale Pillitteri found that real-world costs increased 12% to 27% depending on prompt length, with the largest increases hitting shorter prompts. The list price did not change. The effective price did.

So what: the era of reliably declining AI costs ended in April 2026. Frontier capability now costs more per task than it did six months ago, and the pricing mechanisms are becoming opaque enough that most enterprises cannot accurately forecast their AI spend.

Tokenmaxxing: how agentic AI turned a price collapse into a spending crisis.

A new term entered the enterprise vocabulary in 2026: tokenmaxxing. Coined in developer communities and picked up by outlets from The Observer to The Indian Express, it describes the pattern of maximizing AI token consumption without regard for whether the consumption produces proportional value. ServiceNow's customer chief warned in May that tokenmaxxing 'is an AI hype cycle' rather than a productivity strategy.

The pattern repeats across companies. Engineering teams adopt Claude Code or OpenAI Codex. Developers discover that agentic tools can generate, refactor, and test code autonomously, and begin routing an increasing share of their work through them. Usage grows exponentially because the tools are genuinely useful, because there is social pressure to adopt, and because in many organizations there are no spend limits or visibility tools in place.

Nvidia reported that Codex runs on its GB200 NVL72 systems deliver '35x lower cost per million tokens' and '50x higher token output per second per megawatt.' But the total volume of tokens consumed by its 10,000+ employees using Codex dwarfs anything the company consumed when the tool was a chatbot. Codex reached 4 million weekly users in April 2026, up from 3 million two weeks earlier. OpenAI acknowledged it was actively resetting user usage limits every time the user base grew by a million to manage scale.

So what: tokenmaxxing is not a spending discipline failure. It is a structural consequence of deploying consumption-based tools without consumption-based controls. The tools work. The governance does not.

Consumer AI is priced below cost and the gap is widening.

The cost crisis extends beyond enterprise APIs into the consumer AI business model itself. Multiple analyses have converged on a troubling conclusion: consumer AI subscriptions are priced below cost, and the gap is widening as models grow more capable and users grow more demanding.

One analysis estimated that the breakeven price for OpenAI's current consumer offering, accounting for heavy users and frontier model access, would be approximately $2,000 per month. The actual price of ChatGPT Pro is $200. OpenAI is projected to lose $14 billion by the end of 2026, even as its valuation has soared past $500 billion. The subscription price ladder tells the story: ChatGPT Plus at $20/month, Team at $100/month, Pro at $200/month. The best models and longest context windows are increasingly gated behind higher tiers.

The broader market dynamics compound the pressure. Training runs for frontier models now require on the order of 10^24 to 10^25 FLOPs, attention mechanisms scale quadratically with context length, and the capital expenditure required to stay competitive is consuming an outsized share of operating cash flow. Microsoft's AI-related capex hit nearly $35 billion in a single quarter. Analysts warned that AI capex could consume as high as 94% of operating cash flow in 2025-26. Microsoft also disclosed a $4.1 billion earnings hit tied to OpenAI, a roughly 490% increase from the prior year.

So what: consumer AI cannot scale to billions of users at current pricing without either radical efficiency gains or radical price increases. The market is betting on the former. The data increasingly supports the latter.

95% of enterprise AI pilots produce no measurable P&L impact.

Behind the cost headlines sits a harder question: is any of this spending generating returns? MIT Media Lab's Project NANDA found that 95% of enterprise AI pilots produced no measurable P&L impact. Gartner reported that only 1 in 50 AI investments deliver transformational value, and only 1 in 5 deliver measurable ROI of any kind. These figures predate the agentic era, which has multiplied both the cost and the complexity of measuring returns.

The Littler Annual Employer Survey found that only 6% of workplaces said they were not using AI in 2026, and AI-related policy change expectations among employers surged from 42% in 2025 to 84% in 2026. Adoption is no longer the bottleneck. The bottleneck is proving that the adoption creates value proportional to its cost. The Stanford AI Index reported 55% of companies now have at least one AI use case in production, and PwC's Global CEO Survey found one-third of CEOs have seen concrete results. That still leaves two-thirds who have not.

So what: the spending is accelerating into a ROI vacuum. Companies are doubling budgets for a technology where 95% of pilots produce no measurable financial return. When the correction comes, it will be sharp.

The response: standards bodies, private infrastructure, and the new AI FinOps.

The cost crisis has catalyzed three categories of response. The first is institutional standardization: the Linux Foundation launched the Tokenomics Foundation in June 2026, modeled on the FinOps Foundation that standardized cloud cost management. Its mandate is to create common standards for token usage measurement, billing transparency, cost allocation, and audit. The fundamental visibility problem is that most companies cannot tell you how many tokens they consumed last month, which teams consumed them, or what value was produced.

The second is infrastructure alternatives. Companies are moving inference workloads in-house. Dell and others are building on-premises AI infrastructure that gives enterprises fixed-cost token generation rather than variable API billing. Google's internal TPU pods have historically achieved 40-60% lower cost-per-token than equivalent cloud workloads. Nvidia's Vera Rubin platform, now shipping first customer samples, claims roughly 10x better performance-per-watt than Grace Blackwell. Google's TPU 8i, announced at Cloud Next 2026, includes 384MB of onboard SRAM designed specifically for agentic workloads.

The third is operational optimization. A growing ecosystem of startups and open-source tools enables multi-model routing, where each AI task goes to the cheapest capable model rather than defaulting to the frontier. One analysis showed that a multi-model agent architecture can cut costs by roughly 50% versus using a single high-end model for every step, by reserving frontier models for tasks that actually require frontier capability. Companies are also retroactively building what should have existed from day one: per-user token budgets, team-level spend caps, usage dashboards, and approval workflows for high-consumption tasks.

So what: the correction infrastructure is forming, but the cloud computing industry took roughly five years to build mature FinOps practices. AI cost management is starting that clock now, with the added complexity that token pricing is less predictable than instance-based cloud billing.

Three dynamics will determine whether the cost crisis resolves or triggers a market correction.

First, the hardware cycle. Nvidia's Vera Rubin, Google's TPU 8i, and Baidu's Kunlun M100 all promise step-function improvements in inference efficiency, but none will reach volume deployment before late 2026 at the earliest. The cost relief from new silicon is real but 12-18 months away from material impact on enterprise bills.

Second, the governance cycle. The Tokenomics Foundation, enterprise FinOps tools for AI, and company-level spend controls are all in their infancy. The gap between 'we have AI tools' and 'we can measure what our AI tools cost and produce' is where the next 12 months of enterprise pain will concentrate.

Third, the market correction cycle. If enterprise AI spending contracts because CFOs cannot justify the ROI, the consequences cascade: AI providers lose revenue, infrastructure demand softens, and the capital expenditure cycle that currently sustains chip demand and data center construction decelerates. One analysis warned of a 'Bullwhip Effect' where failing AI companies dump used GPUs onto the secondary market, collapsing demand for new hardware and unwinding the compute boom that sustains Nvidia's $5 trillion valuation. The DeepSeek moment in early 2025, when a one-day $600 billion hit to Nvidia's market value proved the market's sensitivity to cost assumptions, showed what a repricing looks like.

So what: the companies that survive the AI cost crisis will be the ones that treated cost management as a first-class engineering and governance problem from the start, not the ones that hoped the price curve would bail them out. The price curve has reversed. The bill is here.