Key Takeaways:
Microsoft's decision to abandon unlimited AI usage for its enterprise Copilot Cowork product marks a turning point: even the world's second-most-valuable company cannot absorb the token costs of agentic AI at scale.
Microsoft on Tuesday made Copilot Cowork generally available with a usage-based billing model, replacing the unlimited-access approach that had been in place since the product's March preview. The AI agent — capable of independently executing multi-step tasks across Microsoft 365 documents even when a user's computer is off — now charges customers per credit, with each credit priced at $0.01 under the pay-as-you-go plan. A separate $30-per-user monthly license for Microsoft 365 Copilot remains mandatory for large enterprises.
"Some users complete hundreds of tasks per week, which is highly productive — but the cost can escalate very quickly," Charles Lamanna, Microsoft's executive vice president of Copilot, said.
The pricing shift coincides with reports that Microsoft is evaluating a self-hosted version of DeepSeek's V4 open-source model as a lower-cost alternative to the Anthropic and OpenAI models currently powering Copilot Cowork, according to Axios. Customers can already choose between Anthropic's Opus 4.8 and Sonnet 4.6, while those on the Frontier program can access OpenAI's GPT 5.5 and Microsoft's own Cowork 1 model. Adding DeepSeek V4 would give Microsoft a hosted open-source option that could dramatically reduce per-task token costs — a move that Axios reported could be announced within weeks.
The cost pressure is not unique to Microsoft. The Silicon Data Token Index, which tracks AI token pricing across major providers, has fallen in 12 of the past 13 trading sessions, approaching near-term lows as providers compete on price and enterprises push back against rising bills. Mason Daugherty, an enterprise AI consultant, said that in virtually every client conversation over the past two months, organization-wide token expenditure has emerged as a top concern. He predicted that "token economics" will become the dominant theme in AI procurement discussions over the next six to 12 months, as annual enterprise contracts come up for renewal and corporate finance teams question whether premium pricing for frontier models remains justified.
The Architecture Advantage
The competitive battleground is shifting from model intelligence to cost-efficient routing. Arvind Jain, chief executive of enterprise AI platform Glean, said the primary bottleneck for enterprise AI is no longer model capability but "token output efficiency" — how much useful work each consumed token produces. Most AI costs, he noted, come not from the prompt itself but from the surrounding infrastructure: retrieval, tool calls, memory management, and multi-step reasoning. A simple 11-word request can balloon into thousands of tokens once the system gathers context and processes tasks sequentially.
"Frontier intelligence is becoming abundant; efficient execution is not," Jain said. "The real competitive advantage comes from architectures that match the right model and reasoning depth to each task — systems with strong routing, cost controls, and governance."
This diagnosis aligns with Microsoft's strategy. Rather than simply replacing one model with a cheaper one, the company is building a model routing mechanism that can dynamically assign tasks to the most cost-effective option — Anthropic for complex reasoning, DeepSeek or Microsoft's own model for simpler queries. The billing system already breaks down costs into four components: model use, context retrieval, tool calls, and runtime, giving IT administrators granular visibility into where spending accumulates.
Nadella's Framework: Token Capital vs. Human Capital
Microsoft Chief Executive Satya Nadella recently articulated a broader framework that contextualizes the shift. Every company must build what he called "token capital" — its proprietary AI systems and capabilities — alongside "human capital," the knowledge, relationships, and judgment of its workforce. He argued that human capital does not depreciate as token capital grows: "Without human direction, you're just spinning compute in place."
Nadella said the true test of enterprise AI strategy is whether a company can swap out its underlying foundation model without losing the proprietary knowledge and capabilities it has accumulated. "That is the core test of whether you maintain control and sovereignty in the coming era," he said.
The warning carried an implicit tension: Nadella cautioned against allowing all value to concentrate in a handful of dominant models, comparing it to how globalization hollowed out industrial economies. Yet his own company is simultaneously deepening ties with OpenAI and Anthropic while exploring a Chinese open-source alternative — a balancing act that reflects the cost pressures facing every enterprise AI buyer.
For investors, the implications are clear. Premium API providers like OpenAI and Anthropic face margin compression as enterprises demand cheaper alternatives, while open-source model hosts like DeepSeek gain procurement traction. Companies building model routing and cost-optimization infrastructure — middleware that sits between the user and the model — may capture disproportionate value as token economics becomes the central procurement criterion. Microsoft shares trade at roughly 30x forward earnings; the Copilot Cowork transition could save the company hundreds of millions in annual inference costs if DeepSeek V4 is adopted, but it also signals that enterprise AI margins across the industry are tightening faster than many models' pricing power can sustain.
This article is for informational purposes only and does not constitute investment advice.
