Key Takeaways:
The era of unlimited AI token consumption inside Fortune 500 companies is ending, replaced by strict budgets, usage caps, and a scramble for cheaper models as per-employee costs reach $7,500 a month.
AT&T has restricted employee access to Microsoft's GitHub Copilot. Meta tightened spending on Anthropic and other AI services. Uber exhausted its entire 2026 AI coding budget by April and capped each worker at $1,500 per tool per month. Walmart set limits on internal AI agents. Amazon abolished the internal leaderboard that ranked employees by AI usage — after discovering workers were burning through compute just to climb the rankings.
The reversal from "tokenmaxxing" — the practice of maximizing AI token consumption — to "tokenminimizing" is sweeping across the largest corporate users of generative AI, according to people familiar with the matter. At the most AI-intensive companies, per-employee monthly AI costs have reached $7,500, The Information reported, a figure that has forced chief financial officers to intervene.
"Companies are realizing that agentic AI workflows don't scale under flat-rate pricing," said Alex Nguyen, enterprise AI analyst at Edgen. "When a single AI agent can chain 50 model calls to complete one task, the math breaks at enterprise volume."
The $7,500-per-employee math problem
The structural shift traces to the rise of agentic AI tools — software that autonomously chains multiple model calls to complete complex tasks across email, spreadsheets, and messaging apps. Unlike manual chatbot queries, these agents consume tokens in bursts that are difficult to predict or cap.
Microsoft discovered that some engineers were spending $500 to $2,000 a month on token fees from Claude Code alone, according to internal data reviewed by the company. Enterprise AI interaction costs have jumped 30-fold since 2023, and Goldman Sachs projects agentic workflows could drive token demand up 24 times from current levels.
The price gap between premium and open-source models makes the tension acute. Anthropic's latest flagship costs roughly $50 per million tokens, while DeepSeek V4 Pro runs at about $0.87 per million tokens — a 57-fold difference, according to pricing data published by both companies. Microsoft is now exploring a fine-tuned, self-hosted version of DeepSeek V4 as a lower-cost backend for its Copilot Cowork product, Axios reported on June 16.
Not every company is tightening. Databricks imposes no AI budget cap on its engineers, engineering leader Nikita Shamgunov said at a Nebius event last week. Box Chief Executive Officer Aaron Levie said his company never adopted tokenmaxxing in the first place. "We didn't have a leaderboard, so we didn't go astray," Levie said.
The gatekeepers of the new AI budget era
The cost-control wave is creating a new layer of infrastructure demand. Microsoft and Databricks have each launched "gateway" tools that monitor employee AI usage and enforce spending limits. Nvidia-backed Factory, valued at $1.5 billion, released a model router this month that automatically assigns low-complexity tasks to cheaper models.
Palantir and Box executives report growing demand from enterprise clients seeking to shift simple tasks from expensive frontier models to cheaper or open-source alternatives. The pattern mirrors the shift from all-premium to tiered cloud computing that reshaped the public cloud market a decade ago.
Microsoft Chief Executive Officer Satya Nadella framed the trend as a strategic necessity. "None of us want to see a world where every company in every industry cedes value to a handful of 'winner-take-all' models," he wrote on X last week. The comment carries weight given that Microsoft's own productivity software now competes with Anthropic and OpenAI on pricing.
Microsoft's new Copilot Cowork product, which became generally available June 16, embodies the tension. It requires a $30-per-user-per-month Microsoft 365 Copilot license plus additional usage-based charges through Copilot Credits — a dual subscription-plus-consumption model that mirrors Anthropic's enterprise pricing. Microsoft Executive Vice President Charles Lamanna said customers "can choose how to manage costs," including setting per-employee usage caps and swapping Anthropic models for OpenAI or Microsoft's own alternatives.
The question for investors is whether cost controls will blunt the productivity gains that justified enterprise AI spending in the first place. Microsoft shares trade at 33 times forward earnings, with AI-related revenue growth a key pillar of the bull case. If token throttling slows adoption, the revenue forecasts baked into current valuations may prove optimistic. For now, the CFOs have the upper hand.
This article is for informational purposes only and does not constitute investment advice.
