The biggest AI hardware news of October 2026 is a change in what the chips are for. Microsoft and Nvidia, with AMD close behind, spent the month selling hardware built for inference, meaning the work of running a trained model for real users, and for AI agents that call those models in long loops. Every one of those launches leans on a component that is already sold out: high-bandwidth memory.
Below is the news that matters if you build on these systems, with dates and numbers taken from the vendors' own announcements. Vendor numbers are vendor numbers, so where a comparison does more marketing than measuring, this post says so.
Custom AI chips: Microsoft's Maia 300 bets on inference
Microsoft announced Maia 300 on 12 October, the third generation of its in-house AI accelerator. It starts production deployments in 2027, and Microsoft says Maia will become the primary inference system behind its own products and services.
The spec sheet reads like a memory datasheet wearing a compute badge. Each chip carries eight HBM4 stacks with 18.0 TB/s of peak bandwidth inside a 950-watt power budget, and its compute chiplets are built on TSMC's N2P process. A full rack holds 96 accelerators in 167 kilowatts, with 27.6 TB of HBM between them. Microsoft also leans on low-precision formats such as MXFP4, a 4-bit number format that shrinks the model weights and the KV cache, the stored attention state that grows with every token of context.
The memory focus has a mechanical reason. When a model generates text, every new token means reading the weights and the KV cache back out of memory. That decode step waits on bandwidth far more than on arithmetic, so a chip designed for long conversations and agent loops spends its silicon budget on feeding data to the cores.
The headline claim in Microsoft's Maia 300 announcement is 40% more tokens per megawatt than leading GPU-based systems, plus up to 7x the rack-scale throughput of Maia 200. Microsoft does not name the GPU system it measured against. A benchmark against an unnamed rival is the hardware version of a partner who lives in another country and is very real, honestly.
The program has moved past slideware, though. Maia 200 already serves Microsoft Copilot in production, with GitHub Copilot and Microsoft Foundry due to follow by the end of the year. Microsoft says Maia 200 already delivers up to 1.7x more tokens per watt than leading GPU infrastructure for the AI features in Teams and Copilot, which is the number that justified building a successor.
Microsoft is far from alone here. Every hyperscaler now designs its own accelerator, because a chip tuned to your own models and traffic beats a general-purpose part on cost per token, even when it loses on peak benchmarks. Expect Nvidia to keep selling to all of them anyway.
Nvidia Vera Rubin reaches Google Cloud
Vera Rubin is Nvidia's successor to Blackwell, and October is when it turned into something you can rent. At GTC Berlin on 21 October, Google Cloud and Nvidia announced A5X instances, bare-metal machines built on the Vera Rubin NVL72 rack, which links 72 Rubin GPUs into a single system.
The first customer is Ineffable Intelligence, the London lab founded by David Silver, which is using A5X for large-scale reinforcement learning. Google Cloud plans to scale its Vera Rubin fleet to as many as 960,000 Rubin GPUs across multiple sites, with expansion into Europe planned.
Read the small print on dates. The headline of Nvidia's Google Cloud announcement says general availability. The body says general availability is coming next month. Generally available, specifically in November.
If you are planning capacity, treat November as the earliest date and assume the first racks go to customers with large commitments, which is how new GPU generations usually roll out. For background on who actually manufactures these systems, our piece on who builds Nvidia's AI engines walks through the supply chain from foundry to rack.
One detail in the same announcement matters for regulated teams. Google Cloud will offer Nvidia Confidential Computing at rack scale on bare-metal A4X Max instances, the GB300 NVL72 generation, which keeps model weights and prompts encrypted while they are in use. If your legal team has blocked cloud inference over data exposure, that is the feature to ask your account manager about.
AMD Helion and the CPU side of agentic AI hardware
AMD's October news argues that agents need more than GPUs. Helion, announced on 12 October, is an open rack blueprint for agentic AI based on an Open Compute Project design, and AMD plans to contribute the specification back to OCP.
The reasoning holds up. An agent spends much of its time outside the model: calling tools, running code and waiting on APIs. That work runs on CPUs. Helion's hyperscale configuration is a CPU-only rack of 6th Gen EPYC processors with up to 32,768 cores, meant to host agent execution while separate infrastructure serves the models. Its enterprise configuration puts EPYC CPU nodes and Instinct MI350P GPU nodes in a single rack.
AMD's Helion release also mentions an integrated token router that sends agents' model calls across frontier and open-source models, though it says little about how the router decides. A router of this kind sends easy requests to a small, cheap model and hard ones to a frontier model, which is the cheapest optimization most agent stacks are still missing. Partner configurations are planned for November 2026.
This was the most honest framing of the month. Anyone who has profiled an agent loop has watched an expensive GPU sit idle while a Python process waits for a search API to answer. Thirty-two thousand cores is a lot of patience to rent.
The HBM memory shortage underneath every launch
HBM, or high-bandwidth memory, is DRAM stacked in layers and mounted right beside the processor, so data travels a short and very wide path. Every accelerator above depends on it, and there is not enough to go around.
Micron told investors at its 21 October technology webinar that its entire 2026 HBM supply, HBM4 included, was sold out before the event. It also set out a new HBM generation every year through 2030, running from HBM4 this year to HBM6 at the end of the decade, with HBM4E samples targeted for early 2027.
HBM is hungry for wafers. Micron puts the trade at roughly three wafers of conventional DRAM displaced for every HBM wafer, so tight HBM supply drags ordinary server memory along with it. Do the arithmetic on Maia 300: eight stacks per chip across 96 chips is 768 HBM stacks in one rack. At that density, a memory allocation is a deployment schedule.
Supply also has a policy side. Taiwan's economics minister said in mid-October that advanced semiconductors and AI servers will join the island's export control list for China, which would make unauthorized shipments a criminal matter under its Foreign Trade Act. No effective date has been set.
The same bandwidth logic applies on a desk. Our explainer on whether LLMs use Apple unified memory shows why memory bandwidth sets token speed on a laptop as firmly as it does in a data center.
What this AI hardware news means if you build on AI
Most readers will never buy a rack. You will rent capacity from someone who did, and October's news tells you how that rental market behaves through 2027.
- Cheaper tokens arrive slowly. Every vendor this month sold efficiency in tokens per watt or per megawatt. Those gains reach API prices after the hardware ships at scale, which for Maia 300 means 2027.
- Memory sets your bill. Long prompts and big KV caches consume bandwidth on any chip. Trimming context and caching repeated prompt prefixes saves money whoever made the silicon.
- Agents need CPU budget. If your agent spends most of its wall-clock time in tools, a faster GPU will leave its latency where it was.
- Keep a fallback model ready. With HBM sold out, providers will ration their newest capacity, so test a second model before you need it.
Hardware roadmaps share one trait with software estimates: the date on the slide is the earliest date the universe permits. Plan around shipped capacity.
If you want the engineering underneath these choices, LLM Systems Engineering covers how inference loops work and how to plan compute budgets for training and serving. For the cost side from a manager's chair, AI Systems in Business covers total cost of ownership. And if a rented GPU starts to look expensive, our guide to running a local LLM shows what your own hardware can carry.
The rule of thumb for reading next month's AI hardware news: watch the memory announcements more closely than the chip launches. The chips get the keynotes. The memory decides when they ship.