Microsoft plans to scale its custom Azure Maia artificial intelligence infrastructure to multi-gigawatt deployments starting in 2027, positioning the silicon family as the primary inference engine across its product ecosystem.
Scaling Custom Silicon for Next-Generation Workloads
As the industry pushes into the era of Super Intelligence, hyperscale cloud providers face severe power and memory constraints. Microsoft is addressing this bottleneck by anchoring its infrastructure strategy in vertical co-design, pairing custom hardware directly with software applications. The Azure Maia program forms the core of this transition, with Maia 200 currently running production workloads for Microsoft Copilot and supporting the MAI team in training subsequent model generations.
According to pulse2.com, Maia 200 is already operational across data centers in the United States, Europe, and Australia. The chip delivers up to 1.7 times more tokens per watt for native AI capabilities in Microsoft Teams, Copilot, and Cowork environments compared to traditional GPU hardware. Microsoft plans to extend Maia 200 deployment to GitHub Copilot and Microsoft Foundry before the end of 2026. The architecture processes a diverse mix of workloads, ranging from OpenAI models to Microsoft’s proprietary MAI models and prominent open-source model families.
Architecture and Performance of Maia 300
While Maia 200 handles current production demands, the upcoming Maia 300 introduces an inference-first architecture tailored specifically for long-context reasoning, latency-sensitive conversational experiences, and sustained multi-step agentic workflows. By optimizing efficiency across the entire system, Microsoft expects Maia 300 to yield 40 percent more tokens per megawatt than leading GPU-based systems.
Today we are announcing our biggest bet yet on our own silicon.Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services.And we’re introducing Maia 300, which delivers 40%… pic.twitter.com/L490XvOys9
At the rack level, a full 96-accelerator configuration of Maia 300 operates within a 167-kilowatt power envelope. This density delivers 27.6 terabytes of High Bandwidth Memory capacity and 1.7 petabytes per second of aggregate HBM bandwidth across a high-speed scale-up domain. pulse2.com noted that this configuration provides up to seven times the rack-scale token throughput of Maia 200.
Key Hardware Specifications of Maia 300
- Memory Bandwidth: Eight HBM4 stacks deliver 18.0 terabytes per second of peak memory bandwidth within a 950-watt thermal design power envelope.
- Chiplet Design: A multi-chiplet layout separates compute, memory-controller, and I/O functions, with compute chiplets manufactured on TSMC’s N2P process node.
- Precision Formats: Support for advanced low-precision formats, including enhanced MXFP4, reduces memory and bandwidth pressure for KV-cache storage and computation.
- Dataflow Optimization: A software-defined dataflow architecture uses on-die SRAM as a software-managed scratchpad to minimize data movement between compute and memory resources.
Expanding Enterprise Inference Capabilities
The operational data gathered from running large-scale services on Maia 200 directly shaped the specifications of Maia 300. Together, the two generations will distribute workloads across multiple high-demand enterprise environments. GitHub will leverage the infrastructure for multi-step agentic coding tasks, while Microsoft Foundry utilizes the cloud-scale architecture to serve diverse model architectures.
Healthcare workflows will also see integration through Dragon Copilot, which relies on low-latency inference to reduce administrative burdens for clinicians. By combining custom silicon, specialized networking protocols, and software-defined dataflow, Microsoft aims to sustain multi-gigawatt power requirements without sacrificing token throughput as enterprise AI adoption accelerates toward 2027.
