Microsoft to Scale Maia Chip to Gigawatts by 2027

Microsoft announced on Monday, November 12th, that it will take its Azure Maia accelerator family to multi-gigawatt deployments starting in 2027, with the Maia 300 as the centerpiece. According to a blog post signed by Scott Guthrie, Executive Vice President of Cloud + AI, the chip will become the primary inference system for Microsoft's own products and services. The current generation, the Maia 200, is already being installed in data centers across the United States, Europe, and Australia.

The number that the company chose to promote this decision is efficiency, not raw performance. "We expect the Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems," wrote Guthrie. Microsoft did not specify which system it was measured against or the specific workload. This is an engineering target without independent benchmarking and should be read as such.

Why Tokens per Megawatt

This metric reveals where the bottleneck lies. Energy has become the scarce resource for hyperscalers, and those who extract more inference from each contracted megawatt reduce the marginal cost of each response from Copilot or each agent running on Azure. According to reports on the announcement, a rack of Maia 300 consumes about 167 kilowatts, several times the density of a typical corporate rack, which necessitates redesigning entire rooms.

The chip architecture is also changing. The Maia 300 adopts a chiplet design, with computing blocks fabricated using TSMC's N2P process in the 2-nanometer family. This is the same queue of capacity contested by Apple, Nvidia, and AMD in Hsinchu, Taiwan, positioning Microsoft as a relevant customer of the foundry's most advanced node starting in 2027.

The Message to Nvidia

Microsoft is not abandoning GPUs. Bloomberg reported that Nvidia continues to play a role in the company’s server park, and training frontier models still relies on third-party hardware. What changes is inference, which currently represents the fastest-growing segment in the computing bill of any cloud provider.

Microsoft is arriving late to this shift. Google has been utilizing its TPUs for nearly a decade, and Amazon is pushing Trainium for customers like Anthropic. Microsoft comes in with volume measured in gigawatts, not experimental racks, and with the captive demand from products like Microsoft 365 Copilot to absorb the capacity. For Nvidia, the loss won't appear in the next quarter but in the trading margins: a customer with an internal alternative pays less for the next generation.

What Changes for Cloud Buyers

For the CIO, the direct impact is less visible than the headline suggests. The Maia will become the primary inference system for Microsoft's services. However, the inference cost of Copilot, Azure OpenAI, and corporate agents will now depend on silicon that Microsoft controls, allowing for pricing that Nvidia cannot dictate.

The effect extends across multiple markets. In the United States, where most of the Azure installed base resides, the competition is for energy and network connection. In Taiwan, the order reinforces the concentration of 2-nanometer node capacity at TSMC, a geopolitical risk that boards are already treating as an agenda item. In Europe, where the Maia 200 is already operational, self-sufficiency weighs into the conversation about data sovereignty, as it reduces dependency on a single GPU supplier for services that governments and banks contract.

For consulting firms that resell and implement Azure, the shift is one of accounting. Generative AI projects have been budgeted thus far based on rented GPU prices; a layer of inference with Microsoft's own cost alters the comparison with AWS and Google Cloud, which already operate proprietary chips. The total cost of ownership analysis done in 2025 will age quickly.

The caveat lies with the company itself. Microsoft speaks of 2027 for the gigawatt scale, and semiconductor timelines at TSMC's most advanced node frequently slip. Until then, the 40% is a promise without a named comparison, and the market will hold the number against a real GB300, not against a generic GPU.