Microsoft considers DeepSeek models as it looks for cheaper ways to run Copilot

Microsoft is weighing the use of DeepSeek models as part of its broader effort to control the cost of running generative AI services, according to a report by Yahoo Finance. The consideration, described as a potential way to lower the cost of providing Copilot responses, comes as Microsoft pushes Copilot deeper into Office, Windows, developer tools, and cloud offerings, increasing the volume of AI queries the company must serve.

In the Yahoo Finance report, the core rationale is cost. Copilot usage is expected to keep rising, which in turn can increase the amount of compute needed to generate answers in real time. Generative AI “inference” costs, meaning the processing required to produce outputs from a model for each user request, tend to be a major driver of ongoing operating expense. Microsoft’s reported interest in DeepSeek models suggests it is looking at model sourcing and routing decisions not only for capability, but also for unit economics.

Microsoft has not publicly confirmed any specific decision to adopt DeepSeek models, nor has it outlined a timetable for the move described in the report. Nor is it clear from the published account which Copilot experiences or workloads would be targeted first, such as customer support, coding assistance, or enterprise knowledge retrieval. The report frames the issue as “weighing” options rather than announcing an imminent rollout.

The potential shift highlights a broader pattern in the AI industry. As companies scale assistant products, they increasingly compare models on total cost of ownership, not just quality. That can involve evaluating performance at different tasks, energy and hardware requirements, and how well a model integrates with an existing system that includes retrieval, guardrails, and content filtering. In practice, the “best” model can differ by workload, and the lowest-cost option can depend on response length, latency targets, and how many times a system must re-run or revise generations when users are not satisfied with an initial output.

For Microsoft, Copilot is not a standalone app. It is woven into the company’s productivity and cloud strategy, including enterprise deployments where admins may demand strong reliability, compliance, and predictable performance. Cost control therefore has implications beyond margins. If response generation becomes more expensive than planned, it can limit how broadly Microsoft is able to promote usage, how aggressively it can bundle Copilot into subscriptions, and how quickly it can expand features.

At the same time, using a new model family in Copilot can raise operational questions that the company would typically address internally before any customer-facing change. Those questions include how the model is licensed and accessed, how it performs on Microsoft’s safety and policy requirements, and what monitoring is needed to detect quality regressions or unexpected behaviors. Even when a model is cheaper, the company must ensure that the end-to-end experience still meets expectations for accuracy and risk management.

What is not disclosed in the report is as important as what is. Microsoft does not indicate in the available public information whether it has already run formal benchmarks on DeepSeek models for Copilot-like tasks, whether it would use DeepSeek models for specific sections of the workflow (for example, drafting versus final answering), or whether multiple models would be used together. The report also does not specify whether the company’s goal is to cut costs through lower-priced compute providers, more efficient routing, or purely through swapping models.

For the market, the next indicates to watch are any Microsoft statements tying Copilot growth to cost structure, any changes to Copilot pricing or packaging, and any updates from Microsoft on model strategy in its AI and cloud services. If Microsoft moves forward, analysts and customers will likely focus on whether the cost improvements come without a visible trade-off in response quality, latency, or compliance controls.