Maia 200
AI inference accelerator on TSMC 3nm process
Microsoft designed the Maia 200 as a specialised chip for running artificial intelligence models. Built on a 3-nanometer process by TSMC, it focuses on efficient data handling and low-precision computing to lower costs for cloud services.
- Built on TSMC 3nm process with 140 billion transistors
- Delivers over 10 petaFLOPS in 4-bit precision
- Features 216GB HBM3e memory at 7 TB/s bandwidth
- Operates within a 750W power envelope
- Deployed in US Central and US West datacentres
What it is
The Maia 200 is a system on a chip built specifically for artificial intelligence inference. Microsoft created this hardware to improve the cost efficiency of generating text and data from large language models. Unlike general purpose processors, it uses native support for low-precision formats like FP8 and FP4. This design allows it to handle massive models while keeping power consumption within a 750 watt limit. The chip serves as a key component in Microsoft Azure cloud infrastructure.
How it came about
Microsoft developed the Maia series to reduce reliance on third party suppliers for AI compute. The earlier Maia 100, known internally as Athena, supported both training and inference but saw limited deployment. The company shifted focus with the Maia 200, codenamed Braga, to target inference workloads exclusively. This specialisation simplified the design and aimed to outperform competitors in token generation economics. The project used pre-silicon validation to speed up its path to production.
Inside the chip
The architecture prioritises data movement to prevent bottlenecks during model execution. It features a redesigned memory subsystem with 216 gigabytes of HBM3e memory running at 7 terabytes per second. On-chip static random access memory totals 272 megabytes to keep data close to compute units. Specialised direct memory access engines and a network on chip fabric manage high bandwidth traffic. The chip delivers over 10 petaFLOPS in 4-bit precision and over 5 petaFLOPS in 8-bit precision.
How it was made
Taiwan Semiconductor Manufacturing Company produces the Maia 200 using its 3-nanometer process node. Each die contains more than 140 billion transistors packed into the system on a chip. Microsoft employed a cloud-native development approach, validating system elements before the silicon arrived. This method allowed models to run on the hardware within days of receiving the first packaged parts. The rapid deployment reduced the time to rack installation by half compared to previous programmes.
Where it went
This accelerator powers Microsoft Foundry and Microsoft 365 Copilot services. It supports various models, including recent versions from OpenAI such as GPT-5.2. The Microsoft Superintelligence team also utilises the chip for synthetic data generation and reinforcement learning tasks. These workloads help create high-quality, domain-specific data for training next-generation internal models. The hardware is deployed in US Central and US West 3 datacentre regions.
Versions and family
The Maia 200 succeeds the Maia 100, which was designed for both training and inference. While the predecessor had 105 billion transistors and used a 5-nanometer process, the new chip moves to 3 nanometers. The Maia 200 drops training support to focus entirely on inference efficiency. Microsoft describes the accelerator programme as multi-generational, with future designs already in development to continue setting performance benchmarks.
Why it matters
Microsoft claims the Maia 200 offers three times the FP4 performance of the third generation Amazon Trainium. It also reportedly exceeds the FP8 performance of Google seventh generation tensor processing units. The company states it achieves 30 per cent better performance per dollar than its latest existing fleet hardware. This efficiency aims to lower the total cost of ownership for dense inference clusters across the global Azure network.
Questions people ask about Maia 200
What is the primary purpose of the Maia 200?
The Maia 200 is an AI accelerator designed specifically for inference workloads. It helps improve the economics of generating tokens for large language models by offering high performance per dollar.
Which manufacturing process is used for the Maia 200?
The chip is fabricated by TSMC using their 3-nanometer process technology. This advanced node allows the integration of over 140 billion transistors on a single die.
How much memory does the Maia 200 contain?
It includes 216 gigabytes of HBM3e memory with a bandwidth of 7 terabytes per second. Additionally, it has 272 megabytes of on-chip SRAM for fast data access.
What models does the Maia 200 support?
The accelerator serves multiple models, including the latest GPT-5.2 models from OpenAI. It is integrated into Microsoft Foundry and Microsoft 365 Copilot services.
Good reading and sources
There is no English Wikipedia article on Maia 200 yet, so the facts on this page come from the sources listed above: blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-, nextplatform.com/ai/2026/01/28/microsoft-takes-on-other-clouds-with-br. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.