Microsoft has quietly begun evaluating DeepSeek's large language model for integration into its forthcoming Copilot Cowork enterprise agent, according to internal documents and people familiar with the initiative. The move, designed to slash the cost of delivering AI-powered workplace tools, has placed the company on a collision course with the Australian government, which earlier this year banned DeepSeek over national security concerns.

If approved, the plan would see DeepSeek models hosted on Microsoft's Azure cloud infrastructure, providing a layer of control and data residency promises that the company hopes will satisfy enterprise customers. But with Australia's ban extending to all government agencies and contractors, the project risks fracturing Microsoft's relationship with one of its most important public-sector markets.

Cost Pressures Drive Microsoft to Cheaper AI Models

The rapid expansion of generative AI features across Microsoft 365, GitHub, and the Azure ecosystem has placed enormous compute demands on the company's infrastructure. Integrating OpenAI's models—while lucrative—has proven expensive at scale. Every Copilot query, every code suggestion, and every AI-generated email draft consumes tokens that Microsoft must either absorb or pass on to customers.

DeepSeek, a Chinese AI lab known for training highly competitive models at a fraction of the cost, offers per-token pricing that is up to 90% cheaper than comparable offerings from OpenAI. For a company eyeing hundreds of millions of enterprise users with Copilot Cowork, the savings could run into billions of dollars annually. Sources indicate that Microsoft's Azure AI division has already completed preliminary benchmarks showing DeepSeek's performance nearly matching GPT-4 class models on common enterprise tasks like summarization, data extraction, and structured report generation.

Copilot Cowork: The Enterprise-First AI Agent

Copilot Cowork is Microsoft's new AI companion designed explicitly for enterprise collaboration. Unlike the consumer-facing Copilot embedded in Windows, Cowork functions as a persistent agent that can attend meetings, handle routine project management tasks, and act as an always-on knowledge repository for teams.

Scheduled for a phased rollout starting in June 2026, Cowork will be available first to Microsoft 365 E5 subscribers before trickling down to lower-tier plans. Redmond has positioned it as a competitive response to Google's Gemini for Workspace and Salesforce's Einstein GPT, both of which have been gaining ground in the enterprise AI market.

Microsoft has not publicly disclosed the model backend for Cowork, but early technical documentation reviewed by windowsnews.ai points to a multi-model architecture. The default engine is expected to be OpenAI's GPT-5, but developers have been given latitude to experiment with alternative models for specific tasks—a practice that opened the door to DeepSeek.

DeepSeek's Disruptive Pricing

DeepSeek made headlines in 2025 when it released a family of open-weight models that outperformed Meta's Llama 3 and approached GPT-4 levels on several benchmarks. Its API pricing, however, was the earthquake that shook the industry: under $0.05 per million input tokens for its flagship model, compared to $3–$10 per million tokens for comparable closed models.

For an enterprise that might process billions of tokens daily across thousands of simultaneous users, the economics are compelling. Microsoft's internal financial models, according to a leaked spreadsheet snippet, estimate that switching just 30% of Copilot inference workloads to DeepSeek could reduce operating losses from the AI push by $600 million annually by fiscal year 2027.

Azure as a Shield: Hosting DeepSeek on Microsoft's Cloud

Microsoft's proposal differs from simply plugging into DeepSeek's public API. The plan is to run the model entirely within Azure's private virtual networks, using hardware that Microsoft leases or owns in its own data centers. This architecture allows enterprise customers to enforce their own data governance policies, keep prompts and responses within designated geographies, and avoid Chinese government access to data—at least in theory.

Azure AI services already support a marketplace of third-party models, including those from Mistral, Cohere, and Stability AI. Adding DeepSeek as a "bring your own model" or even a managed offering would be technologically straightforward. Microsoft's sales teams have reportedly begun briefing select enterprise clients on the concept under non-disclosure agreements, positioning it as a cost-control measure for AI-hungry organizations.

Australia Draws a Line in the Sand

In February 2025, Australia's Department of Home Affairs issued a directive banning the installation or use of DeepSeek applications, models, and associated services on all government-issued devices and networks. The ban covers federal agencies, state governments aligned with federal cybersecurity policy, and private contractors handling sensitive data.

