Microsoft Eyes DeepSeek V4 for Copilot Cowork, Squeezing Out OpenAI and Anthropic on Cost
Microsoft is weighing a self-hosted version of DeepSeek V4 for Copilot Cowork, citing token cost gaps as high as 57x versus Anthropic and OpenAI. The move accelerates Microsoft's multi-model strategy — but faces US regulatory risk over the model's Chinese origin.
- Microsoft is evaluating DeepSeek V4 for Copilot Cowork to cut token costs by up to 57x versus Anthropic and OpenAI, with plans to self-host the model on Azure
- The Trump administration's potential restrictions on Chinese AI models remain the primary risk to the plan
Microsoft is weighing the integration of China's DeepSeek V4 model into Copilot Cowork, its enterprise AI agent solution within Microsoft 365. The move would partly displace current model suppliers OpenAI and Anthropic. Token cost gaps as wide as 57x are forcing a rethink of model selection across the enterprise AI market.
DeepSeek V4 Pro at $0.87 Per Million Tokens vs Anthropic's $50
Copilot Cowork currently runs on OpenAI and Anthropic models inside Microsoft 365. The cost pressure is real: Anthropic's latest flagship model costs approximately $50 per million tokens while DeepSeek V4 Pro comes in at roughly $0.87 per million — about 57x cheaper. Both OpenAI and Anthropic have abandoned unlimited flat-rate pricing in favor of usage-based billing. What analysts are calling 'tokenmaxxing' pressure is now pushing Microsoft toward model substitution.
Microsoft Plans Self-Hosted, Fine-Tuned DeepSeek on Azure
Microsoft's plan is to run a modified, self-hosted version of DeepSeek V4 exclusively within Microsoft Azure — not on Chinese servers. Customer data would remain inside Microsoft's cloud and be subject to Azure's enterprise security, compliance, and data-residency controls. The company says it has fine-tuned the model and applied safeguards including bias reduction. If deployed, DeepSeek V4 would be an optional choice for enterprise customers.
Nadella: 'A Few AI Players Dominating Would Be Harmful' — Multi-Model Push Accelerates
Microsoft CEO Satya Nadella recently published an essay arguing that 'allowing just a few major players in the AI space to dominate would ultimately be harmful for the economy and society.' The DeepSeek evaluation is a direct extension of Microsoft's strategy to reduce dependence on any single frontier model provider — particularly OpenAI following their public relationship friction. Azure AI Foundry already offers a catalog spanning OpenAI, DeepSeek, Meta, Hugging Face, Grok, Mistral, and Microsoft's own model families. Anthropic models are already embedded in parts of the Microsoft 365 Copilot stack.
Trump Administration's China AI Scrutiny — DeepSeek Ban Risk Flagged
The biggest variable is geopolitical. The Trump administration has increased scrutiny on Chinese AI models and has floated the idea of a DeepSeek ban. It also recently restricted foreign access to Anthropic's latest technology on national security grounds. Should Microsoft proceed with DeepSeek integration, friction with the administration appears likely. The fundamental tension — cost reduction versus geopolitical risk — means the final decision remains unresolved.
Related Stocks & ETFs
US Stocks Microsoft (NASDAQ: MSFT) — Copilot Cowork operator; potential AI cost reduction from DeepSeek integration NVIDIA (NASDAQ: NVDA) — AI GPU supplier; watching DeepSeek's GPU efficiency impact on chip demand Alphabet (NASDAQ: GOOGL) — Google Gemini-based AI competitor ETFs QQQ — Invesco NASDAQ-100 ETF BOTZ — Global X AI & Robotics ETF
Frequently Asked Questions
How does DeepSeek integration affect Microsoft's stock?
It is broadly positive. Lower AI model costs improve Copilot Cowork operating margins and could enhance enterprise pricing competitiveness. The key risk is that US regulatory action against Chinese AI models could force Microsoft to abandon the plan.
How is OpenAI likely to respond to Microsoft evaluating DeepSeek?
Microsoft and OpenAI's relationship has already faced public tension. Microsoft's multi-model strategy is explicitly designed to reduce OpenAI dependence — DeepSeek is an extension of that push, not an isolated event.
Does hosting DeepSeek on Azure resolve the security concern?
Microsoft argues customer data stays within Azure and never touches Chinese servers. However, the Trump administration's concern appears to be the model itself — its Chinese origin — regardless of hosting location. Regulatory risk remains even with Azure deployment.
How is the 57x token cost gap calculated?
Enterprise AI models are billed by the number of text tokens processed. Anthropic's latest model charges roughly $50 per million tokens while DeepSeek V4 Pro costs approximately $0.87 per million — a roughly 57x difference for equivalent token volume (specific conditions may vary).
What does this mean for NVIDIA's GPU demand outlook?
DeepSeek is known for achieving competitive performance with lower GPU resources than OpenAI models. If DeepSeek adoption grows, it could moderate per-workload GPU consumption. However, rising total AI usage volumes may offset per-model efficiency gains, keeping absolute GPU demand intact.
3 Dividend Stocks to Reduce Market Crash Risk: Medtronic, Realty Income, ExxonMobil
MDT
O
XOM+2
Top High-Yield Dividend Stocks: Brookfield Infrastructure, Clearway Energy, Enterprise Products
BIP
BIPC
CWEN+3
Goldman Sachs Raises Tesla Q2 2026 Delivery Forecast to 420K — Europe Up 85%, US Down Double Digits
TSLA
RIVN
LCID+4Wall Street's Defining Moments
See more
Smart Money Briefing
Weekly summaries of Wall Street guru moves and crypto whale activity.
