REDMOND — Microsoft is changing how it sells its AI assistant, shifting Copilot from a flat subscription to billing based on token usage and adding DeepSeek’s models as a lower-cost option for enterprise customers. The changes, detailed in an update to the company’s commercial licensing documentation, mark one of the most significant adjustments to Microsoft’s AI strategy since it began bundling Copilot into its productivity software.
Under the new pricing, customers pay according to how much computing their AI usage consumes rather than a fixed fee per seat, according to people familiar with the plan. TechCrunch and Ars Technica both described the shift in the same terms: Microsoft is moving from selling seats to selling compute, a model that aligns the company’s revenue with the actual cost of running AI workloads.
The pricing change reflects the economics of AI products. Every Copilot interaction consumes expensive inference capacity, and heavy users of the assistant cost Microsoft far more than light users. Flat subscriptions could not capture that difference, analysts said, and the usage-based model lets Microsoft price the product closer to its cost structure while giving customers a way to scale AI spending with actual value received.
The more striking move is the addition of DeepSeek. The Chinese startup’s models, released over the past year, matched Western rivals on many benchmarks at a fraction of the training and inference cost, and its open-weight releases made it a fixture of enterprise evaluations. By offering DeepSeek as an option inside Copilot’s enterprise tier, Microsoft is acknowledging what many customers already concluded: that the most expensive model is not always the right one for every task.
The decision signals a quiet shift in Microsoft’s relationship with OpenAI. The two companies have been partners since 2019, with Microsoft investing billions and integrating OpenAI’s models across its products, and the relationship has been profitable for both. But Microsoft has steadily built a position of independence, offering models from multiple providers inside Azure and now inside Copilot itself. A person close to Microsoft said the company wants to be the marketplace, not the champion of a single model.
For enterprise customers, the change cuts both ways. Usage-based pricing offers flexibility for companies whose AI adoption is experimental, but it also removes the predictability of a fixed subscription and forces finance departments to forecast compute consumption. Analysts said the largest enterprises are likely to negotiate hybrid arrangements, mixing committed usage tiers with per-token pricing for overflow workloads.
The addition of DeepSeek also raises questions that Microsoft will have to manage. DeepSeek’s models are trained in China, and enterprise customers in regulated industries have expressed concerns about data governance. Microsoft said DeepSeek models deployed through Copilot run entirely in Microsoft’s cloud, with the same compliance controls applied to all models, and that customer data does not train the models. Whether that satisfies procurement departments in banking, health care and government remains to be seen.
The broader shift in the market is toward model diversity. Enterprises that once standardized on a single AI provider are now evaluating several, matching models to tasks and price points. Microsoft’s move mirrors that trend: by making Copilot a platform that hosts multiple models, the company positions itself to capture revenue regardless of which model wins the technology race.
Competitors are watching closely. Google has kept Gemini central to its enterprise offerings but has begun offering alternative models through its Vertex AI platform, and Amazon has made Bedrock the aggregation layer for a catalog of models that includes Anthropic, Meta and Mistral. The model marketplace model is becoming the default architecture of enterprise AI, analysts said, and Microsoft’s pricing change is an acknowledgment that the old subscription model belonged to a simpler era.
The pricing change has implications for Microsoft’s financial reporting. Usage-based revenue is less predictable than subscription revenue, and the shift will make Copilot’s contribution to growth harder to forecast quarter to quarter, but it also ties revenue directly to the AI workloads that are driving Azure’s growth. Microsoft has said AI services are adding points to Azure’s growth rate, and the new pricing model is designed to capture that consumption on Microsoft’s own books rather than through third-party channels. Analysts said the change aligns Microsoft’s incentives with its customers’: the company now makes more money when its AI products are actually used, and less when they are licensed but ignored.
Customer reaction has been mixed. Enterprises that rolled out Copilot to thousands of employees on fixed subscriptions face the prospect of bills that vary with usage, and finance departments are asking for tools to cap spending, which Microsoft has said it will provide. Startups and small businesses, which use Copilot lightly, could see costs fall. The deeper question is whether per-token pricing becomes the standard for enterprise AI, the way per-seat pricing became the standard for software in the 1990s, and Microsoft’s decision to move first is likely to shape how rivals price their own AI products. If the model works, Microsoft will have defined the economics of enterprise AI; if it fails, the company will have given its competitors a lesson in what not to do.
The immediate test is adoption. Microsoft has said Copilot usage has grown steadily, and the company’s chief financial officer has described AI as a growing share of revenue. The usage-based pricing model ties that growth directly to consumption, which could make AI a larger and more volatile line item in Microsoft’s financials. For the company’s investors, the question is whether selling compute is as profitable as selling seats once was, and whether customers accept the new meter.
