Microsoft May Turn to Self-Hosted DeepSeek V4 for Copilot Cowork as Enterprise AI Costs Surge
Microsoft is reportedly exploring a major change for Copilot Cowork that could stir debate in Washington: using a self-hosted version of DeepSeek’s V4 model, or a similar open-source AI model, to reduce reliance on increasingly expensive offerings from OpenAI and Anthropic.
The move would reflect a growing problem across the enterprise AI market. As businesses push deeper into generative AI, coding assistants, and agent-based workflows, token costs are becoming harder to ignore. For large companies, the price of running advanced AI systems at scale can rise quickly, especially when employees use models for complex tasks that require long prompts, large context windows, and repeated back-and-forth reasoning.
A token is the basic unit of text or data processed by an AI model. In simple terms, it is usually a small chunk of a word, often around four characters. Every prompt a user enters and every response an AI model generates consumes tokens. The larger the task, the more tokens are required. When AI tools are used for software development, research, automation, and multi-step agentic workflows, usage can grow dramatically.
That is where the business challenge begins. Enterprise AI platforms that once looked predictable under subscription-style pricing are now becoming far more expensive as companies increase usage. Advanced models from OpenAI and Anthropic are powerful, but their pricing structures are putting pressure on customers that need high-volume AI access.
Microsoft’s reported interest in DeepSeek V4 appears to be linked to this shift. Copilot Cowork is designed to combine enterprise Copilot capabilities with advanced AI models to help businesses automate tasks, coordinate workflows, and build more capable AI agents. However, as Microsoft moves Copilot Cowork toward a metered pricing model based on token consumption rather than a flat fee, model efficiency and infrastructure costs become far more important.
A self-hosted model could give Microsoft more control over performance, costs, security, and scalability. Instead of depending entirely on external model providers, Microsoft could run an open-source or open-weight model on its own infrastructure. That would potentially allow the company to offer enterprise customers more predictable pricing while still supporting demanding AI workloads.
DeepSeek has become one of the most closely watched AI companies in the world because of its ability to deliver capable models at lower operating costs. Its open-source approach has made it attractive to developers, researchers, and companies looking for alternatives to closed AI platforms. If Microsoft were to use a self-hosted version of DeepSeek V4 for Copilot Cowork, it would mark a significant moment in the enterprise AI race.
Still, such a decision would likely attract political scrutiny in the United States. DeepSeek is based in China, and Washington has become increasingly cautious about advanced AI technology, model access, cybersecurity risks, and the global competition for AI leadership. Even if Microsoft hosts the model on its own servers, the optics of using a Chinese-developed AI model inside a major enterprise product could trigger concerns among policymakers.
The timing is especially sensitive. U.S. officials have already shown a willingness to impose restrictions around advanced AI models and their capabilities, particularly when national security or cyber-risk concerns are involved. Any Microsoft move involving DeepSeek would likely be examined through that lens.
At the same time, the economics of AI are forcing companies to rethink their strategies. The cost of tokens is no longer a minor technical detail. It is becoming a core business issue. Enterprises want AI tools that are powerful, secure, reliable, and affordable at scale. If premium closed models become too expensive or too restrictive, companies will naturally look for alternatives.
This is where self-hosted AI models could gain momentum. For large cloud providers and software companies, hosting models directly can reduce dependency on third-party pricing decisions. It can also make it easier to customize models for specific enterprise use cases, optimize infrastructure, and manage sensitive data internally.
For Microsoft, the decision would not necessarily mean walking away from OpenAI or Anthropic. Instead, it could signal a more flexible AI strategy where different models are used for different workloads. High-end proprietary models may still handle the most complex tasks, while cost-efficient open-source models could power routine or high-volume enterprise functions.
DeepSeek’s rapid rise adds another layer to the story. The company has reportedly raised billions in funding at a massive valuation and is expanding its computing capacity. That growth suggests DeepSeek intends to compete aggressively in the global AI market, not only with consumer-facing chatbots but also with enterprise-grade models that can challenge established players.
If Microsoft does move forward with a self-hosted DeepSeek V4 model for Copilot Cowork, it could accelerate a broader industry shift. Businesses may begin prioritizing model cost, infrastructure control, and deployment flexibility as much as raw benchmark performance. In the next phase of AI adoption, the winners may not simply be the companies with the smartest models, but the ones that can deliver intelligence at a price enterprises can actually sustain.
For now, Microsoft has not confirmed such a transition. But the reported consideration alone highlights a key reality: the enterprise AI market is entering a cost-conscious era. As token usage explodes and agentic AI becomes more common, companies are searching for ways to keep innovation moving without letting AI budgets spiral out of control.
