What Microsoft's October 12 Maia post schedules

What happened

On October 12, 2026, Scott Guthrie posted on the official Microsoft blog that Azure Maia 200 is now in production and that Azure Maia 300 production deployments begin in 2027. The post says Microsoft will scale the Maia family to multi-gigawatt deployments starting in 2027, and that Maia is to be the primary inference system for Microsoft products and services. It describes Maia 300 as an inference-first design for long-context reasoning, latency-sensitive conversation, and sustained multi-step tasks. A 96-accelerator configuration, the post says, delivers up to seven times the rack-scale token throughput of Maia 200, and the company expects 40 percent more tokens per megawatt than leading GPU-based systems. The page filing that post is Futureweb.

What the post schedules

The schedule has two generations in one paragraph. Maia 200 is described as already in production. Maia 300 is described as beginning production deployments in 2027. The post does not offer a SKU, a price, or a region where a customer can reserve a rack. It says the family is to become the primary inference system for Microsoft's own products and services, and it points readers to a technical paper and a Maia 300 site for architectural detail. This page does not summarize that paper. The chip, as a hardware object rather than a model release, is why the brief sits on Hardware.

What the outlets add

The Verge, the same day, quotes the 40 percent tokens-per-megawatt expectation and the seven-times rack figure, and it says production rollout starts in 2027. CNBC reports that Microsoft intends to rely mainly on its own silicon for incoming requests, that a rack is to hold 96 Maia 300 chips, double the Maia 200 count, and that a general manager declined a timeline beyond the 2027 production start. CNBC also recalls Maia 100 in 2023 and Maia 200 in January. Those earlier dates are the outlet's account, not sentences this page is lifting from the October 12 post. A nameplate figure, which this megawatt claim is not, is filed as what nameplate capacity measures.

Why it matters on this desk

The dependency is a dated company post about who will run inference, and on which silicon, starting in 2027. It is not a benchmark a second lab can rerun, and it is not a purchase order. Readers of the hardware desk can hold three company sentences at once: production deployments in 2027, a 96-accelerator rack compared with Maia 200, and an expectation of more tokens per megawatt than GPU systems. The expectation is labeled as an expectation. What a process node measures, a different hardware fact, is what a process node measures.

What this page will not import

An August report said Microsoft might unveil Maia 300 in September and was seeking foundry capacity for a large 2027 delivery. The October 12 post does not adopt that calendar or that unit count, so this page does not either. CNBC's report that Nadella, in July, described Maia 200 as 30 percent better performance per dollar is a different sentence from a different month. It stays with CNBC. The October 10 post about pausing a model is also a different document, filed as what Satya Nadella's October 10 post asks to be able to stop.

Where the accounts overlap

The blog, The Verge, and CNBC agree that the announcement date is October 12, 2026, that Maia 300 production starts in 2027, that the comparison rack has 96 accelerators, and that the company states a seven-times throughput figure against Maia 200 and a 40 percent tokens-per-megawatt expectation against GPU systems. They do not agree on a customer price, because none of them prints one. They do not agree on a city or a data-center campus, because the post does not name one. The overlap is the schedule and the two comparison figures.

What to watch next

A production notice in 2027, a price, or a region would replace the schedule sentence with a deployment sentence. A correction that withdraws the seven-times figure or the tokens-per-megawatt expectation would change the comparison. Another outlet repeating the blog would not. The October 12 post is the record this page has. An earlier preorder that dated different NVIDIA machines is a separate hardware document, what the RTX Spark preorder posts date.

What would change the record

If Microsoft edits the post, the quotes have to be read again. If the technical paper contradicts the rack count, the paper would become the more specific document and this page would say so. Until then the blog is the company's public schedule, The Verge is the check that a technology desk quoted the same two figures, and CNBC is the check that a business desk reported the same 2027 production start and the same 96-chip rack. No street price is in any of the three.

What stays a 2027 schedule

What stays specific is an October 12, 2026 post by Scott Guthrie: Maia 200 in production, Maia 300 production deployments in 2027, multi-gigawatt deployments of the family starting that year, a 96-accelerator configuration up to seven times Maia 200's rack-scale token throughput, and an expectation of 40 percent more tokens per megawatt than leading GPU-based systems. It is not a chip in a store. It is not the October 10 pause post. The Verge and CNBC report the same schedule and the same two comparison figures. They do not add a price, and this page does not invent one. A reader who wants the architectural paper can open the blog, which points to it. This page files the schedule the blog itself states.