Microsoft Plans Multi-Gigawatt Scale-Up of Custom AI Chips

Microsoft is preparing to scale its custom AI chips across a multi-gigawatt buildout, aiming to cut the cost of running artificial intelligence workloads that currently depend on Nvidia's expensive accelerators. That single word — multi-gigawatt — is the hardest number in the story, and it frames everything else: this is not a pilot project. It's a plan to power entire datacenter campuses with silicon Microsoft designs itself.

The company has already placed its first bets. Its Maia accelerator family, first announced in November 2023 alongside the Cobalt CPU, was built to handle AI training and inference inside Azure. Now, according to the report, Microsoft intends to push those designs into deployments measured in gigawatts of power draw rather than racks.

Why does custom silicon matter to Microsoft's costs?

Nvidia's GPUs dominate AI compute, and they dominate Microsoft's capital spending too. Every point of margin that flows to a chip supplier is a point that doesn't stay with the cloud operator. Designing chips in-house flips that equation. Microsoft pays for design and fabrication rather than paying Nvidia's margin on every unit.

The scale of the ambition is what stands out. A single gigawatt of dedicated AI compute represents billions of dollars in datacenter infrastructure before you count the chips themselves. Multi-gigawatt means Microsoft expects its own silicon to carry a meaningful share of that load, not sit at the edge of it.

The report ties the push directly to cost reduction, not performance bragging rights. That's a commercial signal. Microsoft's AI business lives or dies on the spread between what enterprises pay for Azure AI services and what Microsoft pays to run them. Custom chips widen that spread.

How does this fit the broader custom-silicon race?

Microsoft isn't alone in pulling design work in-house. Google has run its TPU program for roughly a decade. Amazon builds Trainium and Inferentia for AWS. Meta has its own accelerator efforts. Each hyperscaler has reached the same conclusion: at their scale, general-purpose procurement from Nvidia leaves too much money on the table.

For the semiconductor supply chain, the implications run in two directions. Custom designs mean more business for foundries and packaging partners able to serve hyperscaler volumes. They also mean a slower growth curve for Nvidia inside its largest customers — though demand for AI compute remains large enough that both can grow at once.

What happens next?

Microsoft hasn't published a capacity target or a deployment date in the report, so treat the multi-gigawatt figure as a stated direction of travel rather than a confirmed build schedule. What's clear is the intent: more Maia-generation silicon, in more Azure datacenters, at a scale aimed squarely at the cost line — and at the margins Nvidia currently collects from its biggest cloud customer.