Microsoft has announced the Azure Maia 300, the newest member of its custom AI accelerator family, according to UPI. The company confirmed the chip's existence and name but, at the time of the report, had not published full specifications, pricing, or deployment volumes.

The announcement matters because it extends a silicon program Microsoft first placed at the center of its infrastructure strategy in November 2023, when it introduced the original Maia 100 and the Cobalt CPU. A second, then a third generation following that cadence signals that the Maia line is now a standing product family rather than a one-off experiment.

What do we actually know about the Maia 300?

The confirmed fact is narrow but firm: Microsoft has announced a new AI chip called the Azure Maia 300. Everything beyond the name — process node, transistor count, memory bandwidth, power envelope, packaging, and which foundry builds it — remains outside what the company has confirmed in this report.

That restraint is itself informative. When Microsoft launched the Maia 100, it paired the announcement with benchmarks, rack designs, and a clear story about workloads. A name-first announcement suggests a roadmap disclosure rather than a shipping product. Analysts typically treat such disclosures as directional signals, not capacity or revenue commitments, and this report gives no basis to treat the Maia 300 differently.

In plain words: the chip is announced, its silicon is not yet described.

Why does a third-generation Maia matter commercially?

The mechanism is cost and supply. Hyperscale operators buy accelerators in quantities that no single merchant supplier can serve at acceptable price and delivery terms. A custom design lets Microsoft tailor the chip to its own models, its own cooling, and its own software stack, and it removes the margin a merchant chip vendor would otherwise capture on every unit.

The competitive picture has shifted accordingly. Amazon builds its Trainium and Inferentia lines, Google fields its TPU family, and Microsoft now carries Maia through at least a third generation across its custom-silicon effort. Each new generation narrows the functional gap between hyperscaler in-house parts and merchant GPUs for the specific workloads those operators run most.

None of this replaces merchant GPUs outright. Frontier-model training still leans heavily on NVIDIA's highest-end parts, and Microsoft remains one of NVIDIA's largest customers. Custom silicon first captures inference and fine-tuning workloads, where the economics favor purpose-built parts. The takeaway is that Maia 300 deepens a split infrastructure: merchant GPUs for the frontier, custom silicon for volume.

What should buyers and suppliers watch next?

The open questions are concrete. Which foundry manufactures the part, at what node, and in what volumes, will determine whether the Maia 300 is a 2026-scale deployment or a later-decade one. So will the software story: Maia's value depends on how much of Azure's AI serving traffic Microsoft can actually move onto it without re-engineering customer workloads.

For semiconductor suppliers, the watch item is the bill of materials. Custom accelerators still buy high-bandwidth memory, advanced packaging, and networking from the same strained supply chains that serve merchant GPUs, so a volume Maia 300 ramps demand rather than replacing it.

Microsoft has not yet dated volume availability or disclosed performance targets for the Maia 300, so the commercial impact will become measurable only when the company publishes deployment figures for its datacenters.