Microsoft Will Run Its Own AI On Its Own Chips From 2027

In this storyMicrosoftNVIDIA

Microsoft has unveiled the Maia 300, the third generation of its in-house AI accelerator, and said it plans to deploy potentially more than a million of them next year in multi-gigawatt installations.

The chip will become the primary inference system for Microsoft's products and services from 2027, reducing the company's reliance on buying processors from Nvidia. It is aimed at running models rather than training them.

Executive vice president Scott Guthrie said Maia 300 is built with an inference-first architecture optimised for long-context, memory-intensive and decode-heavy workloads, and that a 96-chip rack delivers seven times the rack-scale token throughput of Maia 200. Microsoft expects about 40 per cent more tokens per megawatt than leading GPU-based systems, without naming them.

Each rack of 96 chips draws around 167 kilowatts, against 142 kilowatts for Nvidia's GB300 racks, which hold 72 GPUs. The processor uses chiplets, with key compute components built on TSMC's 2-nanometre process. The Information reported in August that Microsoft was negotiating for more than 300,000 Maia 300 chips for 2027.

Maia 100 arrived in 2023 and Maia 200 was announced in January this year, now supporting production workloads including Microsoft 365 Copilot, GitHub Copilot, Microsoft Foundry and OpenAI models running on Azure. At launch Microsoft said Maia 200 delivered three times the FP4 performance of third-generation Amazon Trainium and FP8 performance above Google's seventh-generation TPU.

The chips are prioritised for Microsoft's internal AI services rather than rented out as external capacity.

The metric Microsoft chose to lead with tells you what the fight is now about.

Not teraflops, not transistor count, but tokens per megawatt. Microsoft claims roughly 40 per cent more of them than leading GPU-based systems, and the comparison it pointedly does not name is Nvidia's. That framing is an admission that compute is no longer the scarce input. Power is.

The rack figures make the point more honestly than the headline does. Ninety-six Maia 300 chips draw about 167 kilowatts against 142 kilowatts for an Nvidia GB300 rack holding 72 GPUs, so Microsoft's rack uses more power, not less. The claim is that it produces disproportionately more output for it. Whether that holds outside Microsoft's own benchmarks is unknowable from here, and vendor efficiency numbers are selected to flatter.

But the direction is right regardless of the precise figure. Data centre expansion is now constrained by grid connections and local opposition rather than by chip supply, which is why Amazon spent last week abandoning NDAs and promising $1 billion to host communities. In that environment, the useful question stops being how many chips you can buy and becomes how much work each megawatt does.

A company that controls the silicon can optimise for that directly. One buying merchant GPUs optimises for whatever the vendor built. Over a multi-gigawatt estate, the difference compounds into something worth designing a chip for.

Three of the four largest buyers of AI compute now design their own, and that changes what the demand signal under this industry actually is.

Google has TPUs, including the four it put into orbit this month. Amazon has Trainium. Microsoft now has Maia at scale. Each was built for the same reason, which is that renting inference capacity from a supplier with 75 per cent margins is intolerable once your own volumes are large enough to amortise a design team.

What that leaves Nvidia is training, where its position remains close to unassailable, plus every customer too small to build silicon. That is still an enormous business. It is a different one from the assumption embedded in the supply chain, which is that hyperscaler capital expenditure converts into GPU purchases at a stable rate.

For the companies downstream of that assumption, the read is mixed. TSMC wins either way, since it fabricates both. Memory makers win either way too, because an inference chip optimised for long-context workloads is memory-hungry by construction, which is part of why Micron expects shortage into 2028.

The firms with most to lose are the neoclouds. CoreWeave and its peers exist to rent GPUs to people who cannot buy them, and their largest potential customers are now building the alternative. Firmus pulling a $5 billion listing last week, after investors declined to fund AI infrastructure at a tripled valuation, sits in the same picture.

For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.