Microsoft is preparing to deploy millions of its self-developed AI chips, according to a report carried by Moomoo, a scale that moves custom silicon from experiment to infrastructure.

The chips in question belong to the company's Maia family, the accelerator line Microsoft has built in-house for its data centers. The report frames the deployment as part of a wider shift, one that runs from Google's TPU line through to Microsoft's Maia parts and is driven above all by AI inference workloads.

What is driving the ASIC expansion?

The Moomoo report identifies inference, not training, as the force behind the surge in custom chip demand. Inference is the phase where a finished model answers user queries, and it repeats at enormous volume. That arithmetic favors purpose-built silicon over general-purpose GPUs, and it explains why hyperscalers are willing to fund their own chip programs rather than rely on a single merchant supplier.

Google's TPU program established the template. Microsoft's Maia effort now appears to be reaching the same stage of maturity, with the reported deployment running to millions of units rather than pilot batches.

Why does the million-unit figure matter?

Volume is the difference between a research project and a supply chain commitment. A million-unit-class deployment requires secured foundry capacity, qualified packaging, and multi-year procurement planning. It also signals that Microsoft expects inference demand to persist long enough to justify dedicated hardware.

For the merchant GPU vendors that currently dominate AI accelerators, the calculus is straightforward: every Maia chip installed in a Microsoft rack is a socket that does not go to an outside supplier.

What comes next?

The report presents the ASIC buildout as an era-defining shift rather than a one-off order. If the deployment proceeds as described, custom accelerators will take a growing share of inference workloads across the major cloud platforms, and competition in AI compute will increasingly be fought between in-house chip families as much as between vendors.