Microsoft has bet billions on OpenAI, integrating its models into Azure, Office, and GitHub Copilot. Now, DIGITIMES reports that the company is turning its gaze toward DeepSeek, a Chinese startup renowned for efficient, open-source LLMs. The driver is not technical but economic: inference costs for GPT models are rising faster than revenue, prompting even hyperscalers to seek less expensive alternatives.

A partnership under financial strain

Microsoft’s investment in OpenAI, estimated at over $13 billion, has delivered a competitive edge but also a costly dependency. Every API call, every Copilot query, every behind‑the‑scenes processing adds marginal cost, and when multiplied across millions of users, it weighs heavily on TCO. Looking at DeepSeek does not signal an imminent divorce; rather, it introduces competitive pressure that can reshape internal negotiations and, over time, diversify Azure AI offerings.

DeepSeek and compute efficiency as a strategic lever

DeepSeek’s models – often based on Mixture of Experts architectures and released under permissive licenses – have earned credibility by delivering competitive performance with a smaller computational footprint. For a given task, they consume less VRAM and generate tokens at lower cost than dense models of comparable capability. This makes them attractive not only for cloud but also for the deployment model most relevant to AI-RADAR readers: on-premise, where controlling capital expenditure and energy costs is paramount.

Beyond Microsoft: what changes for businesses

The Redmond giant’s move confirms a broader trend: large organizations are questioning the economic sustainability of proprietary APIs when open-source, self-hosted, or integrated models offer similar quality at predictable cost. Moving inference to on-premise hardware – or even to regional cloud with an open-source model – shifts the TCO equation and addresses data sovereignty requirements. DeepSeek, in particular, can be quantized and optimized for consumer or edge hardware, lowering the entry barrier significantly.

Outlook and trade-offs

Should Microsoft actually adopt DeepSeek in production, the market would witness a major realignment: Redmond would legitimize enterprise use of non‑OpenAI LLMs, accelerating the maturity of the open ecosystem. Questions remain about governance of Chinese models and compliance in regulated environments – issues that any on‑premise deployment must carefully assess. For anyone planning an AI strategy, this episode shows that flexibility in model mix and the ability to shift workloads between cloud and local infrastructure are becoming indispensable competitive levers.