Enterprises pull back on AI spending as token bills outpace productivity gains
Enterprise AI spending is contracting sharply as chief financial officers demand measurable returns on API bills that have outpaced actual productivity gains, threatening growth projections at OpenAI and Anthropic that depend on expanding token consumption. The reversal marks the end of what the industry calls “tokenmaxxing”, the practice of maximizing AI tool usage regardless of business value, and signals a fundamental shift in how large organizations evaluate artificial intelligence investments.
- Amazon dismantled an internal AI usage leaderboard after determining it generated busywork rather than productive output, with leadership explicitly warning staff against using AI without clear justification.
- Uber depleted its entire 2026 AI coding budget in four months and now enforces a $1,500 monthly spending cap per employee on AI tools after uncontrolled consumption threatened budgets.
- OpenAI and Anthropic face a structural threat as enterprises shift to cheaper models and open-source alternatives, with IDC forecasting 70% of leading AI enterprises will use multiple models by 2028.
- $1,500 Monthly spending cap per Uber employee on AI coding tools, down from unlimited consumption
- 70% Of leading AI enterprises predicted to use multiple models by 2028, versus single-provider reliance today
- 4 months Time required for Uber to exhaust entire 2026 AI coding budget before implementing controls
The pivot away from aggressive AI adoption represents one of the most consequential reversals in enterprise technology spending this year. After months of mandatory AI tool rollouts and aggressive adoption campaigns, major corporations are now systematically auditing their usage and cutting back on the expensive APIs that power large language models.
The shift reflects a hard reckoning between the theoretical productivity gains promised by AI vendors and the actual business impact observed in practice.
CFOs at major organizations are no longer accepting rising token bills as a cost of doing business and are instead demanding evidence that AI investments are actually completing work, saving time, or reducing rework, not simply consuming more computational resources.
Amazon Dismantles AI Leaderboard After Finding System Encouraged Busywork Over Results
Amazon’s decision to dismantle an internal leaderboard tracking employee AI usage crystallizes the problem facing enterprises. The company’s leadership concluded that the system, designed to encourage adoption, was instead driving workers to use AI tools on trivial tasks simply to increase their standing on the leaderboard.
An Amazon senior vice president told staff bluntly: “Please don’t use AI just for the sake of using AI.” This message directly contradicts the adoption-at-all-costs messaging that dominated corporate AI strategy throughout 2024 and early 2025.
Amazon’s experience is not isolated; it reflects a pattern emerging across the enterprise sector.
Uber provides a more dramatic example of uncontrolled AI spending spiraling into crisis. The company burned through its entire 2026 AI coding budget within four months before implementing emergency controls.
In response, Uber imposed a $1,500 monthly spending cap per employee on AI coding tools, a hard constraint that would have seemed impossible just months earlier when unlimited AI adoption was viewed as competitive necessity.
Meta likewise sent an internal memo to approximately 6,000 employees warning of an “exponential increase” in AI usage and flagging that the company faced billions of dollars in internal AI infrastructure costs.
Accenture Reverses AI Adoption Stance as Spending Becomes “Unpredictable” and Unquantified
The contradiction at Accenture underscores how rapidly corporate positions on AI are shifting. The consulting giant previously told employees they could “risk losing out on promotions” if they failed to adopt AI tools, an aggressive mandate designed to drive maximum usage. Within months, that pressure has inverted.
Leaked audio from an internal meeting captured an Accenture executive acknowledging that AI spending is “becoming very unpredictable” and that leadership across finance, operations, and IT levels continues to ask whether the company is “getting value from what we’re spending.”
This confession reveals the core problem: enterprises adopted AI spending without establishing clear metrics for success. CFOs are discovering that high token consumption does not correlate with business outcomes, and in many cases, correlates with waste. The spending is unpredictable because usage patterns are disconnected from measurable work completion.
IBM’s analysis, authored by Adam McDaniel and Markus Eisele, argues that both “tokenmaxxing” and its opposite, excessive cost-cutting focused on minimizing tokens regardless of utility, represent failures of governance. IBM advocates instead for what it terms “valuemaxxing,” which measures completed tasks, time saved, and rework avoided rather than tokens consumed.
This framework reorients corporate AI spending away from consumption metrics and toward business outcomes, which is precisely what CFOs are now demanding to see.
OpenAI and Anthropic Face Threat as Enterprises Shift to Cheaper Models and Self-Hosted Infrastructure
The spending contraction threatens the financial models of the companies that benefited most from the tokenmaxxing boom. OpenAI crossed $25 billion in annualized revenue earlier this year while placing its valuation at $1 trillion.
Anthropic is valued several billion dollars lower but operates under the same assumption: that enterprise customers will continuously increase their token consumption as adoption deepens. Both companies are burning substantial cash on compute infrastructure, research, and hiring while betting that profitable growth will eventually follow from market expansion.
That bet is now under pressure.
Enterprises are already responding to cost pressure by fragmenting their AI vendor relationships. Instead of committing to a single flagship model from OpenAI or Anthropic for all workloads, companies are now reserving expensive, capable models for complex work requiring state-of-the-art performance and routing routine tasks to cheaper alternatives.
Some organizations are deploying open-source models on their own infrastructure, eliminating per-token charges entirely and replacing them with capital expenditure on compute hardware. The International Data Corporation projects that by 2028, 70% of leading AI-driven enterprises will operate with multiple AI models across their operations rather than relying on a single provider.
This structural shift from single-vendor dependency to multi-model strategies undermines the growth projections that OpenAI and Anthropic have built into their business plans. A company consuming tokens across five different models and platforms will almost certainly consume fewer tokens from any single vendor than a company locked into one.
Distributed adoption also means that smaller, more efficient models, whether from competitors or open-source communities, will capture increasing share of workloads that large language models currently handle at significant cost.
CFO Scrutiny Forces Recognition That AI Adoption Metrics Must Connect to Business Value
The underlying shift is structural and unlikely to reverse quickly. CFOs who approved unlimited AI spending during the hype cycle are now facing direct accountability for whether those expenditures produced measurable returns. Board-level scrutiny of AI spending is intensifying precisely as the novelty of the technology is wearing off.
Companies that adopted AI aggressively without establishing ROI frameworks are now in the position of having to justify continued spending to skeptical finance leadership.
This accountability creates a new discipline around AI deployment. Instead of asking “how can we use AI more,” enterprises are asking “what work does AI actually complete better and cheaper than our current process?” The answer to that question is narrower than the tokenmaxxing era assumed.
For routine customer service, basic code generation, and standard report writing, small models and open-source alternatives now deliver sufficient quality at a fraction of the cost of premium APIs.
For genuinely complex reasoning, specialized domain work, and novel problem-solving, flagship models still command premium pricing, but those workloads represent a smaller percentage of total enterprise AI usage than executives initially believed.
The contraction is already observable in API consumption trends among large enterprises.
Accenture’s leaked acknowledgment that spending remains “unpredictable” and that senior leadership cannot yet articulate clear value reflects the reality that many enterprises lack the governance infrastructure to manage AI costs effectively.
Some organizations are implementing formal approval workflows for new AI tool deployments, requiring business case documentation and ROI projections before approving API spending. Others are moving to fixed budgets for AI tools rather than allowing uncapped consumption.
These mechanisms will reduce total token consumption because they force teams to justify spending against actual business requirements rather than adopting AI tools speculatively.
The critical
