Microsoft는 내년부터 자체 개발한 Maia 300 AI 칩을 기가와트 규모로 배치할 계획이며, Copilot 추론 작업도 자체 실리콘으로 옮겨갈 예정이다.

Microsoft is preparing a sweeping shift in how it powers artificial intelligence. The company announced that its newest in-house processor will begin rolling into data centers next year at a scale designed to reduce its long-standing dependence on Nvidia graphics cards.

The company unveiled the Maia 300, its third-generation AI accelerator, at the OCP Global Summit in San Jose on Monday. Executives said the chip is engineered specifically for inference, the computationally intensive work of running trained AI models to answer questions, summarize meetings and generate content. It will become the primary engine for Microsoft's own AI services.

Rani Borkar, who leads Microsoft's data center hardware group, framed the deployment as both a technical and strategic milestone. "We believe we have built the best product for these workloads," she said. "We will deploy Maia at gigawatt scale starting next year, and that is how we are backing up that confidence."

That language reflects how seriously the software giant is treating the economics of AI. Inference now represents a large and growing share of the computing load behind products like Microsoft 365 Copilot and GitHub Copilot. The company says the Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems, allowing the same amount of electricity to support significantly more AI work. A single rack packed with 96 Maia 300 chips, double the density of its predecessor, can push through up to seven times more tokens than a Maia 200 rack.

"We need new architectures that give us better performance and better use of the power we have," said Alistair Speirs, a general manager at Microsoft. "Power translates pretty directly into operational cost for a data center."

Microsoft has been building toward this moment since 2023, when it introduced the original Maia 100. The Maia 200 followed this past January and is already running production workloads across data centers in the United States, Europe and Australia. It supports OpenAI models, Microsoft's own MAI models and several open-source model families, powering AI features in Teams, Copilot and other products.

Chief Executive Satya Nadella told analysts in July that the Maia 200 delivers 30% better performance per dollar than the latest hardware in Microsoft's fleet. Finance Chief Amy Hood said at a Goldman Sachs conference last month that the use of in-house silicon has contributed to margin improvement over the past decade.

Mass production of the Maia 300 is scheduled to begin in 2027, according to a blog post from Executive Vice President Scott Guthrie. Speirs declined to give a precise timeline for when the multi-gigawatt deployments would be complete. "We reserve a little bit of optionality there, because this is an industry that is changing very fast," he said. "We want to deploy the right silicon for the right models at the right time."

**A Different Path From Rivals**

Microsoft's approach diverges sharply from how Amazon and Google have commercialized their custom silicon. Both cloud providers rent access to their in-house AI chips to outside customers, creating a direct revenue stream and giving developers alternatives to Nvidia hardware. Amazon has gone further, exploring sales of chips for customers to install in their own data centers.

Microsoft has no such plans for Maia. The chips will serve the company's own sprawling AI operations rather than being offered as hourly rentals. Borkar put it bluntly: "Our own AI services business is already large enough. We want to serve ourselves first."

The decision means Maia's value to investors will hinge less on direct chip sales or rental income and more on whether it can lower the cost of running AI services at scale. If Microsoft can process more tokens per megawatt, it can either serve more customers with the same infrastructure or improve profit margins on existing AI products.

Analysts at Evercore ISI estimated in a note to clients on Wednesday that Maia will account for roughly 10% of Microsoft's AI chip spending by the first quarter of 2027. The firm recommends buying Microsoft shares.

Patrick Moorhead, CEO of advisory firm Moor Insights and Strategy, noted that the chips remain untested outside Microsoft's own environment. "Nobody has tested them outside of Microsoft yet," he said.

**Nvidia's Role Isn't Disappearing**

Even as Microsoft pushes deeper into custom silicon, Nvidia remains central to its AI strategy. Nvidia GPUs still fill Microsoft's existing data centers and power the training of models from OpenAI and Anthropic. Microsoft also continues to partner with Nvidia and AMD for training workloads, Speirs said.

Nvidia Chief Executive Jensen Huang appeared at a Microsoft event last week where the company announced a Surface laptop containing an Nvidia processor. For general-purpose computing, Microsoft uses its own CPUs alongside chips from Intel and AMD.

The competitive landscape, however, is unmistakably shifting. Hyperscalers including Microsoft, Amazon, Google and Meta, along with AI labs OpenAI and Anthropic, have all either deployed or are developing custom accelerators. For Nvidia, which regularly derives more than half its quarterly revenue from hyperscaler customers, the trend represents a structural challenge.

The Maia 300's efficiency claims are aimed squarely at that market. By positioning the chip as 40% more power-efficient than leading GPU-based systems for inference, Microsoft is signaling that it sees custom silicon as a viable way to control costs as AI usage continues to expand.

For now, Microsoft's strategy is inward-looking: build the chips, deploy them at enormous scale, and let the efficiency gains flow through to its own products and margins. Whether that approach eventually shifts to include external customers remains an open question, one the company is in no hurry to answer.