Microsoft Maia 300

Maia 300 is Microsoft’s next-generation AI inference platform, announced October 12, 2026. Microsoft also said it will expand Azure Maia to multi-gigawatt deployments starting in 2027.Problem: Running AI at scale is limited by power and cost, and Microsoft has to serve a lot of models to a lot of products like Copilot and Teams. Microsoft’s answer is to design its own silicon alongside the GPUs it buys.Solution: Maia 300 is built inference-first. Per Microsoft, it targets long-context reasoning, low-latency conversation and multi-step agent work. Microsoft says it will become the primary AI inference system for its own products and services. Satya Nadella is quoted calling it “our biggest bet yet on our own silicon.”What Microsoft reports about the design:Eight HBM4 memory stacks with 18.0 TB/s peak bandwidth, and a 950 W power ratingCompute chiplets on TSMC’s N2P process, with a multi-chiplet design that separates compute, memory control and I/OSupport for low-precision formats such as MXFP4 to cut memory and bandwidth needsA custom AI transport for scale-up communication across accelerators and racksOne rack holds 96 accelerators in a 167 kW power envelope, with 27.6 TB of HBMPerformance: these are Microsoft’s own claims, with no independent benchmarks. It says Maia 300 delivers 40% more tokens per megawatt. Coverage disagrees on what that is compared against, with some outlets saying the previous Maia and others saying leading GPU-based systems, so I’d check Microsoft’s post. It also expects a Maia 300 rack to reach up to 7x the token throughput of Maia 200 at rack scale.Timeline:Maia 200 is already running production workloads in the US, Europe and AustraliaMaia 300 begins production deployments in 2027Multi-gigawatt scale-up of Azure Maia also starts in 2027 P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends