One Hundred Billion Tokens and No CFO Knows What They Bought
Sam Altman took the stage at OpenAI's enterprise event on June 2, 2026, with a statistic designed to impress: the company's largest internal consumer of tokens processes around 100 billion tokens per month. The room reacted as expected. Then Altman added, almost in passing, that this number is not the world record, because someone outside of OpenAI consumes even more. And there, without quite intending to, he described with precision the problem that is fracturing the economics of artificial intelligence at corporate scale: consumption grew so fast that it outpaced both the imagination of those selling the product and the budgeting capacity of those buying it.
What came after the data point was more revealing than the data point itself. Altman admitted that costs are now the second most frequent complaint from OpenAI's enterprise customers. And he described a meme circulating among corporate executives with more diagnostic precision than any analyst report: "The company spent the entire 2026 budget in the first quarter. Can you make it more efficient?" The question inside the meme is not rhetorical. It is the new state of affairs for dozens of organizations that entered the year with spending assumptions based on 2025 patterns and found that agentic models consume at an entirely different velocity.
Six and a half years ago, OpenAI's most active user processed approximately 100,000 tokens per month. Today, that figure is the global average per person. The company's most active internal user consumes one million times more than that historical record. Altman projects that this expansion will repeat itself. If it does, the artificial intelligence infrastructure that exists today would be to the future market what a pocket calculator is to a data center. But between that projection and the operational reality of corporate buyers, there is a gap that no exponential growth slide deck resolves on its own.
The Budget as the First Indicator of Technological Maturity
There is a pattern in the history of enterprise technology that repeats with enough consistency to serve as a framework: every time a technology moves from an experimental tool to a line-item operational expense, the finance department enters the conversation and changes the rules. It happened with cloud software in the mid-2010s. It happened with data and advanced analytics afterward. With agentic artificial intelligence, it is happening now, in 2026, at a speed that caught off guard even the executives of the most sophisticated companies on the planet.
Uber is the most thoroughly documented case. According to what has been reported, the company exhausted its artificial intelligence budget for all of 2026 in four months. The operational response was immediate: a cap of $1,500 per employee per month for agentic programming tools, including Claude Code and Cursor. But the most significant statement did not come from the CEO but from COO Andrew Macdonald, who said publicly that Uber cannot draw a direct line between that growing token expenditure and concrete improvements for end users — neither for drivers nor for passengers. That statement is, in terms of value architecture, a first-order warning signal. Not because the spending is inherently bad, but because it indicates that the feedback loop between investment and outcome has not yet been closed.
Microsoft cancelled the majority of its internal Claude Code licences before mid-May and redirected its engineers toward GitHub Copilot CLI before the fiscal year close on June 30. The surface-level reading is that Microsoft prefers its own product. The more precise reading is that Microsoft also faced budget overruns on artificial intelligence tools and chose to consolidate spending within its own perimeter before the problem escalated. Amazon eliminated its internal token consumption leaderboard after a senior executive instructed the team to stop using artificial intelligence simply for the sake of using it. Walmart, which had offered unlimited tokens to its employees for its internal artificial intelligence agent, also imposed limits.
The pattern is not coincidence, nor isolated financial panic. It is the signal that the corporate sector has just crossed the threshold where artificial intelligence stopped being a pilot project with an innovation budget and became an operational expense that competes with other operational expenses for return-on-investment justification.
What Token Consumption Reveals About the Distribution of Value
Behind the consumption numbers lies an economic structure that deserves to be examined with precision. Every token consumed is revenue for OpenAI or for Anthropic, compute demand for cloud providers, and investment justification for chip infrastructure. From that angle, the growth of one million times over six and a half years is exactly the narrative that underpins the valuations of infrastructure companies and the logic of large data center contracts.
But that same structure has an asymmetric distribution that the consumption numbers do not capture. The companies paying for the tokens — Uber, Microsoft, Amazon, Walmart — are absorbing the operational costs of growth, while the capture of value in terms of margins, intellectual property over the models, and pricing power is concentrated in the hands of the model providers. That asymmetry is not necessarily unfair by market standards, but it does carry structural consequences for the sustainability of the mass-adoption model.
