Microsoft is advancing a second-generation version of its homegrown Maia AI accelerator to loosen Azure's dependence on Nvidia GPUs, according to a report carried by The Edge Malaysia.

The move lines Microsoft up with Google, Amazon and Meta, all of which have deployed or commissioned custom AI silicon to cut per-token cost and lock in supply.

Why is Microsoft building its own AI chip?

Microsoft disclosed the first-generation Maia 100 in November 2023 alongside its Cobalt 100 Arm server CPU. Microsoft designed the chip end-to-end with its own rack and cooling systems.

The company manufactured Maia 100 on TSMC's 5 nm-class (N5) process and paired it with HBM memory in a board tailored to Azure's existing GPU racks.

Microsoft deployed the chip first on internal workloads, then on Azure-hosted inference for OpenAI and other large-model customers.

The economics drive the strategy. Nvidia's H100 and H200 accelerators have carried list prices well above $25,000 per unit, with allocations vendor-managed through 2024.

What changes in the second generation?

The Edge Malaysia report describes Microsoft's next accelerator as a more direct competitor to Nvidia across both training and inference.

The first-generation Maia 100 tilted heavily toward inference serving. The successor would expand into training-class workloads.

The report omits the next chip's process node, foundry partner, HBM tier and die size. It names no tape-out, sampling or volume production timeline.

What does Microsoft need from the foundry?

A multi-rack Maia ramp at the scale Microsoft has signaled would require sustained access to TSMC's leading-edge capacity and to CoWoS advanced packaging.

CoWoS lines have remained the binding constraint for the AI accelerator market through 2024 and into 2025. Hyperscaler-class chips compete with Apple, AMD and Nvidia for the same slots.

Microsoft's widely reported AI data-center buildout plans have raised the bar for any in-house silicon that can absorb a portion of that spend.

What shifts if Maia reaches volume?

If Microsoft's next-generation Maia accelerator scales to a meaningful share of Azure's AI workloads, the cloud would offer customers a non-Nvidia path for training and inference on Microsoft-managed infrastructure.

That would deepen a trend already underway: AWS, Google Cloud and Azure each carry their own accelerator brand alongside Nvidia's GPUs, with per-token economics that diverge by cloud.

For Nvidia, the threat is not new in shape — Google, Amazon and Meta all built custom AI silicon first. But Microsoft is the largest commercial test yet of whether a hyperscaler can shift meaningful workloads off the GPU incumbent.

The Edge Malaysia piece leaves the next Maia's specifications, tape-out and Azure deployment volumes open. The commercial picture will sharpen as Microsoft confirms its foundry partner, process node and customer rollout for the second-generation chip.