The Maia 300 is Microsoft’s third-generation AI chip, and it’s the company’s plan to run most of its AI services on its own silicon starting in 2027. Microsoft announced it on October 12, 2026, alongside a promise to scale its Maia systems to multi-gigawatt deployments. You can’t buy or rent one yet, but it affects how Copilot gets built and run.
Here’s the short version before the details.
- Announced October 12, 2026, by Microsoft cloud chief Scott Guthrie
- Production deployments begin in 2027, not this year
- Microsoft claims 40% more tokens per megawatt than leading GPU-based systems
- Its predecessor, Maia 200, already serves Microsoft Copilot today
- Not available to rent or buy, and no independent benchmarks exist yet
That last point matters most when you read the big numbers below.
What the Maia 300 actually is
Maia is Microsoft’s in-house line of AI accelerators, built for inference. That’s the work of answering your prompt, not training the model. Every time Copilot writes a reply, a chip like this does the heavy lifting.
According to Unite.AI’s coverage of the announcement, Microsoft calls it a “complete intelligent system” aimed at long-context reasoning, fast conversation, and multi-step agent tasks. It uses a multi-chiplet design on TSMC’s N2P process.
Each chip carries eight HBM4 memory stacks with 18.0 TB/s of peak bandwidth, inside a 950 W power budget. Memory bandwidth is the number that tends to decide how quickly a chip can feed a large model.
Maia 300 vs Maia 200: how much changes
Maia 200 launched in January 2026 and is already doing real work. The table compares the two using figures Microsoft has published.
| Spec | Maia 200 | Maia 300 |
|---|---|---|
| Status | In production, serving Microsoft Copilot | Production deployments start in 2027 |
| Chip process | TSMC 3 nm | TSMC N2P |
| Memory | 216 GB HBM3e at 7 TB/s | Eight HBM4 stacks, 18.0 TB/s peak |
| Power (TDP) | 750 W | 950 W |
| Rack-scale claim | Baseline | Up to 7x Maia 200 token throughput (96 accelerators) |
The jump in bandwidth and the move to a newer process are real changes. The 7x figure is trickier, because part of it comes from packing more chips into a rack, not only from a better chip.
How much is Microsoft claiming, and who can check it?
Microsoft says Maia 300 should deliver 40% more tokens per megawatt than leading GPU-based systems. It also says a full 96-accelerator rack reaches up to 7x the token throughput of a Maia 200 rack.
These are Microsoft’s own numbers. The company hasn’t named the GPU systems it compared against, and the 7x and 40% figures use different baselines, so you can’t stack them. Maia isn’t rentable by the hour, so outside testers have nothing to run yet.
Treat the claims as a direction of travel, not a verdict. We’ll know more when Microsoft shares details at Ignite 2026, where it has said it will go deeper.
What it means for Copilot and everyday users
Maia 200 already serves Microsoft Copilot, and Microsoft says GitHub Copilot and Microsoft Foundry are due to follow by the end of 2026. Maia 300 is meant to become the primary inference system across Microsoft’s products from 2027.
For you, the likely payoff is cheaper and faster AI features over time, though Microsoft hasn’t promised lower prices. It also hasn’t said what share of Copilot requests will run on Maia, or whether customers can pick Maia-backed capacity.
This also fits Microsoft’s push to decide where AI runs. Our guide to Windows Hybrid Intelligence covers the local-versus-cloud side, and our look at Microsoft Copilot Autopilot shows the kind of agent work that needs this much inference.
Why Microsoft keeps building its own chips
The simple answer is cost and power. Running AI for millions of people takes huge amounts of electricity, so Microsoft is measuring its chips in tokens per watt and tokens per megawatt, not only raw speed.
At Maia 200’s launch, Microsoft claimed up to 1.7x more tokens per watt than leading GPU infrastructure for built-in AI in Teams, Copilot, and Cowork. It also claimed 30% better performance per dollar than the newest hardware in its fleet. Those are Microsoft’s own figures too.
Maia 200 is rolling out across the US, Europe, and Australia, starting near Des Moines, Iowa, with Phoenix, Arizona next. Its clusters scale to 6,144 accelerators over standard Ethernet, and Maia 300 uses a custom AI transport on Ethernet-based networking. Microsoft says what it learned from Maia 200 shaped Maia 300.
Frequently Asked Questions
When will Maia 300 be available?
Microsoft says production deployments begin in 2027. It hasn’t given a firmer date or said when customers could use it.
Can I buy a Maia 300 chip?
No. Maia chips are for Microsoft’s own data centers, and you can’t rent Maia capacity by the hour today.
Is Maia 300 better than Nvidia GPUs?
Microsoft claims 40% more tokens per megawatt than leading GPU-based systems, but it didn’t name the hardware. There are no independent tests, so it’s too early to say.
What is Maia 200?
It’s the current generation, introduced in January 2026. It runs Microsoft Copilot and models from OpenAI, Microsoft’s MAI team, and open-source families.
The Maia 300 won’t change your day this week, and Microsoft’s numbers are unproven. Still, a chip that already runs Copilot, with a successor aimed at the whole company, shows Microsoft wants less dependence on outside GPU supply. Watch Ignite 2026 for the independent detail, and don’t read much into the 7x figure until then.
Leave a Reply