Uber, Amazon, JPMorgan, and Meta Are All Cutting AI Token Budgets. The Enterprise Spending Backlash Has Arrived.

Uber burned through its entire 2026 AI tool budget by April. The company responded by capping monthly token spending at $1,500 per employee, according to Bloomberg. Uber is not alone. Across the enterprise landscape, the token-spending free-for-all that defined early 2026 is over.

Amazon told employees to stop using AI tools without clear business justification, according to the Financial Times. JPMorgan circulated internal memos on excessive AI spend after some employees ran up AI bills larger than their own salaries, per Yahoo Finance. Meta shifted from what Alberto Romero, writing in The Algorithmic Bridge, calls “token-maximizing” (rewarding engineers who consumed the most tokens) to “token-minimizing” (eliminating waste), according to The Information. Microsoft canceled its internal Claude Code licenses entirely after engineers in its Experiences and Devices group burned past the annual AI budget months ahead of schedule, shifting developers to its own Copilot CLI by June 30, per Forbes.

The pattern is consistent: explosive, uncontrolled adoption through early 2026, followed by a sharp correction as CFOs examined the bills.

The Scale of the Problem

The enterprise AI spending binge was not subtle. Engineers at multiple companies ran autonomous agent loops competing on internal leaderboards, driving massive token consumption with no direct business output, according to Romero’s analysis. The dynamic resembled what organizational theorist Steven Kerr described decades ago as “rewarding A while hoping for B”: companies incentivized AI usage (A) while hoping for productivity gains (B).

The bill arrived simultaneously across sectors. The Information reported that customers are actively cutting their OpenAI and Anthropic bills because the costs are unsustainable. This is happening even as both companies report record revenue growth. Anthropic and OpenAI together capture approximately 89% of the entire AI startup ecosystem’s revenue, per The Information. Epoch AI data shows annualized revenue curves for both companies that are nearly vertical.

That concentration creates a fragile picture. If enterprise cost discipline causes significant customer churn at the top, the entire AI startup revenue base contracts with it.

Why Agents Made It Worse

The shift from chatbots to autonomous agents accelerated the spending crisis. A chatbot query costs a fixed amount of tokens. An agent loop can run indefinitely, consuming tokens on each iteration as it reasons, acts, observes, and reasons again. An engineer using Claude Code to refactor a codebase might burn 50,000 tokens. An autonomous agent retrying failed approaches across a complex repository can burn millions.

This is the economic structure that caught enterprises off guard. AI tools that billed per seat (like GitHub Copilot at $19/month) created predictable costs. Token-based billing on agent workflows created open-ended exposure. When Uber’s engineers adopted Claude Code enthusiastically, nobody had modeled what happens when hundreds of engineers run agent sessions simultaneously, all day, for months. The answer: you exhaust the annual budget in four months.

The same dynamic hit Microsoft internally. Forbes reported that Microsoft let thousands of engineers use Claude Code on the company’s expense in December 2025. Within its Experiences and Devices group, Claude Code became the preferred tool. Then the bill arrived, and it had run past the annual AI budget months ahead of schedule. Microsoft’s response was to build its own coding model and move engineers off Claude Code by June 30. Cost pulled the trigger, but what cost exposed was the structural problem: renting intelligence by the token at enterprise scale is not sustainable without strict governance.

The Governance Response

The correction is not just budget cuts. It is creating a new category of enterprise infrastructure: agent spending governance. Runlayer, which raised $30M in Series A funding from Felicis and Khosla Ventures on June 24, is building exactly this: a centralized control layer for deploying and managing AI agents across organizations, per AlleyWatch.

The governance stack that enterprises now need includes token budgets per user and per project, approval workflows for agent deployments, real-time spending dashboards, and kill switches for runaway loops. Six months ago, none of this infrastructure existed because nobody needed it. Now every large enterprise that adopted agentic AI tools needs it.

This is the predictable second act of any enterprise technology adoption cycle. The first act is excitement and uncontrolled adoption. The second act is the CFO asking why the bill tripled. The third act is the governance and procurement infrastructure that makes the technology sustainable. Enterprise AI just entered act two.

The Revenue Quality Question

Romero’s analysis raises a deeper question about what the record revenue numbers at OpenAI and Anthropic actually represent. Run-rate revenue (monthly revenue multiplied by 12) can be misleading when a significant fraction of that revenue comes from API customers who can cut usage frictionlessly. Unlike SaaS subscriptions with annual contracts, API billing can drop to zero overnight.

Romero identifies three risk factors in the current revenue structure. First, large customers like Microsoft can and do shut off access abruptly, as the Claude Code cancellation demonstrates. Second, many customers are trying Claude Code and Codex seriously for the first time, meaning a portion of current revenue is trial-period spending. Third, a meaningful fraction of Anthropic and OpenAI’s revenue comes from other AI companies building on their APIs, creating circular dependency: if the broader AI market contracts, the revenue base contracts with it.

The question is whether the “honeymoon revenue,” as Romero calls it, converts into stable, recurring spending. The enterprise backlash suggests that at least some of it will not. Uber is not abandoning AI tools. It is capping them at $1,500 per month. That is still significant spending, but it is a fraction of what uncontrolled usage produced.

The Reliability Gap

Cost is not the only problem. Enterprises are also discovering that the cost-benefit analysis for current AI tools yields unclear results in many production scenarios. Romero points to a structural limitation he calls “jaggedness”: models that can disprove a century-old mathematics conjecture but score 0.43% on ARC-AGI 3, a pattern-recognition test designed for children.

For enterprises, jaggedness means AI tools that produce spectacular results on some tasks and unreliable results on others, with no clear way to predict which category a given task falls into. When AI bills were small enough to ignore, this inconsistency was tolerable. When they exceed employee salaries, the ROI calculation becomes urgent.

What Comes Next

The spending backlash does not mean enterprise AI adoption is declining. It means it is maturing. The companies cutting budgets are not abandoning AI. They are implementing the controls that every enterprise technology eventually requires: procurement workflows, usage policies, cost attribution, and spending limits.

For AI startups, the implications are structural. Token-based pricing, which drove record revenue growth in early 2026, will face increasing downward pressure as enterprises optimize usage. Agent platforms that cannot demonstrate clear ROI on a per-task basis will lose budget allocation. Governance and cost-management tooling, which barely existed six months ago, is becoming a prerequisite for enterprise sales.

The venture capital flowing into vertical AI agent startups (JustAI’s $17M Series A for agentic martech, Probook’s $34M from a16z and Sequoia for home services automation) suggests investors are betting that focused, industry-specific agent deployments can clear the ROI bar that horizontal agent tools are failing. Infosys disclosed approximately $1 billion in annualized AI services revenue at its Annual General Meeting on June 24, signaling that enterprise demand remains massive even as cost discipline tightens, per HDFC Sky.

The AI industry’s first real stress test is not a technology failure. It is a spreadsheet. And the spreadsheet says the current spending trajectory is not sustainable without better tools, better governance, and better proof that the agents are worth what they cost.