Update, 12 October 2026 (REPORTED): Microsoft wants to be able to order high-bandwidth memory years ahead, according to an interview with the head of its Maia chip programme. Andrew Wall, general manager of Microsoft’s Maia programme, spoke to Seoul Economic Daily (SEDaily) in a video interview on 7 October; SEDaily published it on 12 October at 14:55 BST. This is SEDaily’s report, not a Microsoft announcement, and the quotes below are SEDaily’s English rendering.
High-bandwidth memory, or HBM, is a type of fast memory made by stacking memory chips on top of each other and placing them right next to an AI chip, so data can move in and out quickly. Every Maia accelerator needs it, and only a few companies make it.
“It is very important for us to establish a high level of trust such that we can place orders as much as five years in advance,” Wall said, according to SEDaily. He also said: “Microsoft’s supply chain team is working very closely with all memory companies to ensure sufficient HBM supply not only for Maia 300 but also for next-generation roadmaps beyond it.”
SEDaily reports that Wall described the HBM industry as being in a very constrained situation. “HBM is clearly a major issue and constraint across the industry right now,” he said. Microsoft does not disclose which companies supply its HBM; Wall said only that it is “working with all the major players”, SEDaily reports.
CONFIRMED: Microsoft announced Maia 300, its next in-house AI chip system, on its official blog on 12 October 2026. All performance comparisons below are Microsoft’s own claims or expectations and have not been independently tested (COMPANY CLAIM).
Microsoft says it will make its own Azure Maia chips the main system it uses to run AI for its products, at very large scale. In a blog post on 12 October 2026, Scott Guthrie, Executive Vice President of Cloud + AI, wrote [1]:
“We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.”
Microsoft did not give a total gigawatt figure, a number of sites or a cost.
What is inference? Training is how an AI model is built. Inference is the work of running a finished model to answer requests, such as a Copilot reply. Every answer costs computing power and electricity.
Maia 200 is already in production
Microsoft says Maia 200 “is now serving Microsoft production workloads, including Microsoft Copilot”, with GitHub Copilot and Microsoft Foundry “to follow by end of year” [1]. It says Maia 200 runs OpenAI models, Microsoft’s own MAI models and open-source models, and is deploying across the United States, Europe and Australia [1].
COMPANY CLAIM: Microsoft says Maia 200 delivers “up to 1.7x more tokens per watt” for built-in AI features in Microsoft Teams, Copilot and Cowork, “compared with leading GPU infrastructure” [1]. It does not say which GPUs.
What are tokens per watt? AI models read and write text in small pieces called tokens. Tokens per watt measures how much AI output a system produces for each unit of power. Higher means more work for the same electricity.
Maia 300: what Microsoft announced
Microsoft calls Maia 300 the third generation of its Maia program, with an “inference-first architecture” built for long conversations, fast replies and multi-step “agentic” tasks [1].
COMPANY CLAIM: “we expect Maia 300 to deliver 40% more tokens per megawatt compared to leading GPU-based systems,” Microsoft says. In a full 96-accelerator setup, it says Maia 300 “delivers up to 7x the rack-scale token throughput of Maia 200” [1].
Specs Microsoft gave [1]:
- Eight HBM4 memory stacks with 18.0 TB/s of peak memory bandwidth, within a 950-watt TDP (the chip’s rated power and heat budget).
- A multi-chiplet design, with compute chiplets built on TSMC’s N2P process. Chiplets are smaller pieces of silicon combined into one chip.
- Support for low-precision number formats, including “enhanced MXFP4”, which Microsoft says cuts memory and bandwidth needs.
- Ethernet-based networking linking chips across racks.
- Racks of 96 accelerators within a 167-kilowatt power envelope, with 27.6 TB of HBM and 1.7 PB/s of total HBM bandwidth.
What is HBM? High-bandwidth memory is fast memory stacked in layers right next to the chip. AI inference often waits on memory, so how quickly data moves in and out matters as much as raw compute. Microsoft says “moving data efficiently, rather than raw compute power, is increasingly the bottleneck in large inference systems” [1].
Microsoft also describes Maia 300 as delivering “industry-leading HBM bandwidth per watt” and “a significant performance-per-watt advantage over traditional GPU architectures” (COMPANY CLAIM) [1].
When
“Maia 300 will begin production deployments in 2027,” Microsoft says [1]. It says it will share more at Microsoft Ignite 2026 in San Francisco next month [1].
In its results for the quarter to 30 September, Microsoft reported $39.4 billion of additions to property and equipment, double the figure a year earlier.
Sources
- Scott Guthrie, “Microsoft to expand Azure Maia system to multi-gigawatt deployments”, The Official Microsoft Blog, 12 October 2026. https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- The 13 Billion Bet: Why Microsoft Is Building Its Own AI Stack
- Microsoft’s 900MW Power Grab: The AI Infrastructure War Just Escalated
- Meta’s Modular AI Chip Gambit: Why Everyone Is Building Their Own Silicon to Escape Nvidia
- Counterpoint: Most Nvidia Rubin HBM4 Likely 12-Layer, Not 8
- What Is Custom HBM? Why Nvidia Wants to Move the Memory Controller Off the Chip
- Microsoft Revenue Rises 19% to $92.3bn; Azure Grows 48% to $33.2bn
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X





