AMD has put a CPU-first rack at the center of its agentic AI pitch. At the OCP Global Summit in San Jose on October 12, 2026, AMD unveiled Helion, an open reference design that can hold up to 128 EPYC 9996 "Venice" processors, or 32,768 Zen 6 cores, in one rack. Tom's Hardware reported the details, including a revised ship date of November 2026.
In the same keynote, CEO Lisa Su reportedly previewed a next-generation 6th Gen EPYC chip called Verano. This post separates the hardware that is documented from the claims that are still thin, and explains why a company that sells GPUs is pushing racks full of CPUs.
AMD Helion at a glance
| Question | Answer |
|---|---|
| What is it? | An open, standardized rack reference design for agentic AI, contributed to OCP |
| Main configuration | CPU-only: up to 128 EPYC 9996 (256-core Venice) CPUs, 32,768 cores |
| Hybrid option | Up to 16 Instinct MI350P GPUs and 102 Tbps networking, with fewer CPUs |
| Standards | OCP ORv4 rack, 44 OpenU trays, DC-MHS compliant, air or direct liquid cooling |
| Does AMD sell it? | No; OEMs and ODMs bring it to market |
| Timing | Updated to November 2026 (originally first half of 2027) |
| Verano | Next EPYC, teased; details unconfirmed |
What Helion is, and how it differs from Helios
Last year at OCP, AMD showed Helios, an open rack for GPU training and inference. Helion is its counterpart. Tom's Hardware describes it as a standardized 21-inch rack reference design that complements Helios, built on OCP's ORv4 specification, which is slightly wider and taller than a standard IT rack. It supports 44 OpenU trays, each 21 inches wide and 48mm tall, and supports both air-cooled and direct liquid-cooled deployments. It is also compliant with OCP's DC-MHS standard for modular, interoperable server hardware.
The key point is what goes inside. The CPU-only version uses the dense 256-core EPYC 9996, and 128 of them yield 32,768 cores. The hybrid version trades CPUs for up to 16 MI350P accelerators and 102 Tbps of networking, a configuration AMD targets at enterprises that want one rack for data management, networking, and AI. For the GPU side of AMD's roadmap, our earlier coverage of the AMD and Anthropic 2 GW MI450 deal explains why Helios and Helion are complementary rather than competing.
What changed on the ship date
Tom's Hardware's article carries an update note: AMD originally said Helion would ship in the first half of 2027 but revised the release date to November 2026, and the outlet corrected an error that had listed the tray width as 21mm instead of 21 inches. The body text of the article still says first half of 2027 in one place, so treat the November date as the later correction. Because AMD contributes designs and does not sell racks, the date likely refers to the reference design and early partner systems, not to a shelf product.
Partners
Tom's Hardware says AMD did not share supporting or deployment partners in its pre-briefing, though it suggested Helios customers will also deploy Helion racks. The outlet said CEO Lisa Su would announce partners in the keynote. We could not confirm a partner list from a readable primary source, so we do not name any; check AMD's own materials.
Why agents need CPUs
The reasoning behind Helion is practical. Tom's Hardware notes that the personal-agent model in use today, as in Meta's Muse and OpenAI Dots, gives each user a sandbox with a slice of compute on a host machine, backed by an inference server. Some work happens on CPU-only servers, such as building software, while heavy inference runs elsewhere. See our analysis of what Muse and Dots promise on privacy for how those sandboxes are described.
Small shapes moving inside a low green-walled square, representing an AI agent sandbox running on Helion CPU servers
An agent's loop is mostly not matrix math. It reads files, runs tests, calls APIs, parses output, waits on the network, and keeps state. Multiply that by thousands of simultaneous users and you need many cores with decent memory and fast networking, while the accelerator pool handles only the model calls. That is the gap Helion targets: a dense, open CPU rack that sits beside the GPU racks.
Nvidia showed its own open rack designs for Vera Rubin at the same event; see our coverage. Both vendors are using OCP to make the surrounding rack, power and cooling an open standard.
Why the Open Compute Project matters here
The Open Compute Project is a community working on open data center and hyperscale infrastructure standards, first initiated by Facebook, now Meta, in 2011. Its 2026 Global Summit runs October 12 to 15 in San Jose, and the project's own materials point to rack power densities approaching 1 MW. That context explains why vendors bring rack designs to OCP: a shared specification lets operators buy trays, power shelves and cooling from several suppliers instead of one. For Helion, that means the value is as much in the standardized envelope as in the 256-core chip inside it.