When Uber's COO says he cannot connect token expenditure to product improvements, he is describing a return-on-investment problem that, if it persists, will not be resolved with more tokens but with a renegotiation of the entire spending architecture. The fact that Anthropic has surpassed OpenAI in enterprise corporate spending, according to Altman himself, adds another layer to this analysis. It means that the competition between models is generating a proliferation of platforms that engineering teams adopt in parallel, which multiplies costs without necessarily multiplying results. The consolidation that Microsoft is executing internally — forcing the use of a single tool — is a rational response to that problem, even if it arrives wrapped in the rhetoric of product preference.
The case of Peter Steinberger, an external developer who according to reports consumed 603 billion tokens in 30 days, and the OpenAI employee who reportedly used 210 billion tokens in a single week, illustrates something different but related. When individual consumption surpasses the global average by several orders of magnitude, flat or semi-closed pricing models begin to generate cross-subsidies that appear on no balance sheet but that distort the economics of the service. Not all tokens carry the same production cost, nor the same value for the person consuming them.
The Equation That Doesn't Balance in the Mass-Adoption Model
The dominant narrative in the sector since 2023 was one of frictionless adoption: provide broad access, eliminate price barriers, scale consumption, and capture value later through dependency, data, and network effects. That playbook worked to build massive user bases. The problem is that in the enterprise segment, "dependency" has a counterweight that does not exist with the same intensity for the individual consumer: a CFO and an annual budget cycle.
Altman described the shift in attitude as something that arrived "suddenly." In early 2026, by his own account, no one cared about costs. Everyone was comfortable with their level of spending. That description, delivered by the CEO of the most valuable company in the artificial intelligence sector, is itself a diagnosis of how the adoption phase was structured: without buyers having any clarity about the cost curve they were implicitly accepting as they scaled agentic usage.
Agentic models, unlike point-query chatbots, have a characteristic that makes them structurally costly at scale: they execute tasks in chains, which means that each step in the process consumes tokens — including the intermediate steps of reasoning, verification, and error correction. A task that a human resolves with a single decision may require dozens of calls to the model before producing a result. That multiplier was not evident in pilots with moderate usage. It became visible when companies deployed these tools at the scale of hundreds or thousands of employees simultaneously.
The result is a gap between the perceived value during the experimental phase and the real cost during the operational phase. And that gap does not close with marginal efficiency improvements. It requires either radically different pricing models, or a thorough review of which tasks genuinely merit being resolved with artificial intelligence agents and which can be resolved more cheaply with simpler processes.
The Next Cycle Will Not Be Won by Whoever Sells the Most Tokens
The most direct conclusion emerging from Altman's statements and the simultaneous behavior of the world's largest companies is that the enterprise artificial intelligence sector is entering its second phase. The first phase was adoption driven by enthusiasm — funded by innovation budgets and with a high tolerance for uncertainty about returns. The second phase is adoption driven by justification, where spending on artificial intelligence competes at the same table as spending on infrastructure, personnel, and operations, and must demonstrate the same kind of measurable return.
That transition is not negative for the sector. But it does change who wins within it. In the first phase, the winners were those who offered the most capable model and the smoothest experience. In the second phase, the winners will be those who can demonstrate with precision how much each outcome costs and how much it is worth. That favors providers who develop observability tools, cost control mechanisms, and outcome attribution capabilities — not just those who scale the raw capacity of the model.
Altman projects another one-million-times growth in token consumption. If that growth materializes without the cost structure becoming more transparent and controllable for buyers, what will occur is not a sustained expansion of the market but a series of budgetary corrections that will fragment adoption. The corporate meme he himself cited — the annual budget consumed in the first quarter — is not a charming anecdote. It is the precise description of the structural limit of the current monetization model based on token volume, which grows in revenue for sellers in exactly the proportion that it generates unsustainable pressure for the buyers who finance it.
The architecture that would allow both curves to coexist without one cancelling the other does not yet exist with any clarity. Until it does, every token consumption record will simultaneously be good news for the infrastructure layer and a warning signal for the continuity of the corporate spending that funds it.
