Microsoft Maia 200 enters large-scale production, begins running its own largest models
At the Ignite 2026 conference, Microsoft announced that its second-generation in-house AI accelerator, Maia 200, has entered large-scale production and has begun running Microsoft's largest frontier AI large models in its own data centers, taking less than a year from its first unveiling in January of this year to volume deployment. The chip uses TSMC's 3nm process, delivers over 10 petaFLOPS of FP4-precision compute, and is paired with 6 HBM3e stacks providing 216GB of capacity and 7TB/s of bandwidth, with a single cluster scalable to 6,144 chips. Microsoft also confirmed that the third-generation Maia 300 will enter mass production in 2027, and that starting in 2027 it will expand Azure Maia to multi-gigawatt scale.
Why read this
SelectedAI score 77
The original article provides the specifications, deployment pace, and roadmap for Maia 200, along with industry skepticism about the tokens-per-watt metric, making it easier to assess the actual progress of cloud providers' in-house compute.
Topics
Body · AI translation
Microsoft Maia 200 Enters Mass Production, Begins Running Its Largest In-House Models Source: Semiconductor Horizons Published: 2026-11-18 15:26 Microsoft chip design generated poster From its first unveiling in January this year to volume deployment, it took less than a year.
Recently, at the Ignite 2026 conference, Microsoft announced that Maia 200, its second-generation in-house artificial intelligence accelerator, has officially entered mass production and has begun running Microsoft's largest frontier AI large models in its own data centers. From its first unveiling in January this year to volume deployment, this chip took less than a year, and the pace of cloud providers' in-house computing power is clearly accelerating.
According to the specifications published by Microsoft, Maia 200 uses TSMC's 3nm process, integrates more than 140 billion transistors on a single chip, has a die area of 26mm by 33mm, close to the lithography limit, and integrates HBM and the compute die on the same silicon interposer through CoWoS-S advanced packaging. The entire chip delivers more than 10 petaFLOPS (10145 TFLOPS) of compute at FP4 precision, 5072 TFLOPS at FP8 precision, and 1268 TFLOPS at BF16 precision, with 272MB of on-chip SRAM, and six external HBM3e stacks providing 216GB of capacity and 7TB/s of bandwidth, with an SoC thermal design power of 750 watts.
Rather than simply stacking up compute, the design focus of Maia 200 is placed more on data movement. Microsoft calls this architecture a software-defined dataflow architecture. Inside the chip, the high-bandwidth data network and the independent control network are physically separated, and dedicated DMA engines, on-chip SRAM and a custom NoC jointly serve the tensor compute units to minimize data round trips between storage tiers. Microsoft's claim is that the bottleneck of large-model inference is no longer just FLOPS; whether weights and KV cache can be fed to the compute units in time is the key to determining Token throughput.
At the system level, Maia 200 does not use a proprietary interconnect; its scale-out foundation is Ethernet. A single chip integrates 28 400Gbps network controllers, with a total bidirectional bandwidth of 1.4TB/s, and a single card provides 2.8TB/s of bidirectional scale-out bandwidth externally. The four accelerators within each tray interconnect directly in pairs, and the same in-house transport protocol is used within and across racks, allowing a single cluster to scale to 6144 chips. According to figures given in Microsoft's technical paper, such a cluster can reach 62 exaflops of FP4 compute, 43PB of aggregate memory, and 8.6PB/s of Ethernet bandwidth.
In terms of performance comparisons, Microsoft's claim is that on built-in AI features such as Teams, Copilot and Cowork, Maia 200 processes up to 1.7 times more Tokens per watt than existing leading GPU infrastructure, and delivers 30% more performance per dollar compared with the latest-generation hardware in its fleet. Internal data disclosed in the technical paper shows that, thanks to hardware-software co-design, Maia 200 can save about 30% in total cost of ownership and 15% in energy consumption compared with Microsoft's other active AI accelerators, and is not overfit to a single problem.
