Your solution is Maia 200

Microsoft has presented the Maia 200, its second self-designed AI accelerator aimed at model inference, that is, executing them once trained. The chip, manufactured in TSMC’s 3 nanometer process, seeks to improve efficiency and reduce data center operating costs and the rest of the company’s AI-dependent services. Below these lines we tell you all the details.

What makes this chip special. According to the company, the Maia 200 integrates more than 140,000 million transistors and is optimized to work with large language models. Microsoft promises 30% more performance per dollar than the previous generation, the Maia 100. The company also claims that it outperforms FP4 Trainium3 from Amazon and TPU Google’s seventh generation in FP8 precision.

Image: Microsoft

Why inference matters. Inference is the process of running an already trained model to generate answers, and it is becoming an increasingly important expense for AI companies. Unlike model training, which requires raw computing power over concentrated periods, inference is a process that must operate continuously and efficiently so as not to compromise the experience of millions of users.

Highlighted technical features. The chip incorporates 216 GB of HBM3e memory with a bandwidth of 7 TB/s and 272 MB of integrated SRAM. According to the company, the chip can achieve more than 10 petaflops in 4-bit precision (FP4) and approximately 5 petaflops in 8-bit precision (FP8), all with a consumption of 750W. Just like has shared Microsoft has also designed a hierarchical memory system that promises to distribute workloads more intelligently between SRAM and HBM to keep models fed with data at all times.

Where and what it will be used for. Microsoft has already begun deploying the Maia 200 in its Azure US Central data center near Des Moines, Iowa, with the US West 3 region in Phoenix as the next destination. The chip will be used to run models like GPT-5.2 of OpenAI in services such as Microsoft 365 Copilot and Microsoft Foundry. Microsoft’s Superintelligence team will also use it to generate synthetic data and reinforcement learning tasks.

Less dependency. With the Maia 200, Microsoft joins a growing trend among large technology companies: designing its own accelerators to reduce dependence on NVIDIAwhose chips dominate the market and have a high cost. Google has its TPUs, Amazon has Trainium, and now Microsoft reinforces its hardware with this second chip after the Maia 100 launched in 2023. According to the specifications, the Maia 200 works at almost half the energy consumption of the NVIDIA Blackwell B300 Ultra (750W vs. 1400W), although the two chips are designed for different use cases (inference vs. training + inference).

Between the lines. The launch of the Maia 200 is really late. According to they point from Tom’s Hardware, the chip known internally as Braga, was scheduled for 2025 and could have come out before NVIDIA’s B300. Microsoft’s messaging repeatedly emphasizes efficiency and performance per dollar, so it aligns with the company’s strategy to keep AI operating costs in check as much as possible. It also coincides with Satya Nadella’s recent statementsCEO of Microsoft, on the need for the industry to maintain “social permission” to continue expanding its data centers.

And now what. Microsoft is already working on future generations of the Maia and, according to share According to Tom’s Hardware, the next chip could be manufactured with Intel Foundry’s 18A process. Meanwhile, the deployment of the Maia 200 will allow the company to test its ability to compete with Amazon and Google on its own infrastructure, while containing the operational costs of running its AI services at scale.

Cover image | Microsoft

In Xataka | The number of new apps coming to the App Store has skyrocketed. We have a culprit: “vibe coding”