Microsoft considers using Chinese model DeepSeek in Copilot to reduce AI costs

What happened?

According to Axios, reported by The Next Web, Microsoft is exploring the use of a self-hosted and fine-tuned version of the Chinese model DeepSeek V4 —or another open-source model— to power Copilot Cowork, the agentic assistant of its Microsoft 365 suite. The main reason is cost: OpenAI's proprietary models, which currently power Copilot, are significantly more expensive to run at enterprise scale. Microsoft confirmed to Axios that it is looking for cheaper options without compromising service quality. A spokesperson noted: 'We are actively evaluating ways to reduce AI costs without sacrificing user experience.' The company has not made a final decision, but the fact that it is publicly considering alternatives to OpenAI marks a milestone in their relationship.

Why is it important?

This move, if realized, would represent a major strategic shift. Microsoft has invested over $13 billion in OpenAI since 2019, integrated its models into Azure, Office, Bing, and Windows, and holds a significant stake in the startup. Turning to a Chinese open-source model like DeepSeek would indicate that cost efficiency is taking precedence over technological exclusivity. DeepSeek V4, released in 2025, has demonstrated competitive performance with GPT-4 in reasoning, coding, and language understanding tasks, but with inference costs up to 10 times lower according to industry estimates. The 1.5 trillion parameter model uses a Mixture-of-Experts (MoE) architecture that activates only a fraction of its parameters per token, drastically reducing computational consumption. Additionally, DeepSeek is open-source, allowing Microsoft to host it on its own infrastructure, avoiding OpenAI's per-token API costs.

The decision also has geopolitical implications. DeepSeek is developed by Chinese company High-Flyer Capital Management, and its use by a US company could create regulatory tensions, especially under the current administration that has tightened restrictions on Chinese technology. However, being an open-source model, Microsoft could argue it is not transferring sensitive technology.

Market consequences

Microsoft's potential migration to DeepSeek could trigger a price war in enterprise AI. OpenAI, which relies on Microsoft as its largest customer (estimated to generate over $1 billion in annual revenue from Azure OpenAI Service), could be forced to lower prices or risk losing its key partner. This would benefit startups and companies using OpenAI APIs, though it could also reduce the company's margins. For the AI model market, it would signal that even tech giants are willing to abandon exclusivity for efficiency, accelerating AI commoditization. Companies like Google, Amazon, and Meta could follow suit, exploring open-source models like Llama or Mistral to cut costs. For DeepSeek, it would be a huge endorsement, validating its technology globally and potentially boosting its adoption in the West. However, it could also attract stricter regulatory scrutiny over the security of Chinese models.

Historically, Microsoft has already diversified its AI model sources. In 2024, it integrated models from Meta (Llama 3) and Mistral into Azure AI Studio, and developed its own small model family Phi-3 for low-cost scenarios. But DeepSeek V4 represents a qualitative leap as a cutting-edge model with world-class performance. The company has also experimented with open-source models for specific tasks, such as code generation with Code Llama. However, replacing OpenAI in a flagship product like Copilot Cowork would be unprecedented.

What readers should know

Nothing is confirmed yet; Microsoft is only evaluating options. Copilot Cowork is a more advanced version of Copilot for enterprises, with task automation capabilities like summarizing emails, scheduling meetings, and generating reports. If DeepSeek is implemented, it would likely be used for less critical tasks or as a complement to OpenAI models, not a full replacement. The move would not affect end consumers of Copilot on Windows or Edge, at least for now. Microsoft could adopt a hybrid approach: using OpenAI for tasks requiring maximum precision and DeepSeek for routine tasks where cost is more relevant. Additionally, the company could fine-tune DeepSeek with proprietary data to improve its performance in enterprise scenarios.

It is important to note that DeepSeek V4 has been criticized for potential biases and security issues, common in models trained on Chinese data. Microsoft would need to implement additional filtering and alignment layers to ensure it meets Western standards. There is also a risk that the model is more vulnerable to adversarial attacks, though being self-hosted, Microsoft would have full control over the infrastructure.

“Microsoft is actively looking for ways to reduce AI costs without sacrificing user experience,” a spokesperson told Axios.

Historical context

This is not the first time Microsoft has diversified its AI models. It has already experimented with models from Meta (Llama) and Mistral. However, DeepSeek represents a qualitative leap as a cutting-edge Chinese model. The company has also developed its own small models (Phi-3) for low-cost scenarios. In 2024, Microsoft launched Copilot for Microsoft 365 at $30 per user per month, but inference costs were so high that the company barely made a profit. The pressure to monetize AI has led to this exploration. Additionally, the relationship with OpenAI has become more complex after Sam Altman's departure and return, and Microsoft has sought to reduce its dependence. In 2025, OpenAI released GPT-5, but its costs remain high. DeepSeek V4, on the other hand, offers similar performance at a much lower cost, making it attractive for large-scale applications.

Impact on users and businesses

For businesses using Copilot, the change could translate into lower prices or additional features without cost increases. If Microsoft manages to reduce its inference costs, it could pass those savings to customers, making AI more accessible to SMEs. For end users, the experience could be similar, though perhaps with subtle differences in response quality. DeepSeek V4 is known to be faster in certain tasks but may have gaps in up-to-date knowledge or cultural biases. Developers could benefit from a more accessible open-source model, as Microsoft might release fine-tuned versions of DeepSeek to the community. However, there is also a risk that the decision creates uncertainty among enterprise clients who have invested in the OpenAI ecosystem, who might wonder if Microsoft is committed to that technology long-term.

In market terms, the move could pressure OpenAI to innovate faster or lower prices. It could also incentivize other cloud providers like Google Cloud and AWS to offer alternative models. For investors, the news is a sign that AI is becoming a commodity, where cost efficiency is the key differentiator. Startups building on proprietary models could be affected if large clients opt for cheaper alternatives.