What is known about Verano
Verano is the thinly sourced part of this story. A relayed report of the keynote said Su teased Verano as a next-generation 6th Gen EPYC data center processor for 2027, with an LPDDR memory interface. We could not confirm that from a readable primary source, and Tom's Hardware lists a separate report that Verano, described as an AI host CPU, will reportedly use a special SB1 socket. Specs, core counts, pricing, and exact timing have not been published in anything we found, so treat Verano as a name and a direction, not a product.
Memory bandwidth is the constraint that matters for agent hosts, which is why any CPU aimed at them gets scrutinized on memory. For the broader memory squeeze, see how memory costs are hitting local AI hardware.
What this means if you buy or build infrastructure
The practical question for a platform team is where agent sandboxes should run. If your agents mostly execute code, run test suites and call tools, the bottleneck is core count, memory per sandbox and network, not accelerator FLOPs. A dense CPU rack lets you size that tier independently of your inference tier, and upgrade each on its own schedule. The trade-off is operational: a 21-inch OCP rack needs facility support for its power and cooling, so the decision is as much about the data hall as the silicon.
- Plan CPU capacity for agents, not just GPUs. If you run agent fleets, sandbox density and per-user CPU slices will decide cost as much as token prices.
- Open racks reduce lock-in. Helion and Helios sit on OCP standards, so partners can build to one spec. Check which OEMs ship what, and when.
- Do not plan around Verano yet. Treat it as direction, not a purchase spec.
- Watch the November date. If AMD's revised timing holds, partner systems could arrive alongside Venice ramps; if not, the original first-half 2027 window is the fallback.
Questions people are asking
Is Helion a product I can order from AMD? No. Tom's Hardware stresses AMD will not sell racks. It contributes the reference design to OCP, and OEMs and ODMs would build and sell systems based on it.
Does a CPU-first rack mean AMD is backing away from GPUs? The reporting points the other way. Helios remains the GPU rack for training and inference, and the Helion hybrid option includes Instinct accelerators. Tom's Hardware suggests Helios customers will also deploy Helion, so the two are meant to sit side by side in the same data hall.
What is the 256-core EPYC 9996? It is the dense Venice part AMD uses to reach 32,768 cores in a rack. The arithmetic is simple: 128 processors times 256 cores equals 32,768. Clock speeds, power draw and pricing for that part were not in the sources we reviewed.
How does it compare with an all-GPU design? Per the same article, a CPU-only Helion would look like an agent sandbox tier connected to a separate inference tier, because model serving still runs on accelerators. The hybrid version tries to put some of both in one rack, at the cost of CPU capacity.
Why does 21-inch matter? The OCP-style 21-inch rack allows denser, shared power and cooling than a standard 19-inch rack. It is part of why operators building new AI halls favor it, though it also means the design cannot simply drop into an existing 19-inch room.
Claimed versus verified
Documented by Tom's Hardware from AMD's press briefing: Helion's ORv4 basis, 44 OpenU trays, DC-MHS compliance, air and liquid cooling, 128 EPYC 9996 CPUs for 32,768 cores, hybrid option with up to 16 MI350P GPUs and 102 Tbps, AMD not selling racks, and the date revision to November 2026.
Relayed from the keynote, not confirmed: Verano's LPDDR interface and 2027 date.
Unknown: Verano specs and pricing, which GPU powers the all-in-one version, and which customers deploy Helion.
What to watch next
AMD's investor and newsroom pages should carry the keynote details. Next, look for OEM announcements with real shipment dates, any hyperscaler adoption statements, and how this compares to Nvidia's CPU and rack strategy. For the supply side of this race, see our note on optical parts sold out through 2029.
Related reading
- AMD and Anthropic: 2 GW MI450 deal
- Nvidia Vera Rubin open designs at OCP
- Meta Muse and OpenAI Dots privacy promises
- AI chip supply chain: TSMC, Samsung, GlobalFoundries
- Optical parts sold out through 2029
- Nvidia DGX Spark price and memory cut
Source: Tom's Hardware on AMD Helion.
Details are as of October 13, 2026 and come from reports of the OCP keynote; AMD may revise specifications and dates.