In terms of deployment pace, Maia 200 has already landed in data centers in the United States, Europe and Australia, handling production-grade workloads such as Microsoft Copilot, and supporting the MAI team in developing next-generation models; GitHub Copilot and Microsoft Foundry will be connected before the end of this year. The models running on these chips include OpenAI's GPT-5.2 as well as Microsoft's in-house MAI series and mainstream open-source model families. To lower the migration barrier, Microsoft has simultaneously opened a preview of the Maia SDK, which includes PyTorch integration, the Triton compiler and optimized operator libraries, as well as the NPL language for low-level programming and a cost simulator.
More noteworthy is the degree to which the roadmap has been made public. Microsoft also confirmed that the third-generation Maia 300 will enter mass production deployment in 2027, and that starting in 2027 it will expand the Azure Maia system to multi-gigawatt scale, making Maia the primary inference platform for Microsoft products and services. Maia 300 switches to a multi-chiplet architecture, with compute chiplets based on TSMC's N2P process, paired with eight HBM4 stacks providing 288GB of capacity and 18TB/s of bandwidth, and a thermal design power of 950 watts. A single rack integrates 96 accelerators, consumes 167 kilowatts, and has 27.6TB of aggregate HBM capacity and 1.7PB/s of bandwidth. Microsoft expects its Token output per megawatt to be 40% higher than leading GPU systems, and at a full configuration of 96 chips, rack-level Token throughput can reach 7 times that of Maia 200.
Looking at the competitive landscape, the reference frame for this chip is quite clear. In its launch materials, Microsoft directly benchmarks against Amazon's third-generation Trainium and Google's seventh-generation TPU, with FP4 performance about three times that of the latter and FP8 performance higher than TPU v7. The fact that three hyperscale cloud providers have pushed in-house accelerators to mass production and actual operation of frontier large models within the same time window means that the supply structure of AI compute is shifting from single-source GPU procurement to a hybrid form in which GPUs and in-house ASICs coexist.
Questions also exist. Some industry practitioners point out that the Tokens-per-watt metric lacks a unified definition, that there is no stable conversion relationship between the definition of a Token and compute, and that efficiency metrics measured in watts tend to reward models that output more verbosely. Others ask over how long a context the 40% improvement was measured, since in long-context and multi-step agent scenarios the real constraint is often memory rather than compute. Microsoft's statements in the two materials are also not entirely consistent: the engineering blog uses "expected," while an executive's personal post uses "achieved." In addition, all currently disclosed workloads are Microsoft's own businesses, and the company has not said whether customers can directly order Maia compute.
For the upstream supply chain, the significance of Maia 200 lies in bringing 3nm capacity, HBM3e supply and advanced packaging demand further into the cloud providers' in-house development cycle. As multi-gigawatt-scale deployment unfolds in 2027, cloud providers' bookings of HBM4 and N2P capacity will become earlier and more concentrated, and the bargaining structure of memory and foundry will change accordingly. Microsoft mentions in its technical paper that Maia 200 took only a few days from the first packaged samples to running real models, and that the time from first tape-out to first rack deployment was more than halved compared with similar projects; this end-to-end pace is itself a new variable that suppliers need to adapt to.
This article is reproduced from media reports or online platforms and represents the author's personal stance or views. Our reproduction is solely for sharing and does not mean that we approve of or endorse it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service in a timely manner, and as a neutral platform service provider we will promptly correct, delete or handle it in accordance with the law. Comments No user comments yet Latest News More Related Recommendations
High-end sealing component selection strategies under advanced processes for AI chips 2026-11-18 15:57
Easy to build a plant, hard to win orders: India's packaging and testing faces many challenges 2026-11-18 15:55
Hon Hai buys Sony's automotive camera module for 7.24 billion yen, placing another piece on its automotive electronics chessboard 2026-11-18 15:30
AI Trends