Microsoft wants to run its own AI on its own silicon. The company unveiled Maia 300 on Monday, its third-generation custom AI chip, and said it plans to put potentially more than one million of them to work across its infrastructure in 2027.

That’s not a typo. One million chips, in multi-gigawatt installations, all bought from nobody.

The announcement, covered by Reuters, CNBC, and Microsoft’s own news center, marks the most aggressive step yet by a hyperscaler to loosen Nvidia’s grip on the AI compute market. From 2027, Microsoft says Maia will become the primary inference system for its products and services, meaning the Copilots, the Azure AI workloads, and even some OpenAI models running on Microsoft’s cloud will increasingly run on chips Microsoft designed itself.

Built for inference, not training

Here’s the important distinction: Maia 300 isn’t trying to out-train Nvidia’s best GPUs. It’s built for inference: the unglamorous, high-volume business of actually running AI models once they’re trained. Every time someone asks Copilot to summarize a document or GitHub Copilot to finish a line of code, that’s inference. And inference, not training, is where most of the money now gets spent.

Microsoft’s executive vice president Scott Guthrie said Maia 300 uses an inference-first architecture optimized for long-context reasoning, memory-heavy work, and decode-heavy workloads, the exact profile of agentic AI tasks where models call tools, reason over long documents, and hold conversations.

Rani Borkar, Microsoft’s president of Azure hardware systems and infrastructure, told Reuters the company is tuning the whole system around efficiency: the goal, she said, is better performance for every watt, dollar, and unit of memory.

What’s inside the chip

Maia 300 breaks with the “one giant chip” approach. It uses a chiplet design, with specialized components stitched into a single package — with the key compute chiplets manufactured on TSMC’s 2-nanometer process and supporting components on a cheaper 3-nanometer node.

The memory story is arguably more impressive than the compute. Each chip carries eight stacks of HBM4 high-bandwidth memory: 288 gigabytes total at 18 terabytes per second of bandwidth. That matters because inference on large models is mostly a memory problem: getting the model weights in and out fast enough — and Microsoft is betting that feeding memory faster beats raw floating-point muscle.

At the rack level, the numbers look like this: 96 chips per rack drawing about 167 kilowatts. Microsoft claims seven times the rack-scale token throughput of the previous Maia 200, and roughly 40% more tokens per megawatt than leading GPU-based systems. The company declined to name which systems it compared against, so take the 40% figure the way you’d take any vendor benchmark: interesting, but unverified by anyone outside Redmond.

For context, Nvidia’s competing GB300 racks hold 72 GPUs and draw around 142 kilowatts. Microsoft is packing more chips per rack and spending more power per rack, but claims to squeeze more tokens out of every megawatt.

A family, not a one-off

Maia 300 didn’t arrive out of nowhere. Maia 100 debuted in 2023; Maia 200 was announced this January and already powers live services like Microsoft 365 Copilot, GitHub Copilot, Microsoft Foundry, and OpenAI models hosted on Azure. The 200, which is already running production workloads, is scheduled to reach the rest of the Copilot fleet and Foundry by the end of this year. Microsoft’s own chips are prioritized for its internal AI services, not sold or rented as external capacity the way AWS rents out Trainium or Google rents out TPUs. Back in August, The Information reported Microsoft was negotiating for more than 300,000 Maia 300 chips for 2027. The new figure of over a million suggests those talks, and the ambition, have grown.

And the money behind it is staggering. Microsoft expects its capital spending to pass $175 billion in calendar 2026. In the most recent quarter alone it spent $41 billion, with roughly two-thirds going to shorter-lived assets, mostly CPUs and GPUs. Maia 300 is, at heart, a cost-control project: a way to bend that curve.

Why this matters

Every hyperscaler is now making its own chips for the same reason: Nvidia’s margins are someone else’s costs. Google has TPUs, Amazon has Trainium and Inferentia, and Meta is on its third generation of training-and-inference chips. Microsoft was late to silicon and has leaned hardest on Nvidia. Maia 300 is its attempt to catch up fast.

Don’t read this as Microsoft divorcing Nvidia. A million chips sounds enormous until you remember the size of Microsoft’s fleet: that $175 billion capex budget still has Nvidia’s name written all over it. The real play is bargaining power. Every workload Microsoft moves to its own chips is margin it keeps, and a stronger hand in its next GPU deal.

The deeper signal is about where AI computing goes next. For years, the industry obsessed over training: bigger models, bigger clusters, bigger headlines. Maia 300 is a bet that the next decade of AI economics will be decided at inference time: cheaper, more efficient, running everywhere. Whoever controls the cheapest token wins the agentic era.

FAQ

What is the Microsoft Maia 300 AI chip? Maia 300 is Microsoft’s third-generation custom AI accelerator, built for AI inference, running trained models for Copilot, GitHub Copilot, and OpenAI models on Azure. It uses a chiplet design with TSMC 2-nanometer compute components and was announced on October 12, 2026.

How many Maia 300 chips will Microsoft deploy? Microsoft plans to deploy potentially more than one million Maia 300 chips in 2027 across multi-gigawatt installations, making Maia the primary inference system for its products and services from 2027 onward.

Does the Maia 300 replace Nvidia GPUs at Microsoft? Not entirely. It reduces dependence rather than eliminating it. Microsoft says the chips will handle most of its internal AI services, and the company’s $175 billion in planned 2026 capital spending still includes large GPU purchases. The goal is cost control and negotiating power, not a full breakup with Nvidia.

How efficient is the Maia 300? Microsoft claims about 40% more tokens per megawatt than leading GPU-based systems, though it hasn’t named the systems compared. A rack of 96 Maia 300 chips draws around 167 kilowatts and delivers seven times the rack-scale token throughput of the previous Maia 200, according to the company.

Sources: Reuters, CNBC, Microsoft News Center