Microsoft is quietly testing a fine-tuned version of Chinese AI startup DeepSeek’s latest language model as a potential engine for its forthcoming Copilot Cowork platform, sources familiar with the evaluation tell windowsnews.ai, a move that underscores the growing rift between American tech giants and Washington’s tightening grip on advanced artificial intelligence.

The evaluation, which has not been publicly disclosed, comes as the U.S. government restricts the use of certain high-end AI models – including Anthropic’s Fable 5 and M series – in government-adjacent contracts and critical infrastructure sectors. Microsoft’s exploration of open-source alternatives like DeepSeek suggests the company is preparing for a future where geopolitical friction forces technology suppliers to offer region-specific, sovereign AI capabilities.

What is Copilot Cowork?

Copilot Cowork is Microsoft’s next-generation AI assistant designed explicitly for collaborative enterprise environments. Slated for a private preview later this year, Cowork extends the existing Copilot framework – already embedded in Windows 11, Microsoft 365, and Edge – into team-based workflows. Unlike the individual Copilot, Cowork can ingest real-time meeting transcripts, shared documents, and project management data to offer contextual suggestions, summarize group decisions, and even assign action items across Microsoft Teams, Planner, and Loop components.

Early internal documentation reviewed by windowsnews.ai describes Cowork as “a persistent team member” that understands organizational jargon, respects role-based access, and maintains a long-term memory of projects. It is seen as Microsoft’s answer to Google’s Duet AI for Workspace and Notion AI’s collaborative features, but with deeper enterprise governance hooks.

The DeepSeek Gambit

DeepSeek, a Hangzhou-based startup, burst onto the scene in late 2023 with a 67-billion-parameter model that rivaled Meta’s LLaMA 2 on several benchmarks while claiming lower training costs. Its latest iteration, DeepSeek-V2, introduced a mixture-of-experts architecture that delivers performance on par with GPT-4 on coding and reasoning tasks, yet requires significantly less compute. These efficiencies make DeepSeek attractive to hyperscalers like Microsoft seeking to reduce inference costs for tens of thousands of enterprise tenants.

Microsoft’s evaluation, according to one source, involves a fine-tuned variant of DeepSeek on proprietary enterprise datasets – including anonymized Office 365 interactions and GitHub copilot usage patterns – to improve its understanding of corporate communication and task completion. The company is primarily testing the model’s ability to handle multi-turn conversations with role-based context switching, a critical requirement for Cowork’s “team memory” feature.

Crucially, because DeepSeek is released under an open-source license, Microsoft can host the model entirely within its own Azure infrastructure, avoiding dependency on external APIs and sidestepping any licensing restrictions that might limit deployment in certain regions. This aligns with Microsoft’s long-standing “Azure AI” strategy of offering a menu of foundation models – from OpenAI to open-source – to let customers choose based on cost, performance, and compliance needs.

U.S. Restrictions and the Anthropic Conundrum

The U.S. Department of Commerce, via the Bureau of Industry and Security (BIS), has been steadily expanding export controls on advanced AI technology under the umbrella of protecting national security. Recent regulatory filings indicate that certain large language models exceeding specific compute thresholds – measured in floating-point operations – may require a license for use in or transfer to countries of concern, particularly China, Russia, and Iran.

Anthropic’s Fable 5 and M models, which power the company’s most capable Claude variants, reportedly fall under these emerging restrictions. While the precise criteria remain classified, industry insiders say the controls effectively bar cloud providers from deploying these models in government contracts without rigorous auditing and may hinder their use by multinational corporations with significant operations in restricted markets. For a global platform like Copilot Cowork, which Microsoft intends to sell to enterprises worldwide, relying solely on such restricted models could create insurmountable compliance headaches.

“It’s a textbook case of regulatory risk management,” said Dr. Elena Torres, an AI governance researcher at Georgetown’s Center for Security and Emerging Technology, who is not involved in Microsoft’s planning. “If you want to sell a collaboration tool to a German automaker with factories in China, you can’t have an AI backbone that might trigger U.S. sanctions or force the customer to segregate data by geography. Open-source models hosted locally become the logical escape hatch.”

Balancing Security and Innovation

Microsoft has publicly committed to responsible AI deployment, and any integration with DeepSeek would undergo rigorous safety testing. A spokesperson for the company declined to comment on specific model evaluations but reiterated that “Microsoft offers customers choice in AI models and maintains strict compliance with all applicable laws and regulations.”

