On March 9, 2026, Microsoft launched Copilot Cowork, a redesign of its flagship AI product built on Anthropic's Claude Cowork technology. In January 2023, Microsoft invested $10 billion in OpenAI for what was supposed to be an exclusive partnership. The journey from that bet to this product tells the story of the fastest commoditization in enterprise software history.
The Thesis
Microsoft's OpenAI investment was built on a simple premise: exclusive access to the best model gives you the best product. In 2023, that was a defensible bet. GPT-4 was the undisputed frontier. No one else was close. Microsoft wired OpenAI's models into everything — Bing, Office, GitHub, Azure — and announced Copilot as the future of productivity software.
The total investment eventually reached $13 billion. For that, Microsoft got priority access to OpenAI's models, the right to resell them through Azure, and what appeared to be an insurmountable head start in enterprise AI.
The Crack
The first sign came not from a competitor but from OpenAI itself. In November 2023, the board fired Sam Altman. For 72 hours, Microsoft's entire AI strategy depended on the outcome of a boardroom fight it didn't control. The Financial Times later reported that Microsoft began diversifying its AI investments immediately after the dispute — the dependency was suddenly visible.
What followed was a systematic unwinding of exclusivity, measured not in press releases but in product decisions:
-
DEC 2024Sources: Microsoft is working on adding more first- and third-party AI models to 365 Copilot and other products.
-
MAR 2025Microsoft is testing ways to replace OpenAI's models in Copilot with models from xAI, Meta, and DeepSeek. Separately, Microsoft has completed training its own family of models, codenamed MAI.
- SEP 2025 Sources say Microsoft will use Anthropic's models for some AI features in its products alongside OpenAI.
-
NOV 2025Anthropic commits to buy $30 billion in Azure capacity; Microsoft invests up to $5 billion in Anthropic. Mustafa Suleyman's team starts building Microsoft's own models.
- JAN 2026 Microsoft is spending nearly $500 million per year on Anthropic's API. It is now one of Anthropic's top clients.
-
FEB 2026Suleyman tells the Financial Times that Microsoft is pursuing "true self-sufficiency" in AI. The same week: Copilot's paid user share as first choice has fallen from 18.8% to 11.5%.
Every step moved in one direction: away from OpenAI as the sole foundation of Microsoft's AI products.
The Number
The Wall Street Journal reported in February that Copilot's position as paid users' primary AI tool had fallen nearly 40% in six months. The headline said it plainly: "Microsoft's Pivotal AI Product Is Running Into Big Problems." Gemini, which Google bundles into Workspace, rose from 12.8% to 15.7% over the same period.
The model was supposed to be the differentiator. GPT-4 was supposed to make Copilot the default. Instead, the best model in the world produced a product that was losing ground to a competitor running a different model. The variable that mattered wasn't which LLM powered the product — it was how well the product integrated into work.
The Swap
Today's Copilot Cowork launch makes this explicit. Microsoft didn't just add Claude as an option. It integrated Anthropic's Claude Cowork technology as a core component of Microsoft 365 Copilot, alongside a new layer called Work IQ that grounds the AI's actions in the user's actual work data — emails, documents, calendars, meetings.
The architecture reveals the thesis shift. The old Copilot was "GPT-4 inside Microsoft products." The new Copilot is "Microsoft's work graph powering whatever model is best." The model became a component. The integration layer became the product.
Three days ago, Microsoft publicly stated it would keep Anthropic's AI tools embedded in its client products even as the Pentagon moved to designate Anthropic a supply chain risk under the Defense Production Act. Google and Amazon made similar declarations the next day. Today, Anthropic sued the Department of Defense to block the designation.
Microsoft deepened its Anthropic integration — launching a product built on Claude — on the same day Anthropic entered open legal conflict with the U.S. government. Whether or not the timing was deliberate, the product architecture speaks for itself: a system designed to swap models can survive any single supplier being cut off. The Pentagon's attempt to designate Anthropic a supply chain risk matters less when your product doesn't depend on any single model.
The Lesson
Microsoft spent $13 billion to learn that the model isn't the moat. The moat is everything the model plugs into.
"True self-sufficiency" — Suleyman's phrase — doesn't mean building the best model. It means building a product layer so valuable that you can swap the model underneath without users noticing. That's what Copilot Cowork is: the moment the swap became visible in the product itself.
In January 2023, Microsoft paid for exclusive access to GPT-4 because it believed the model was the scarce resource. By March 2026, it rebuilt the same product on a competitor's model because it learned the scarce resource was something else entirely: the enterprise context that makes any model useful.
The $13 billion bought time. It did not buy a moat.