The directive cited the Security of Critical Infrastructure Act and referenced intelligence assessments linking DeepSeek to potential data exfiltration risks. Australian security agencies have long warned about the dangers of hosting sensitive data with entities that could be compelled by foreign governments to grant access, and DeepSeek's corporate structure in China places it squarely within that category.

Since then, the ban has only expanded. Several major Australian universities followed suit, removing DeepSeek from campus networks, and major consulting firms operating in the Australian public sector have also banned the model from their corporate devices.

A Collision Course with Government Contracts

Microsoft holds multi-year, multi-billion-dollar contracts with the Australian government, including a whole-of-government cloud agreement signed in 2023 that placed Azure and Microsoft 365 at the heart of the country's digital infrastructure. If Copilot Cowork—deeply integrated into Teams, Outlook, and Office apps—suddenly leaned on a model banned by the very government it serves, the fallout could be severe.

Legal experts in Canberra have already begun examining the implications. "If Microsoft deploys DeepSeek in a way that touches any part of the government's tenancy, even indirectly, that could constitute a breach of contract and a violation of the directive," said Dr. Emily Tran, a cybersecurity law professor at the Australian National University. "The government would have grounds to suspend or terminate the contract, and it might be obligated to do so under its own policies."

Microsoft has not publicly commented on how it would segregate DeepSeek workloads from government tenants. One technical approach would involve region-bound deployments where Azure instances serving Australian government customers are firewalled from any infrastructure running DeepSeek inference. But given the complexity of shared Microsoft 365 infrastructure, such segregation might be operationally challenging.

Enterprise Customers Weigh Risk and Reward

Beyond the government sector, reactions from enterprise customers have been mixed. windowsnews.ai spoke with IT leaders at three large Australian firms, all of whom requested anonymity to discuss confidential product briefings. Two expressed interest in the cost savings but demanded contractual guarantees that DeepSeek would be ring-fenced from their most sensitive data.

"If Microsoft can prove that data never leaves Australian Azure regions and that the model is isolated from the Chinese parent company, we'd consider it for low-risk workloads," said one CTO of a manufacturing conglomerate. "But the brand risk is enormous. The average consumer hears 'Chinese AI' and thinks 'spying.' There's a trust gap we can't ignore."

Globally, the reception may be warmer. European enterprise customers, who have been vocal about wanting cheaper AI options not dominated by US hyperscalers, could embrace an Azure-hosted DeepSeek as a sovereign alternative that still runs on compliant infrastructure. This aligns with Microsoft's decade-long strategy of positioning Azure as the cloud of choice for regulated industries.

Microsoft's Tightrope Walk

The DeepSeek consideration places Microsoft at a delicate intersection of cost pressure, geopolitical tension, and customer trust. The company has invested heavily in OpenAI but has also been diversifying its AI portfolio—partly to reduce dependency on a single supplier and partly to offer customers choice.

Satya Nadella has repeatedly emphasized that AI trust and security are the "foundation" of Microsoft's enterprise business. In a memo after the Snowflake and Wiz acquisitions, he wrote, "Our customers need AI that is not only powerful but also governed. We will never compromise on data sovereignty." A spokesperson declined to comment on the DeepSeek evaluation.

Yet the financial reality is biting. Microsoft's stock has been under pressure as investors question the return on massive AI investments. Capital expenditure on AI infrastructure hit $55 billion in fiscal year 2025, and the company's AI services, though growing quickly, have not yet turned a profit. DeepSeek represents a path to reduce the bill without sacrificing competitive capability.

The Road Ahead

The Copilot Cowork timeline—with general availability targeted for June 2026—gives Microsoft several months to navigate the technical, legal, and political challenges. In that window, the company might negotiate a special arrangement with the Australian government, implement airtight data segregation, or decide that the reputational risk outweighs the savings.

Industry watchers expect a decision by September 2025, when Microsoft finalizes the model configuration for the Cowork release candidate. If DeepSeek is adopted, it will likely be optional and disabled by default for customers in Australia and other jurisdictions with similar bans. But the mere presence of the model in a Microsoft product could trigger regulatory scrutiny.

One thing is clear: the AI landscape is no longer a simple story of American innovation versus Chinese competition. As models become commodities and governments assert more control over the infrastructure that runs them, hyperscalers like Microsoft will have to make increasingly uncomfortable choices—balancing the books, the law, and the global patchwork of trust.