Microsoft has revealed a new version of its Maia artificial intelligence accelerator, the news site Axios reported. The move positions the software company as the latest hyperscaler building custom silicon in direct competition with its largest supplier, Nvidia.

The report frames the disclosure as the closing of a circle. Every major Nvidia customer now operates an in-house AI chip program, and each program targets a different part of the same compute stack that Nvidia dominates. The competitive line, the report argues, has effectively dissolved.

What is the Maia chip?

Microsoft first introduced the Maia line in 2023 as part of a strategy to bring AI training and inference workloads in-house for its Azure cloud platform. The chip is intended to accelerate large language model operations, including those that underpin Microsoft's Copilot products. Microsoft ranks among the largest single buyers of Nvidia's H100 and Blackwell generation GPUs. The internal chip does not replace that purchasing commitment in the near term.

Which Nvidia customers are building their own chips?

The list of Nvidia buyers running parallel silicon programs now includes each of the four largest U.S. cloud operators.

• Google has shipped multiple generations of its Tensor Processing Unit, or TPU, since 2015.
• Amazon Web Services has deployed Trainium for training and Inferentia for inference workloads.
• Meta has developed the Meta Training and Inference Accelerator, or MTIA, for recommendation and ranking systems.
• Microsoft, with Maia, has now joined the group, according to the Axios report.

Each program targets a different layer of the AI compute stack, and each operates at a different scale. None of these companies has announced plans to end Nvidia purchases.

What does the competitive overlap change?

For Nvidia, the overlap changes the structure of customer relationships rather than the immediate revenue picture. The four cloud operators together account for a substantial share of data center GPU shipments. Custom chips typically handle a defined fraction of total AI workloads, with the remainder still running on Nvidia systems.

The competitive risk sits in second-generation deployments. Once a custom design is proven in production, customers can scale it for specific model architectures that would otherwise have demanded a new Nvidia product cycle. The economic case for Microsoft rests on volume. A chip used across millions of Azure inference calls can amortize design and fabrication costs that would otherwise flow to Nvidia's gross margin.

How does the supply picture fit?

The development lands against a backdrop of constrained advanced packaging capacity, particularly for chip-on-wafer-on-substrate, or CoWoS, the technology that links Nvidia GPUs to high-bandwidth memory. Taiwan Semiconductor Manufacturing Company operates the dominant share of CoWoS lines, with capacity expansion plans under way. Microsoft's internal chip does not displace that bottleneck. It does, however, add a parallel demand stream for foundry and packaging capacity at the same constrained nodes.

What to watch next

The next signal will be the deployment share of Maia inside Azure, measured against continued Nvidia GPU orders. Microsoft has not disclosed that ratio publicly. The wider question, as the Axios report frames it, is whether the line between Nvidia's largest customers and its closest competitors has effectively disappeared for good.