AI news videos

Microsoft's Maia 300 Is Announced, Not Available

Microsoft announced Maia 300, its newest custom AI accelerator, on October 12, 2026. Microsoft also says the chip will launch in 2027, so this is an announcement rather than a part anyone can run today. Everything below comes from Microsoft's own announcement and its own engineering blog.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• Establishes WHAT was announced, WHO announced it, and WHEN: Microsoft, on its own newsroom host, on Oct 12 2026, introducing a custom chip called Maia 300. Backs the hook's subject ("Microsoft announced its newest custom AI chip today") and the proper noun in scene 1 ("Microsoft calls it Maia 300"). Backs scene 5's narrated clause that the announcement page contains no specifications, which is a direct observation of this page in both the live fetch and the same day snapshot. Supplies NO numeric claim to this Short, deliberately, because it publishes none.(THE ALERTED FIRST PARTY URL, AND WHY IT IS NOT THE FACT SOURCE. Microsoft newsroom page titled "Introducing Microsoft's latest custom chip: Maia 300", dated Oct 12 2026, detected by our watcher at 2026-10-12T13:18:36.158Z, four minutes after publication. This page was fetched LIVE at production time (HTTP 200) and is also held as our watcher's snapshot at scripts/data/article-cache/5a6d73fa70f63e19.txt. Both copies agree on what it is: a PRESS KIT or media resources page, not an article. Its 534 words are almost entirely site navigation chrome. It contains no specifications, no performance numbers, no availability date and no quoted sentences. What it does contain is the title twice, a "Watch the video" player plus "More videos", downloadable image captions ("Azure Maia 300 System", "Azure Maia 300 chip render", "Azure Maia 300 chip to server to rack render", three "Azure Maia 300 server" items, three "Azure Maia 300 prototype chip" items and "Azure Maia 300 board"), five video titles naming Microsoft executives (Scott Guthrie, Saurabh Dighe, Rani Borkar, Mustafa Suleyman and Jim Kleewein), a "More resources" list ("Maia 300 industry analyst quotes", "Explore Azure infrastructure from silicon to systems", "The yield imperative: Turning infrastructure into useful intelligence", "Video transcripts: Download the file") and two links to real Microsoft articles about this chip. Because the page itself carries no facts, the Short does not script from it; it scripts from the two first party Microsoft articles it links, which are the other two entries in this list. Scene 5 states this on camera in the narrated words "the announcement page itself contains no specifications", and the scene 5 card spells out the full chain. NO prompt injection or instruction like content was found anywhere in the live page or the snapshot; both were treated throughout as untrusted data informing facts only.)
• Backs scene 2's narrated claim in full ("more rack scale token throughput and better performance per watt than Microsoft's previous Maia chip"), which is a word for word plain English restatement of this article's own sentence, with the attribution left visible by the phrase "the headline claim is". Backs scene 3's narrated claim that the central comparison is Microsoft measuring its new chip against the one it already built: every multiplier in this article names Maia 200 as the baseline, and Maia 200 is Microsoft's own earlier accelerator. Supplies every digit on the scene 2 and scene 3 cards. Backs the cross check recorded on the scene 5 card, since 288 GB x 96 accelerators equals the 27.6 TB per rack figure in the other article and 18 TB/s x 96 equals about its 1.7 PB/s aggregate. Narration speaks none of these digits; they are on silent cards only.(PRIMARY FACT SOURCE 1, the Azure architecture blog. First party Microsoft engineering article on the Maia 300 architecture, published Oct 12 2026, the same day as the alert, authored by Rani Borkar (President, Azure Hardware Systems and Infrastructure) and Saurabh Dighe (Corporate Vice President, Systems Planning and Architecture). Reached from the alerted press kit's own link and fetched live at production time. This is the article the press kit points to and it is where every specification in this Short comes from: 96 accelerators per rack (doubled from Maia 200), multi-chiplet architecture on the TSMC N2P process, 288 GB HBM and 18 TB/s per accelerator, eight HBM4 stacks, a NUMA architecture across chiplets, a 36% increase in compute FLOPs against Maia 200, 2.5x greater memory bandwidth, 33% more capacity, 2x HBM bandwidth per watt, a superpod of up to 12,288 accelerators, up to 750 GB/s peer-to-peer bandwidth inside a superpod, a 167 kW rack power envelope, 24 four-accelerator trays, six high-radix switch ASICs in three liquid-cooled trays, a 50V busbar, liquid cooling removing more than 80% of the thermal load, the Software-Defined Locally Accessed Dataflow Architecture, enhanced MXFP4 and sub-3-bit storage formats, and optimized vector engine kernels. It also carries the two headline multipliers verbatim: "Maia 300 delivers up to 7 times greater rack-scale token throughput and more than 2.5x performance per watt vs. Maia 200." It is first party and therefore interested: Microsoft is measuring silicon Microsoft designed, and the Short says so on camera in scene 3.)
• Backs the hook's second sentence ("It does not arrive until 2027"), which is this article's own stated launch year for Maia 300 and nothing more; the hook does not characterise the delay, it reports the year. Backs scene 1's narrated claim that the chip will run in Microsoft's own data centers, since this article describes Maia as Azure infrastructure and names the Microsoft products its predecessor already serves rather than any external buyer. Backs scene 4's narrated claim in full: Microsoft does claim more tokens per megawatt than leading GPU based systems, and this article's own phrase "leading GPU-based systems" is the entire identification of the comparison, so "it names no competing system" is a direct observation of the text. Backs scene 5's narrated claim that no price and no independent benchmark were published, which holds across this article, the architecture blog and the press kit; none of the three states a price or cites an outside test.(PRIMARY FACT SOURCE 2, Scott Guthrie's expansion announcement, "Microsoft to Expand Azure Maia System to Multi-Gigawatt Deployments", published Oct 12 2026 by Scott Guthrie, Executive Vice President, Cloud + AI. Reached from the alerted press kit's link aka.ms/AA13e23q, which returns a 301 to this blogs.microsoft.com address, Microsoft's own host, so this is first party and not an aggregator or a link out stub. This is the article that carries the TIMING, which is the load bearing honesty fact of this Short: Microsoft will scale the Azure Maia family to multi-gigawatt deployments beginning in 2027, the Maia 200 is currently in production supporting Microsoft Copilot, GitHub Copilot and Microsoft Foundry, and the Maia 300 will launch in 2027. It also carries the only outward facing efficiency claim Microsoft makes, "40% more tokens per megawatt" compared to leading GPU-based systems, with no competing vendor, product or configuration named and no methodology given; the internal figure "1.7x more tokens per watt" for built-in AI capabilities in Teams, Copilot and Cowork experiences; the restated specifications of eight HBM4 stacks at 18.0 TB/s peak memory bandwidth, a 950-watt TDP, 96 accelerators per rack in a 167-kilowatt envelope, 27.6 TB HBM capacity and 1.7 PB/s aggregate HBM bandwidth per rack, and the TSMC N2P process node; and the Guthrie quote "model and silicon choice are inherent in our vision for Frontier Transformation".)