Two rack stories, one OCP week. Microsoft is selling inference density inside Azure. AMD is selling an open blueprint that treats agent execution as its own compute tier.
Maia 300 for 2027; Helion for November
On 12 October 2026 Microsoft's Scott Guthrie post says Azure Maia 200 is already serving production workloads including Microsoft Copilot, with GitHub Copilot and Microsoft Foundry to follow by end of year on that generation. The same post says Microsoft will scale the Maia family to multi-gigawatt deployments starting in 2027 and make Maia the primary inference system for Microsoft products and services. Maia 300 is described as an inference-first system for long-context reasoning, latency-sensitive conversation, and multi-step agentic tasks. Company figures on that post: expect 40% more tokens per megawatt versus leading GPU-based systems, and up to 7× Maia 200 rack-scale token throughput in a full 96-accelerator configuration. Rack numbers printed: 96 accelerators inside a 167-kilowatt envelope, 27.6 TB of HBM capacity, and 1.7 PB/s aggregate HBM bandwidth across the scale-up domain. Per-accelerator memory on the same post: eight HBM4 stacks at 18.0 TB/s peak bandwidth within a 950-watt TDP. Compute chiplets are named on TSMC N2P. Production deployments for Maia 300 begin in 2027 on Microsoft's calendar.
Microsoft's Azure architecture post the same day adds system detail: Software Defined Locally Accessed Dataflow Architecture (SDLA), NUMA across chiplets, enhanced MXFP4 and sub-3-bit storage formats, a superpod of up to 12,288 accelerators, and claims of more than 2.5× performance per watt versus Maia 200. Those ratios are Microsoft's.
Also on 12 October 2026, AMD introduced Helion, an open agentic AI rack blueprint based on an Open Compute Project spec-design. Example configurations on the newsroom post: a CPU-only rack on sixth-generation EPYC with up to 32,768 CPU cores for hyperscale agent execution tied to separate inference; and an all-in-one rack mixing EPYC nodes, Instinct MI350P GPU nodes, and general-purpose compute for enterprise agentic AI. AMD says it plans to contribute the Helion specification to OCP. Partner configurations are under development and planned available in November 2026. Named ecosystem quotes on the post include Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, and Wiwynn. AMD's product page frames Helion versus Helios: Helios for frontier-scale training and large-scale inference; Helion enterprise for agentic infrastructure with inference in a single rack without a unified scale-up GPU fabric between the two MI350P nodes in the example layout.
Who can buy which rack story
Microsoft ships Maia inside Azure for its own services first; the Guthrie post is a capacity and silicon roadmap for Microsoft products, not a public SKU price list for third-party buyers. AMD ships a reference blueprint and partner builds; customers buy through the named OEM/ODM path once November configurations land. OCP gets the Helion spec if AMD follows through on the contribution pledge. Operators running only NVIDIA MGX or other OCP racks are not bound by either announcement until they change bill of materials.
Inference-first Maia density vs CPU-first Helion agent racks
A dated Maia 300 rack density and 2027 production window from Microsoft, plus an AMD Helion OCP blueprint that separates hyperscale CPU agent execution from optional on-rack MI350P inference. It is not a joint product. It is not proof that GPU training racks are obsolete.
Tokens per megawatt stay on the vendor blogs
Date: 12 October 2026 posts; Maia 300 production 2027; Helion configurations planned November 2026. Price: not disclosed for Maia as a merchant SKU on these posts; Helion is a blueprint with partner availability, not a single AMD list price. Hardware: Maia numbers above are Microsoft's; Helion core counts and MI350P mix are AMD's examples. Independent measurement of 40% tokens/MW, 7× rack throughput, or 2.5× perf/W versus Maia 200 is not in the posts. Whether Helion ships on the November date at every named partner is unsettled beyond AMD's plan language.
Update the 2027 inference BOM and the November agent-rack RFP
Azure-heavy buyers should treat Maia 300 as Microsoft's stated primary inference path from 2027 and ask which workloads stay on GPUs versus Maia when Foundry and Copilot capacity plans refresh. On-prem and colo teams writing agent infrastructure RFPs should ask OEMs whether Helion CPU-only or MI350P mix configs are on the November build list and how the token router is licensed. Ignore "40% more tokens per megawatt" as a contract term until you see the comparison harness.