Microsoft's Maia 300 chip cuts its reliance on Nvidia GPUs

The News

Microsoft has unveiled Maia 300, the third generation of the in-house accelerator it is positioning as the "primary AI inference system" for its own products and services, in a move designed to pare back its dependence on Nvidia and AMD. The chip will begin rolling out to production AI workloads in 2027.

The headline claim came from Scott Guthrie, executive vice president of Microsoft's cloud group. He said the company expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration, Guthrie added, it reaches up to 7x the rack-scale token throughput of its predecessor, Maia 200.

Maia 200, announced only in January, already runs live services including Microsoft Copilot, GitHub Copilot, Microsoft Foundry and the company's own models. The new part is built on TSMC's N2P process node and uses a chiplet design that splits compute, memory controller and I/O onto separate silicon. Guthrie said Azure would scale the Maia family to multi-gigawatt deployments from 2027.

Why It Matters

Custom silicon is how the hyperscalers are trying to escape the margin squeeze of the Nvidia era. Every token a reasoning model or an agent produces costs power, and power has become the binding constraint on AI growth. Microsoft framing Maia around tokens-per-megawatt rather than raw benchmarks is telling: the contest has shifted from who owns the fastest chip to who can serve inference most cheaply at gigawatt scale.

The playbook echoes Amazon's Trainium and Inferentia and Google's TPU line, first shipped back in 2016, both built to blunt Nvidia's pricing power. What is different here is intent. Microsoft AI chief Mustafa Suleyman said the company has to "work across the full stack" down to the silicon, tying the chip to its ambition to catch Google DeepMind, OpenAI and Anthropic on model quality. Owning the hardware lets Microsoft co-design chips and models together, something a merchant GPU buyer cannot.

Indian Angle

For Indian technology buyers, this is a story about the price of a token. Azure runs some of its largest regions out of Pune, Chennai and Hyderabad, and Indian SaaS and AI startups increasingly rent inference there in dollars. If Maia 300 cuts the energy cost of serving a model by anything close to Guthrie's claim, those savings should eventually reach the cloud list prices Indian developers pay in a softening rupee. Cheaper inference is, for them, as much a forex story as a technology one.

It also sharpens the contrast with India's own compute build-out. Homegrown model makers such as Sarvam and Ola's Krutrim still lean heavily on imported Nvidia GPUs, the very hardware Microsoft is now designing around. India's answer so far is the IndiaAI Mission's shared GPU pool and the India Semiconductor Mission's fabs at Dholera and Sanand, but those plants will begin on mature nodes, far from the TSMC N2P process behind Maia 300.

There is a talent dimension too. Microsoft employs thousands of engineers across Hyderabad, Bengaluru and Noida who work on Azure infrastructure and model engineering. A deeper in-house silicon programme tends to pull more high-end chip-and-systems work towards the hubs that already host it.

FAQ

When will Maia 300 be available?

Microsoft says the chip begins rolling out to production AI workloads in 2027, with the Azure Maia family scaling to multi-gigawatt deployments from that year. It succeeds Maia 200, announced in January 2026, which already runs live Microsoft services.

How much more efficient does Microsoft claim it is?

The company expects 40 percent more tokens per megawatt than leading GPU-based systems, and up to 7x the rack-scale token throughput of Maia 200 in a 96-accelerator configuration. These are vendor figures, not independent benchmarks.

Does this mean Microsoft is dropping Nvidia?

Not immediately. The efficiency gains are meant to let Microsoft rely mainly on its own chips over time rather than Nvidia and AMD, but the shift starts with the 2027 rollout and will run alongside existing GPU fleets.

What does it mean for Indian cloud customers?

Lower inference costs on Azure could, over time, ease the dollar-denominated compute bills Indian startups and enterprises carry, though any benefit depends on how far Microsoft passes the savings on through pricing.

Where can I read the original announcement?

The Verge's report by Tom Warren carries Microsoft's full set of claims and quotes. The link appears in the attribution below.

This story was reported by The Verge. Read the full original coverage at The Verge.