
Quick Summary
On Monday, October 12, 2026, Microsoft unveiled Maia 300, the third generation of its custom AI chip family. The chip is built for inference, which means running AI models after they have been trained. Microsoft says Maia will become the primary inference system behind its products and services, with large-scale deployment starting in 2027. Most performance figures below are Microsoft’s own projections and have not been independently tested.
What Is Maia 300?
Maia is Microsoft’s in-house line of AI accelerators for its Azure cloud. The first chip appeared in November 2023, and Maia 200 followed in January 2026. Maia 300 is the successor.
Inference is the everyday work of generating answers, code and summaries for users. It runs constantly, so it drives much of the ongoing cost of AI. Microsoft says the Maia system is designed around the workloads it already runs, including long-context reasoning, fast conversational replies and multi-step agent tasks. Earlier Maia chips already support Microsoft 365 Copilot and run OpenAI models on Azure alongside Microsoft’s own models.
Key Features
- Efficiency: Microsoft expects 40% more tokens per megawatt of power. A token is a small unit of text that a model produces. Media reports describe the comparison as against leading GPU-based systems, but it is a company estimate, not an independent benchmark.
- Throughput: At rack scale, Microsoft expects up to seven times the token output of Maia 200.
- Rack design: Microsoft reportedly fits 96 chips in one rack with a power envelope of about 167 kilowatts.
- Chiplet architecture: Instead of one large main chip surrounded by memory, as in Nvidia’s approach, Maia 300 combines several smaller specialized chips in a single package. Reports say the core compute uses TSMC’s 2-nanometer process.
- Scale: Microsoft plans multi-gigawatt deployments beginning in 2027. Reuters reported that this could mean more than one million chips.
Why It Matters
Running AI is expensive, and the bill grows with every Copilot request. A chip tuned for Microsoft’s own workloads could lower that cost and free up scarce third-party hardware for paying Azure customers.
The move also fits an industry pattern. Google and Amazon already build their own AI chips, and OpenAI has developed one with Broadcom. Microsoft started later, and its second-generation chip arrived after delays, so Maia 300 is widely read as a larger bet on its own silicon.
Still, this is not a break with Nvidia. Microsoft continues to buy chips from Nvidia and AMD, so Maia adds to its mix rather than replacing it. Analysts at Evercore ISI estimate Maia could make up about 10% of Microsoft’s AI chip spending by early 2027. Because large-scale rollout starts in 2027, financial benefits are unlikely to appear sooner.
Limits and Open Questions
- Unverified numbers: The 40% and sevenfold figures come from Microsoft and are described as expected results. No outside tests have confirmed them.
- Missing specs: In the sources reviewed, Microsoft has not published per-chip details for Maia 300, such as memory capacity or peak compute. Treat unofficial figures with caution.
- Manufacturing risk: Reports have flagged tight advanced chip-packaging capacity, which could limit how fast Microsoft can scale.
- Availability: Microsoft describes Maia 300 as the engine for its own products. The sources reviewed do not announce a date for outside customers to rent it directly.
Maia 300 vs Maia 200
| Feature | Maia 200 | Maia 300 |
|---|---|---|
| Announced | January 2026 | October 12, 2026 |
| Main role | AI inference | AI inference (primary system for Microsoft products from 2027) |
| Manufacturing process | TSMC 3nm | Reported 2nm for core compute |
| Design | Single large chip, 140+ billion transistors | Chiplet-based package (multiple smaller chips) |
| Peak compute (per chip) | 10+ petaFLOPS in FP4; about 5 in FP8 | Not disclosed |
| Memory | 216 GB HBM3e, 272 MB on-chip SRAM | Not disclosed |
| Power | 750 W per chip | Not disclosed per chip; about 167 kW per 96-chip rack (reported) |
| Efficiency claim | 30% better performance per dollar than prior fleet hardware | 40% more tokens per megawatt (expected) |
| Rack-level output | Baseline | Up to 7x Maia 200 (expected) |
| Rollout | Live in Iowa and Arizona data centers | Large-scale deployment from 2027 |
Figures are manufacturer claims or press reports, not independent measurements.
Frequently Ask Question
What is Microsoft Maia 300?
Maia 300 is the third generation of Microsoft’s custom AI chip, unveiled on October 12, 2026. It is designed for inference, the work of running trained AI models for services such as Copilot.
When will Maia 300 be available?
Microsoft plans large-scale deployment starting in 2027. It has said more details will come at its Ignite conference in November 2026.
How is Maia 300 different from Maia 200? Microsoft expects up to seven times the rack-level token output of Maia 200 and 40% more tokens per megawatt. Reports also describe a chiplet design and a 2-nanometer process for core compute, while Maia 200 uses a 3-nanometer process.
Does Maia 300 replace Nvidia chips?
No. Microsoft still buys chips from Nvidia and AMD, and Maia adds to its mix.
Can I rent or buy Maia 300?
The sources reviewed do not announce outside availability. Microsoft describes it as the main inference system for its own products and services.
What is AI inference?
Inference is running an already-trained AI model to produce answers, code or summaries for users, as opposed to training the model.
Plan: What to Watch Next
- November 2026: Microsoft says it will share more details at its Ignite conference in San Francisco. Watch for per-chip specs and customer availability.
- 2027: First large-scale Maia 300 deployments begin, and Maia becomes the main inference system for Microsoft products.
- Beyond 2027: Maia 200 is expected to stay in service for several more years, according to a Microsoft Azure Maia executive quoted by Techzine.
- Your next step: Follow Microsoft’s official announcements for confirmed specifications before relying on any third-party number.
Official Sources
- Microsoft Source EMEA: Microsoft announces Maia 300
- Official Microsoft Blog: Scott Guthrie’s posts, including Maia 200
- Microsoft Source Asia: Maia 200, the AI accelerator built for inference
Comments
What do you think: will custom chips like Maia change how much AI costs?
Share your view below.
Author: Atia Arefin | Published: October 14, 2026 | Last updated: October 14, 2026
Sources: Microsoft Source (official announcement), The Official Microsoft Blog, GeekWire, The Next Web, Techzine
Image: Original illustration created by Navigate Explained using Canva.
How we check and correct our articles: Editorial Policy. To report an error, contact us.
This article is for informational purposes only. Specifications, performance figures, availability and rollout plans may change.













































Leave a Reply