AMD Helion: an open rack design for agentic AI with up to 32,768 Zen 6 cores

On October 12, 2026, the opening day of the OCP Global Summit in San Jose, AMD introduced Helion, an open rack reference design for agentic AI. It is aimed at AI datacenters that need lots of ordinary CPU capacity alongside GPUs. In its CPU-only form, a Helion rack holds up to 32,768 Zen 6 cores, using 6th Gen EPYC processors, code-named "Venice." An enterprise version mixes CPU nodes with up to 16 Instinct MI350P GPUs in the same rack. AMD won't sell the racks itself: partners including Cisco, Dell, HPE, Lenovo and Supermicro will build them.

This story is based on AMD's press release and Helion product page, checked October 12, 2026, plus coverage from Tom's Hardware, StorageReview and DCD.

The short version

• What it is: a rack blueprint built on Open Compute Project (OCP) standards, which AMD plans to contribute to OCP.
• Two configurations: a CPU-only "Hyperscale" rack (up to 32,768 cores) and an all-in-one "Enterprise" rack (EPYC CPUs plus up to 16 MI350P GPUs).
• Why CPUs: AMD's pitch is that agents spend much of their time running tools, code and sandboxes, which is CPU work, not GPU work.
• When: AMD says configurations are "planned to be available in November 2026." Tom's Hardware reported availability "starting in the first half of 2027." Expect partner systems from late 2026 into 2027.
• Not disclosed: prices, per-rack power and benchmark results.

Why build a CPU rack for AI agents?

GPUs generate the tokens, but an agent does much more than generate text. It runs code, calls tools, browses, queries databases, and works inside isolated sandboxes (see our explainer on what "agentic" means). Each of those steps runs on general-purpose CPUs. AMD's Dan McNamara, senior vice president and general manager for Compute and Enterprise AI, put it this way in the release: "Agentic AI is not a single workload," adding that it "takes optimized compute and infrastructure that can adapt to different workloads."

Helion is AMD's answer for that CPU side. It complements Helios, AMD's GPU rack with 72 Instinct MI455X accelerators, which AMD says OpenAI expects to bring online from Q4 2026 (see our coverage of Vera Rubin and AMD). In AMD's hyperscale design, a Helion CPU rack runs the agents and hands model inference to a separate GPU rack such as Helios, according to StorageReview.

What's in a Helion rack?

Last verified: October 12, 2026 (AMD Helion product page and press release).

The 32,768 figure is 128 processors at 256 cores each. Tom's Hardware and StorageReview identify the dense part as the EPYC 9996. AMD's press release describes an "integrated token router that directs agents' model calls across frontier and open-source models," but gives no detail on how it works. The product page lists the OCP standards the design follows, including ORv3, DC-MHS, DC-SCM, OCP NIC 3.0 and the Caliptra root of trust. Tom's Hardware described the rack as based on ORv4. AMD's own page says ORv3, and so does Supermicro's statement.

The MI350P is not a new chip. AMD announced it on May 7, 2026: a passively cooled, dual-slot PCIe card with 144 GB of HBM3E, 4 TB/s of bandwidth and a 600 W maximum board power, configurable to 450 W. That makes the enterprise rack a fit for air-cooled datacenters that can't take Helios-class liquid-cooled GPU racks.

Who's building it, and when?

AMD's release names eight partners with statements: Cisco, Dell Technologies, Hon Hai Technology Group (Foxconn), HPE, Lenovo, QCT, Supermicro and Wiwynn. Dell says it supports the reference design and can extend deployments with Dell PowerRack. Supermicro says it will sell Helion through its Data Center Building Block Solutions. StorageReview also lists MiTAC, Inventec, ASUS, ASRock Rack and GIGABYTE among AMD's partners.

On timing, the sources disagree:

• AMD's press release: configurations "are currently under development with ecosystem partners" and are "planned to be available in November 2026." DCD and StorageReview repeat that date.
• Tom's Hardware, from AMD's pre-briefing: availability "starting in the first half of 2027," with Zen 6 deployments ramping later in 2026.

These may describe different milestones, such as first partner configurations versus volume shipments, but AMD hasn't said. Treat November 2026 as AMD's official target and expect broader availability in 2027. No partner has published pricing or a ship date for a specific product.

What it means

For hyperscalers, Helion is AMD's bet that agent fleets will need dedicated CPU capacity at rack scale. The argument is that each running agent may need its own sandbox, browser or code interpreter, and those jobs run on CPU cores, not GPUs. For enterprises, the all-in-one rack is a way to keep sensitive agent workloads on-premises without rebuilding a datacenter for liquid-cooled GPUs.

Sandboxes also matter for safety. Isolating what an agent can touch is one of the main defenses discussed in our case against unattended agents. Helion's sandbox nodes are hardware for that isolation; the isolation itself still depends on the software running on them.

AMD's claims about agentic performance are untested so far. AMD published no benchmarks for Helion, and the Zen 6 parts it depends on are only now ramping. For the bigger picture of where this capacity is going, see our tracker of gigawatt-scale datacenter projects.

What's confirmed, reported and uncertain

• Confirmed (AMD): the announcement date; the two configurations; up to 32,768 Zen 6 cores; up to 16 MI350P GPUs; 102 Tbps networking; 21-inch ORv3 rack with liquid or air cooling; plans to contribute the spec to OCP; the eight partners; the November 2026 target.
• Reported: EPYC 9996 as the dense part, and that AMD won't sell racks (Tom's Hardware); availability starting 1H 2027 (Tom's Hardware); the extra partner names (StorageReview).
• Unknown: pricing, rack power, which partner ships first, and how the token router works.