Microsoft DeepSeek: a Chinese model for Copilot Cowork?

      Microsoft is contemplating the integration of a Chinese AI model into its enterprise Copilot, primarily for financial reasons. The company informed Axios that it is exploring a self-hosted, fine-tuned version of DeepSeek V4 or another open-source model as a more economical alternative to support Copilot Cowork, the intelligent assistant within its Microsoft 365 suite. It anticipates launching a lower-cost model in the coming weeks.

      Simultaneously, Microsoft is transitioning Copilot Cowork to a usage-based pricing model, charging companies based on their actual compute usage instead of a flat fee.

      This change reveals insights into the economics of agentic AI. Tools like Copilot Cowork, Anthropic’s Claude Code, and OpenAI’s Codex repeatedly call a model while completing tasks, which proves both powerful and costly. Charles Lamanna, Microsoft’s executive vice-president for Copilot, agents, and platform, noted, “We have users who perform hundreds of tasks a week, which is excellent for productivity, but it can lead to significantly high costs.”

      Copilot Cowork presently utilizes Anthropic and OpenAI models, both of which have increased their prices and moved away from unlimited plans. Microsoft has already implemented metering for GitHub Copilot for the same reasons, making a more affordable open-source engine a logical next step.

      The most economical option happens to be Chinese. DeepSeek V4, released in April, is open-source, favored by developers, and considerably cheaper to operate, which is why it is on Microsoft’s radar.

      However, its Chinese origin complicates matters. Political tensions have led to discussions in Washington about banning DeepSeek, threats against Chinese AI companies, and a recent requirement for Anthropic to restrict access to its top models for users outside the U.S., escalating into crisis talks with the Commerce Department.

      Microsoft is clearly cognizant of the implications. It states that any DeepSeek option would be optional for customers and fully hosted on Azure, ensuring that data remains within Microsoft’s cloud, adhering to its security, compliance, and residency protocols. Additionally, the model has been fine-tuned and enhanced with safeguards aimed at reducing bias.

      The broader perspective indicates that Microsoft aims to reduce reliance on any single lab. Having moved away from its exclusive, often strained partnership with OpenAI, it is now adopting a multi-model strategy, integrating different engines internally.

      As of now, this is still an evaluation rather than a definitive choice, and Microsoft plans to announce its decision when the cheaper tier is launched. However, the willingness to even mention DeepSeek as a potential option in the current climate highlights the mounting pressure from the costs associated with running these AI agents.