The "tokenmaxxing" era is officially over. For two years, enterprise AI operated on a simple philosophy: use more, optimize later. Developers were rewarded for pumping tokens into AI models. Internal leaderboards celebrated heavy usage. CFOs signed off on AI budgets with minimal scrutiny. That era ended in April when Uber's CTO revealed the company had burned through its entire annual AI budget in four months.
This is not an Uber problem. This is your problem.
The Wake-Up Call Every CIO Needs to Hear
Uber's Chief Technology Officer Praveen Neppalli Naga disclosed that the ride-sharing giant had exhausted its AI spending allocation for the full calendar year — by April. Teams across the company had wide latitude to experiment with AI tools. Those teams used that latitude aggressively. The result: a budget crisis that forced Uber to implement monthly per-employee spending tiers starting at $1,500, with higher access requiring explicit management approval and justification.
Think about what that means operationally. Thousands of engineers and business analysts who had unrestricted access to AI tools suddenly faced hard limits. Projects built on AI-powered workflows had to justify their token consumption. One of the most data-sophisticated technology companies on the planet found itself rationing AI like a scarce resource.
It is not alone. Not by a long shot.
The Pattern Now Spreading Across Industries
CNBC has identified the root cause as "tokenmaxxing" — the enterprise tendency to maximize AI consumption without benchmarking costs against measurable outcomes. What started as encouraged experimentation became uncontrolled infrastructure spend.
The correction is happening simultaneously at multiple levels.
At the startup level, AI spend is now exceeding payroll at companies of all sizes. The CEO of one 25-person AI startup told CNBC he moved 100% of his company's traffic from a top-tier AI provider to a lower-cost open-weight alternative. The switch will save the company millions within months. Even after the switch, AI costs are still projected to exceed the company's entire payroll. "It's a matter of survival for the business," he said.
At the enterprise level, Wall Street analysts are publicly flagging the risk. Gil Luria, an equity analyst at D.A. Davidson, told CNBC that current AI adoption growth rates may represent the peak velocity for major AI providers — calling it a basic math problem. His specific concern: large enterprise customers will begin imposing hard limits on what he described as "out-of-control" token expenditure. When major analysts start using that language in public, CFO attention follows.
At the board level, CFOs are now directly engaged in AI budget discussions that 18 months ago were handled entirely by engineering and product teams. What was treated as an R&D experiment line item is appearing on quarterly earnings calls alongside headcount and real estate. The CFO-CTO conversation on AI has fundamentally changed character.
The Numbers Behind the Shift
Here is the paradox that makes this moment genuinely complex: enterprise AI adoption is accelerating at the same time companies are pulling back on spend. These two dynamics are not contradictory — they reflect a market maturing from experimentation to accountability.
OpenAI's Chief Revenue Officer disclosed in April that business customers now represent more than 40% of total revenue, with enterprise on pace to reach parity with the consumer business by the end of 2026. The company's APIs are processing 15 billion tokens per minute. Its coding tool reached 3 million weekly active users — up five times since January 2026.
Anthropic's annualized revenue run rate stood at $47 billion as of May. Both Anthropic and OpenAI have reportedly filed confidentially for IPOs, according to CNBC. That means both companies face an imminent reckoning: public market investors will demand proof that growth is durable, not just fast. The budget pressure mounting from their largest enterprise customers is arriving at exactly the wrong moment.
The analyst assessment from D.A. Davidson is pointed: these growth rates may be the fastest either company ever posts. Not because AI adoption will reverse, but because the unconstrained experimentation phase — where enterprises threw budget at AI with minimal accountability — is ending.
The Open-Source Wildcard Changing the Leverage Dynamic
One factor accelerating the budget reckoning: enterprise buyers now have credible alternatives.
Eighteen months ago, if a company needed frontier-level AI performance, the practical choices were OpenAI or Anthropic. Microsoft and Google offered access to similar capabilities through their cloud platforms, but the core model dependency was tight. That market structure gave AI providers substantial pricing power.
That pricing power is eroding faster than most observers expected.
Open-weight models have closed meaningful performance gaps on specific task types — particularly coding assistance, document analysis, and structured data extraction. Cloud providers are adding model flexibility, enabling enterprises to route different workloads to different models based on cost-performance profiles. What was a binary choice (premium model or nothing useful) is now a spectrum.
The startup example above — switching 100% of traffic to save millions — is not an outlier. It reflects a new negotiating dynamic. Enterprise procurement teams can now credibly threaten to switch providers, and AI providers know it. The ability to shop models has shifted leverage in ways that weren't possible at this scale a year ago.
For CIOs evaluating infrastructure: the performance gap between frontier models and capable open-weight alternatives has closed significantly for high-volume routine tasks. That changes the build vs. buy calculation for many workloads.
