OpenAI clients burn 100 billion tokens monthly as costs surge

Enterprise AI spending has shifted from an afterthought to a budget crisis in just six months, OpenAI CEO Sam Altman acknowledged.

The cost of running artificial intelligence has become a "huge issue" for enterprise clients, OpenAI Chief Executive Officer Sam Altman said, as token consumption surges 1 million-fold from levels just six years ago.

"Cost was never brought up six months ago," Altman said at an OpenAI enterprise event on June 2. "Now it's a huge issue." The CEO disclosed that OpenAI's single largest customer now consumes about 100 billion tokens monthly — roughly 75 billion words — compared with 100,000 tokens for the top user six and a half years ago.

The explosion in usage has exposed a structural problem: OpenAI spends $1.35 for every $1 of revenue, with losses driven primarily by inference costs rather than model training. Uber Technologies Inc. exhausted its entire 2026 AI budget in the first four months of the year, forcing hard token caps, while individual engineers at the ride-hailing company racked up monthly AI bills ranging from $150 to $2,000. Amazon.com Inc. has closed internal token leaderboards to discourage runaway consumption.

The shift to token-based billing by OpenAI and Anthropic in the first quarter of 2026 has turned a previously opaque cost line into a measurable, per-task expense — and the early results are alarming corporate finance teams. Gartner projects AI agent software spending will reach $207 billion in 2026, up 139% from 2025, but that trajectory assumes enterprises continue expanding AI spend. The Uber signal, and a pattern of companies quietly pulling back token consumption, suggest the trajectory is under pressure.

The Token Trap

The root of the cost crisis lies in the industry's pricing structure. For most of the generative AI era, flat-fee subscriptions absorbed unlimited token burn, making the actual cost of any given task invisible. When Anthropic and OpenAI moved enterprise customers to usage-based billing in Q1 2026, the hidden costs became suddenly legible. One Anthropic enterprise customer accidentally spent $500 million in a single month after failing to set spend limits.

The problem has two layers. First, output quality remains unpredictable — large language models hallucinate, loop and fail in ways that are difficult to anticipate, and every failed run costs tokens regardless of outcome. Second, there is no standard unit for measuring the cost of an AI task, because the same task can consume wildly different token counts depending on the prompt, model version, context window and whether the agent makes wrong turns.

GitHub Copilot's move to token-based billing in June 2026 provided the clearest retail-level evidence yet. Users on the promotional tier reported burning 30% to 60% of monthly credits in a handful of prompts. One user said Copilot went from their favorite subscription to their most stressful overnight.

The ROI Reckoning

Uber's experience illustrates the broader challenge. Chief Operating Officer Andrew Macdonald acknowledged at a May 25 conference that despite 95% of engineers using AI tools monthly, he could not draw a line between that token spend and meaningful consumer-facing product improvements. "That link is not there yet," Macdonald said.

Microsoft Corp., facing Claude Code bills running $500 to $2,000 per engineer monthly, began canceling direct Claude Code licenses and routing engineers back to GitHub Copilot.

Anthropic Chief Executive Officer Dario Amodei has acknowledged the timing risk explicitly. In a February interview, he warned that if AI revenue growth forecasts are off by even a year, "then you go bankrupt." He was referring to Anthropic's own infrastructure bets, but the logic applies to enterprise customers too. If token-based billing reveals that productivity gains do not justify the cost, enterprises do not go bankrupt — they just stop renewing.

For investors, the token billing transition is the first real price discovery mechanism the AI industry has produced. Flat-fee subscriptions created convenient optics: costs were low, adoption was high, and return on investment was a question to be addressed later. Usage-based billing has changed that calculus overnight. Companies that can measure and demonstrate AI return on investment will determine whether the current capital stack holds. Those that cannot will be the first to renegotiate and rethink.

Nvidia Corp., whose graphics processing units power the majority of AI training and inference, faces a potential demand-side shock if enterprise customers universally cap spending. The company's data center revenue has grown more than 200% year-over-year for five consecutive quarters, but that growth assumes ever-expanding token consumption. A sustained pullback in enterprise AI budgets could compress that trajectory.

This article is for informational purposes only and does not constitute investment advice.