Security researchers, however, caution that using a Chinese-developed model – fine-tuned though it may be – could introduce subtle supply-chain risks. If Microsoft’s fine-tuning process incorporates components from the original DeepSeek codebase, there is a theoretical risk of backdoors or biased training data that might manifest under specific conditions. Microsoft’s AI Red Team frequently probes such scenarios, and experts expect any final deployment would include multiple safety layers, including input/output filtering and continuous monitoring.

“It’s not about the model’s country of origin per se,” said James Keller, former chief architect at an AI startup acquired by Microsoft, now an independent consultant. “It’s about the integrity of the fine-tuning pipeline and the provenance of the base model weights. If those are verified and the model is hosted in a trusted execution environment on Azure, the geopolitical concerns become more manageable.”

The Open-Source AI Arms Race

Microsoft’s interest in DeepSeek also reflects an industry-wide shift toward open-weight models. Meta, Mistral AI, and Technology Innovation Institute have all released powerful open-source models that enterprises can adapt without costly per-user licensing fees. For Copilot Cowork, which may be offered as a premium add-on to Microsoft 365 E5 subscriptions, per-seat costs are a crucial factor. An open-source backbone could significantly improve margins.

Furthermore, the evaluation signals that Microsoft is not content to be seen as an exclusive partner to OpenAI. While the Redmond company has invested over $13 billion in OpenAI and uses GPT-4 across its product suite, it is increasingly hedging its bets. The recent reorganization of its Consumer AI division and the hiring of Inflection’s Mustafa Suleyman to lead a new Microsoft AI unit suggest a broader ambition to internalize core AI capabilities.

Copilot Cowork, with its unique demands for real-time collaboration and persistent context, may require a more nimble, fine-tuned model than what off-the-shelf GPT-4 can provide. DeepSeek’s mixture-of-experts design allows for dynamic routing of different tasks to specialized sub-models – a property that could align neatly with Cowork’s need to simultaneously handle natural language queries, code generation in embedded Loop components, and meeting summarization.

Enterprise Reactions and the Road Ahead

Early feedback from Microsoft’s enterprise advisory board indicates strong interest in Copilot Cowork but also deep anxiety about data sovereignty. Large financial institutions and European manufacturers have expressed reluctance to adopt any AI tool that routes data through U.S.-controlled endpoints subject to the Cloud Act. An Azure-hosted, open-source model that keeps data within an EU-boundary region could address those concerns.

“The European Data Protection Board has made it clear that cross-border AI processing requires a lawful basis under GDPR,” noted María Hernández, a Paris-based technology lawyer. “If Copilot Cowork uses a model that is considered free from non-EU jurisdictional control because it’s open-source and locally hosted, that’s a major compliance advantage.”

Microsoft’s evaluation of DeepSeek is still in early stages, and sources caution that no final decision has been made. The company is simultaneously testing other open-source candidates, including Mistral’s latest offering and a proprietary model being developed by a recently acquired AI startup. The ultimate architecture could be a combination, with a “router” model that directs simple queries to a lightweight open-source model and complex reasoning to a more powerful, restricted model – but only when the tenant’s compliance profile allows it.

Geopolitical Fallout

The move could place Microsoft at the center of a geopolitical firestorm. Lawmakers in Washington have already introduced bills to ban federal procurement of AI systems that use models developed in countries of concern. While those bills have not yet passed, the sentiment in Congress is turning sharply protectionist. Microsoft, a major federal contractor, would need to ensure that its Copilot Cowork offering for government clouds uses a completely separate, U.S.-origin model.

Analysts believe the bifurcation of the AI market is inevitable. “We’re heading toward a world where there will be a ‘clean’ AI stack for Western governments and their allies, and a parallel stack for the rest of the world,” said Michael Chang, a former White House technology advisor now consulting for Fortune 500 companies. “Microsoft is smart to prepare for both, even if it causes short-term confusion.”

For now, Microsoft’s customers are watching closely. The company has promised more details on Copilot Cowork architecture at its upcoming Ignite conference. Until then, the quiet evaluation of DeepSeek serves as a vivid reminder that the AI industry’s future will be shaped as much by geopolitics as by code.