Microsoft just took a swing at the company that sells it the AI hardware. Its new Maia 300 chip is a direct shot at its biggest supplier: on Oct. 12, the company unveiled the third generation of its in-house AI accelerator and said it will scale the family to multi-gigawatt deployments starting in 2027.
Here’s the part that should make Nvidia’s investors sit up. Microsoft says Maia will become the primary inference system for its own products and services. Satya Nadella called it “our biggest bet yet on our own silicon,” in a post on X.
What the Maia 300 Chip Is Actually Claiming
Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems. That’s a big number. It’s also Microsoft’s own projection, not an independent benchmark (a company grading its own homework is a time-honored tradition, but still).
Per a summary of the company’s blog post, a full 96-accelerator configuration is meant to deliver up to seven times the rack-scale throughput of the current Maia 200. Again, those are the company’s figures. Nobody outside Redmond has tested them yet.
The chip is also built to cut down on unnecessary data movement between the processor and its memory. In plain English, it wastes less energy shuffling data around, which matters a lot when power is the real bottleneck. Elon Musk’s camp sees the same squeeze, as SpaceX’s push to build its own turbine parts shows.
Why Every Big Nvidia Customer Is Now a Rival
Microsoft is far from alone. Axios lays out the pattern: Nvidia’s largest customers are all building their own silicon.
- Google has its TPUs.
- Amazon has Trainium, which Anthropic uses.
- OpenAI is working with Broadcom on a chip called “Jalapeño.”
- Meta runs an internal chip program.
- Anthropic has confirmed it is exploring its own designs.
The target is the same for nearly all of them: inference, the stage where a trained model actually answers questions. Axios notes that inference is where margins are highest. Nvidia still dominates training.
That’s the real logic. Training a model is a giant one-time bill. Inference is the meter that never stops running, so even small efficiency gains add up fast. That’s why in-house AI silicon has become the hottest project in Big Tech.
Here’s the Thing: This Is Not a Nvidia Obituary
Before anyone writes Nvidia’s eulogy, look at the numbers. Axios reports that Nvidia’s revenue more than doubled to $96 billion in the second quarter. Overall demand for AI compute leaves room for both Nvidia and the in-house chips.
In fact, Microsoft says as much. “We prioritize Maia for inference,” Rani Borkar, who runs Azure Hardware Systems and Innovation, told Axios. “For us, it’s not an either-or strategy.”
There’s another catch. Custom-chip makers still rely on the same scarce manufacturing capacity, advanced packaging and high-bandwidth memory that Nvidia does. Everyone is shopping in the same crowded store, and the shelves are thin.
And Microsoft’s chip is a private club. Unlike Google’s and Amazon’s chips, Maia isn’t available for outside customers to rent, GeekWire reports. It exists to run Microsoft’s own stuff.
The Azure Maia Track Record So Far
The first Maia arrived in 2023. Maia 200 followed in January 2026 after delays, and as of this summer it reportedly ran in just two U.S. data centers, per GeekWire’s citation of The Information. Now Microsoft says Azure Maia is deployed across the U.S., Europe and Australia.
The chip is already serving Microsoft Copilot and supporting the company’s own model development. GitHub Copilot and Microsoft Foundry are due to follow by the end of 2026. Meanwhile, Microsoft has other fights on its plate, including Vance’s green card suspension threat over H-1B use.
So the real question is simple: can Microsoft turn a bumpy start into multi-gigawatt scale by 2027?
Nvidia will be watching, probably while counting its $96 billion quarter.
