Microsoft drops its first custom AI chips

For years, big tech's artificial intelligence addiction had one obvious gatekeeper: Nvidia's GPU supply.

At its Ignite developer conference, Microsoft finally ended years of rumors by unveiling its first in-house silicon for the cloud. The announcement features two distinct chips: the Azure Maia 100 AI accelerator and the Azure Cobalt 100 CPU.

Packing 105 billion transistors for OpenAI workloads

The star of the show is Maia 100, custom-built specifically for large language models and generative AI tasks. Fabricated on a 5-nanometer process, it packs a massive 105 billion transistors, making it one of the largest chips ever built on that node.

Maia is not just an experimental test run. Microsoft confirmed it will power workloads behind GitHub Copilot and run models from OpenAI, where Microsoft has invested $11 billion. Both Maia 100 and Cobalt 100 are paired with 200 gigabits-per-second networking, capable of pushing 12.5 gigabytes per second in data throughput.

Alongside the AI accelerator sits the Azure Cobalt 100, an in-house 64-bit CPU featuring 128 cores built on the ARM architecture. Microsoft claims it slashes power consumption by 40% compared to previous ARM-based instances on Azure, and it is already running live workloads for Microsoft Teams and Azure SQL.

The last cloud giant to build its own silicon

Microsoft is actually the last among the Big Three hyperscalers to design its own silicon. Google got into the custom accelerator game back in 2016 with its TPU, while Amazon Web Services has spent years rolling out Graviton, Trainium, and Inferentia.

Even with custom silicon in hand, Microsoft made sure to clarify that it is not breaking up with Nvidia or AMD. The company announced plans to deploy Nvidia's flagship H200 and AMD's MI300 next year.

Demand is simply too high to rely solely on internal chips. GitHub Copilot paid corporate customers jumped 40% quarter-over-quarter, reaching over 1 million paid users across 37,000 organizations with massive traction outside the United States.

Why it matters

We think Microsoft's move into custom silicon is less about replacing Nvidia overnight and more about gaining desperately needed leverage. As AI inference costs threaten to eat into cloud margins, designing bespoke silicon gives Microsoft a critical path to better unit economics.

For enterprise customers and developers, you will not suddenly see an option to buy raw Maia chips. The real payoff will come downstream in the form of more stable cloud pricing, lower latency for tools like Copilot, and less severe capacity constraints when OpenAI deploys next-generation models.

It looks like a pragmatic hedge. By powering high-volume background services like Teams with Cobalt and offloading select AI inference to Maia, Microsoft can reserve its priciest Nvidia clusters for frontier training tasks. The playbook is smart, but keeping software developers happy on custom silicon remains an ongoing battle.