Microsoft weighs self-hosted DeepSeek V4 model to cut Copilot agent costs
Microsoft $MSFT is weighing a self-hosted version of China's DeepSeek-V4 as a lower-cost option for its Copilot agent, as rising inference costs push the company toward usage-based pricing and a multi-model strategy in Microsoft 365.
Microsoft $MSFT is considering a Microsoft-hosted version of DeepSeek-V4, the latest model from the Chinese AI lab, as a lower-cost option inside Copilot Cowork, its enterprise AI agent. The move accompanies a shift toward usage-based pricing and a broader multi-model strategy across Microsoft 365. The driver is economics. Agentic AI is expensive to run, and Microsoft has reportedly concluded that Copilot Cowork cannot be offered responsibly on an all-you-can-eat basis. That leaves three choices: charge customers by usage, absorb runaway inference costs, or route lower-stakes work to cheaper models. A modified, self-hosted DeepSeek-V4 would let Microsoft reserve the most expensive models from OpenAI and Anthropic for tasks that truly need them. The strategic and political stakes are larger than cost control. This is a test of whether Microsoft can make AI economics work without being permanently dependent on a single frontier-model supplier. It is also a test of whether Washington will tolerate a major US software company wrapping a Chinese-origin open-weight model in Azure compliance controls and selling it to enterprise customers. How regulators and customers react will shape whether Chinese open models become a normal part of Western enterprise software stacks.