AI Chips · Product page

Microsoft Maia 200: The Complete Guide

The chip that arrived six months late and still made it from first silicon to a production rack faster than Microsoft’s own past programs. Simple to expert, every number links to its source.

Stuck at any point? Ask an AI about this page →
Ask an AI about this chip:CAGPX
Microsoft Maia 200 10 sections · 2 levels 10 linked sources Checked September 2026
How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.

1.At a glance

SimpleStart here

Maia 200, announced 26 January 2026, is Microsoft’s current AI chip, built specifically for inference — running already-trained models rather than training new ones 1 2. By June 2026 it was in production in Microsoft data centres near Des Moines, Iowa, and Phoenix, Arizona, running OpenAI’s GPT-5.2 models and Microsoft 365 Copilot 3 1.

Deep diveThe technical detail

Microsoft calls Maia 200 “the most performant, first-party silicon from any hyperscaler” and says it delivers “30% better performance per dollar than the latest generation hardware in our fleet today” 1. Like every Maia chip, it is not for sale or rent — Microsoft uses it only for its own services.

2.Launch and history

SimpleStart here

Maia 200 was late. Press reporting from June 2025, more than six months before launch, said the chip — then code-named Braga — had slipped at least six months into 2026, citing design changes including features OpenAI requested and staff turnover 4.

Deep diveThe technical detail

Microsoft says that despite the delay, Maia 200 went from first silicon to its first data-centre rack in “less than half that of comparable AI infrastructure programs” 1. By its fourth-quarter fiscal-2026 earnings call (July 2026), CEO Satya Nadella said the chip “continues to scale” and now supports both OpenAI’s and Microsoft’s own MAI models 5.

3.What’s inside it

SimpleStart here

Tuned for very small number formats (4-bit and 8-bit) that let it do more calculations per second on inference workloads, with 272 MB of fast on-chip memory 1.

Deep diveThe technical detail

Maia 200 links chips over standard Ethernet rather than a proprietary fabric: 2.8 TB/s of scale-up bandwidth per chip, four chips per tray joined by direct links, and a two-tier network supporting clusters of up to 6,144 chips 1. At Hot Chips 2026, Microsoft said the chip has no separate scale-out network at all: “it is all scale-up networking” 6.

4.Full spec table

SimpleStart here
SpecMicrosoft Maia 200
AI computeOver 10 PFLOPS FP4; over 5 PFLOPS FP8 1
Memory216 GB HBM3e 1
Memory bandwidth7 TB/s 1
On-chip memory272 MB 1
TransistorsMore than 140 billion 1
Power750 W (SoC TDP) 1
Process nodeTSMC 3 nm 1
Die sizeAbout 820 mm², per press coverage of Microsoft’s Hot Chips 2026 talk 6
Deep diveThe technical detail

Against Maia 100 (see that page on this site): memory grew from 64 GB to 216 GB, memory bandwidth from 1.8 to 7 TB/s, and the process node moved from TSMC 5 nm to 3 nm 7 1. Press reporting names SK hynix as Maia 200’s sole HBM3e supplier; Microsoft itself does not name a memory supplier 8.

5.Where it’s made

SimpleStart here

Designed by Microsoft; manufactured by TSMC on its 3 nm process, its most advanced node yet for a Maia chip 1.

Deep diveThe technical detail

Microsoft does not name which TSMC factory makes Maia 200. At Hot Chips 2026, Microsoft reported measured figures for the chip including 1.65 petaflops of effective attention throughput and about 1.3 TB/s for a BF16 all-reduce operation 6.

6.Which systems use it

SimpleStart here

Running in Microsoft’s own Azure regions: US Central (near Des Moines, Iowa) and US West 3 (near Phoenix, Arizona), both confirmed in production by June 2026, with Italy, Australia and South Korea named as next 3 1.

Deep diveThe technical detail

Maia 200 uses a second-generation closed-loop liquid-cooling unit, an upgrade from the “sidekick” cabinet cooling used on Maia 100 (see that page on this site) 1. Microsoft has also opened a Maia 200 software development kit in preview, inviting developers, AI start-ups and academics outside the company to use it — no equivalent outside-developer preview was published for Maia 100 1.

7.Official pricing

SimpleStart here

Not for sale or rent. There is no Maia 200 virtual machine on Azure and no published chip price; Microsoft uses it only for its own services 1.

Deep diveThe technical detail

Press reported in May 2026 that Anthropic was in early talks to use Maia, which would make it the first outside company to run workloads on Microsoft’s own AI silicon; as of the check date neither company has confirmed a deal 9.

8.Real-world performance

SimpleStart here

Microsoft’s own comparisons: “three times the FP4 performance of the third generation Amazon Trainium”, and FP8 performance “above Google’s seventh generation TPU” 1.

Deep diveThe technical detail

These are Microsoft’s own comparisons against competitors’ own published figures, not an independent third-party benchmark. See the Amazon Trainium and Google TPU guides on this site for those chips’ own figures. Microsoft also says it is “seeing 40% better performance per watt when running MAI models on Maia 200” — Microsoft did not state what baseline that 40% is measured against, so this page does not assume it means “versus Maia 100” 5.

9.What came before, what came next

SimpleStart here
Came before: Maia 100, covered on its own page on this site. Came after: nothing confirmed — press reports point to a possible Maia 300, but Microsoft had not announced it as of this page’s last check.
Deep diveThe technical detail

Maia 200 arrived more than two years after Maia 100’s November 2023 announcement. Press reporting from August 2026, citing The Information, said Microsoft was seeking TSMC capacity for more than 300,000 Maia 300 chips for 2027 delivery and could unveil the chip “as early as September” 2026 10. As of this page’s last check, no Microsoft announcement, blog post or filing confirming Maia 300 had been found, so it is treated here as unconfirmed, not shipped.

10.Hidden in plain sight

SimpleStart here

Maia 200 has no scale-out network at all — everything runs over the same scale-up Ethernet fabric, a design choice Microsoft called out explicitly at Hot Chips 2026 6.

Deep diveThe technical detail

Microsoft’s own performance-per-dollar claim (30% better than its existing fleet) is framed against Microsoft’s own hardware mix, not against a single named competing chip — a narrower and harder-to-verify comparison than the specific FP4-versus-Trainium3 and FP8-versus-TPU-v7 figures Microsoft publishes elsewhere in the same announcement 1.

11.Sources

10 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.

  1. Official Microsoft Blog: Maia 200: the AI accelerator built for inference (26 Jan 2026)Official
  2. Microsoft Tech Community: Deep dive into the Maia 200 architecture (26 Jan 2026)Official
  3. Microsoft: Microsoft Build 2026 live blog (2–3 Jun 2026)Official
  4. Tom’s Hardware (press, citing The Information): Microsoft’s own AI chip delayed six months (Jun 2025)Press
  5. Microsoft Investor Relations: FY2026 fourth-quarter earnings callOfficial
  6. ServeTheHome (press): Microsoft’s Maia 200 accelerator at Hot Chips 2026 (25 Aug 2026)Press
  7. Microsoft at Hot Chips 2024: Inside Maia 100 (Xu and Ramakrishnan)Conference
  8. TrendForce (press): Maia 200 on TSMC 3nm; SK hynix reportedly sole HBM3E supplier (27 Jan 2026)Press
  9. DatacenterDynamics (press, citing The Information): Anthropic in talks to use Microsoft’s Maia chips (May 2026)Press
  10. TrendForce (press, citing The Information): Microsoft eyes Maia 300 launch in September (11 Aug 2026)Press

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.