OpenAI's Strategic Response: The Unified Agent Layer
OpenAI is not standing still. The company's enterprise strategy has pivoted around a platform called OpenAI Frontier, designed to give organizations a single plane to deploy and manage AI agents across entire business functions — not isolated point tools.
Several large enterprises are already using Frontier to deploy agents that move across internal systems and data sources. The commercial pitch is coherent: instead of managing separate AI tools for legal, finance, HR, and engineering — each with separate pricing, data governance requirements, and integration costs — organizations run agents through a single managed platform. Agents retain context across complex multi-step tasks through a stateful runtime environment developed jointly with AWS.
The partner network behind Frontier is substantial. McKinsey, Boston Consulting Group, Accenture, and Capgemini are consulting partners. AWS, Databricks, and Snowflake are infrastructure partners. This is not incidental — OpenAI is betting that enterprise deployment requires the same professional services ecosystem that made Oracle and SAP dominant decades ago.
The second strategic pillar is a unified daily interface for employees — a single place where workers interact with AI agents throughout the workday, rather than jumping between specialized tools. The company's coding tool is the early proof point, with customers building multi-agent systems capable of executing complete engineering workflows end to end.
The broader hypothesis is that platform consolidation reduces both cost and complexity for enterprise buyers. If token consumption runs through a single managed layer, enterprises gain the visibility they currently lack: which teams consume what, allocated to which business outcomes, measured against what ROI benchmarks.
Whether that platform strategy can sustain current revenue growth trajectories through an IPO is the central question for OpenAI's next twelve months.
Five Things the Best-Run Enterprises Are Doing Right Now
The companies navigating this transition successfully are not the ones spending the most on AI. They are the ones that built governance infrastructure before budget pressure forced them to.
1. Tag AI consumption like cloud compute. The discipline that solved cloud cost sprawl applies directly here. Every AI API call should be tagged to a team, project, and business outcome. If you cannot see which teams are consuming what volume, you cannot make intelligent allocation decisions. Implement tagging before budget review season, not during it.
2. Separate high-stakes tasks from high-volume tasks. Not every AI request requires frontier-model performance. Document summarization, meeting transcription, internal FAQ responses, and first-draft generation often perform adequately at significantly lower cost. Reserve premium access for tasks where quality genuinely matters: regulatory compliance analysis, customer-facing outputs, complex code generation with security implications.
3. Build or buy a model routing layer. Sophisticated enterprises are deploying middleware that routes tasks to different models based on complexity, latency requirements, and cost profiles. A straightforward employee question routes to a fast, cheap model. A contract review with regulatory implications routes to a premium model. The routing logic can itself be rules-based or AI-assisted. Either approach delivers meaningful cost reduction without degrading output quality on what matters most.
4. Tie AI budgets to business outcomes, not team headcount. Uber's fix — the $1,500 per-employee monthly tier — is a blunt instrument that treats AI consumption as a cost to control rather than a capability to optimize. A more durable approach allocates budget to measured outcomes: $X for customer support deflection rate, $Y for developer productivity uplift, $Z for sales research time savings. Budgets tied to outcomes can be evaluated, adjusted, and defended to boards.
5. Run a formal open-source evaluation on high-volume workloads. If your organization is spending more than $500,000 annually on AI API costs, a structured evaluation of capable open-weight alternatives is now worth the engineering investment. The evaluation framework is straightforward: cost per task, quality score on a representative test set, latency under load, and data governance implications. The findings will either confirm that premium model performance justifies premium pricing for your specific workloads, or surface meaningful savings opportunities.
The Bottom Line for Leaders
What is happening is not a retreat from AI. It is the inevitable maturation of a technology category that moved faster than enterprise governance could keep pace with.
The parallel to cloud computing holds. In cloud's early enterprise years, organizations dramatically over-provisioned because running out of capacity felt catastrophic. FinOps eventually emerged as a discipline. Reserved instances and spot pricing became standard tools. Cost optimization became as important as scaling. AI is following the same arc, compressed into months rather than years.
The enterprises that emerge from this inflection point in the strongest position will not be the ones that spent the most on AI during the experimentation era. They will be the ones that built the governance infrastructure — budget visibility, model routing, outcome measurement, vendor optionality — before a crisis forced the conversation.
Uber burned its annual AI budget in four months. The board conversation that followed was painful. The $1,500 per-employee tier is a symptom of what happens when adoption outpaces governance.
Every CIO and CFO reading this is now having that board conversation in advance — or will be having it in arrears. The choice is which version you want.
Sources: CNBC, Fortune, MarketScale, D.A. Davidson research commentary, OpenAI CRO Denise Dresser (April 2026 enterprise note)
