Microsoft introduced the Maia 300, its second-generation AI accelerator, alongside a multi-gigawatt Azure capacity commitment. The chip succeeds the Maia 100, which the company unveiled at Ignite in November 2023.

What changes with the second-generation silicon?

The Maia line marks Microsoft's vertical push into custom AI hardware. Microsoft and OpenAI engineered the first-generation Maia 100, which TSMC fabricated on a 5-nanometer process.

Each Maia 100 server pairs two accelerators over a high-bandwidth interconnect. Microsoft deployed the chips in early 2024, primarily to back Azure OpenAI Service inference.

The Maia 300 extends that rack-scale architecture into a new generation. Detailed specs — transistor counts, memory bandwidth, interconnect speeds — remained undisclosed at the time of the announcement.

How large is a multi-gigawatt Azure commitment?

One gigawatt roughly equals the continuous output of a large nuclear reactor. At full load, it could power tens of thousands of H100-class sockets running GPT-4-class training jobs.

Microsoft capex has surged through fiscal 2024 and 2025 to fund new shells, GPUs, networking fabric and liquid-cooling systems. Azure now spans more than 60 regions and roughly 300 data centers.

Active construction extends across Iowa, Wisconsin, Arizona and the United Kingdom. Microsoft's AI infrastructure capex has climbed into tens of billions per quarter through fiscal 2025.

The Maia 300 build-out slots into that pipeline. Multi-gigawatt-class sites typically bundle training, fine-tuning and inference capacity behind a single grid interconnect.

Why pursue in-house AI silicon now?

Microsoft has invested more than $13 billion in OpenAI across multiple tranches, alongside long-term compute commitments. Owning part of the silicon stack reduces cost per token and cushions margins from NVIDIA price swings.

At peak GPU shortage in 2023, H100 lead times stretched beyond 12 months and secondary-market prices ran well above MSRP. Custom silicon lets Microsoft tune hardware to internal model workloads and inference traffic shapes.

Google's TPU program and AWS Trainium and Inferentia chips have followed the same playbook for years. Microsoft is now scaling toward similar depth after a measured start.

What does this signal for AI infrastructure buyers?

Hyperscaler customers should expect a two-tier accelerator market inside Azure. NVIDIA H100 and H200 parts will keep handling peak training workloads. Maia-class parts will absorb steady-state inference at scale.

AMD's MI300X remains the third supplier line, with Microsoft adding capacity through 2025. The three-way sourcing strategy gives Microsoft negotiating leverage against NVIDIA, which still controls roughly 90 percent of the AI accelerator market by unit volume.

The commercial question is pricing. If Microsoft bills Maia 300 capacity below leading NVIDIA rates, custom silicon becomes a procurement line item customers actively pull. If it holds near parity, workloads only land on the custom stack when Azure routes them there.

What to watch next

The first live Maia 300 deployments will likely surface in Azure regions tied to the new multi-gigawatt sites. Customers should track Azure pricing sheets for the first custom-silicon SKUs and benchmark them against H200-equivalent rental rates.

Microsoft also needs to demonstrate Maia 300 software maturity. CUDA-equivalent toolchains, model-porting paths and reliability track records will determine whether the second-generation chip attracts third-party model traffic, or stays largely an internal inference engine.

Watch for Microsoft financial disclosure on gross-margin movement at the Intelligent Cloud segment. Custom silicon savings should show up there within two to three quarters of mass deployment.