# AMD Helion Racks: 73,728 Threads, 3 Variants, 0 Prices

> Source: <https://shattered.io/amd-helion-racks-73728-threads-three-variants-zero-prices/>
> Published: 2026-11-16 02:48:48+00:00

AMD’s Helion rack family moved from blueprint to buyable hardware this month, and the first numbers are large. On November 10, 2026, AMD declared general availability of its 6th Gen EPYC server processors on the SP7 platform and named three Helion deployment options, with some high-density configurations reaching up to 73,728 threads per rack. [Data Center Knowledge reported](https://www.datacenterknowledge.com/data-center-chips/amd-launches-epyc-sp7-helion-as-ai-rack-demand-rises) that the first AMD Helion racks were planned for availability during November.

The catch is price. AMD has published no list price for Helion Liquid, Helion Air or Helion Air Inference, and its own materials describe Helion as an open blueprint rather than a fixed, priced rack product. Buyers get thread counts and cooling options but no cost per rack. That gap matters because the silicon inside is not cheap. AMD’s footnotes list a two-socket EPYC 9996 setup at $14,904 against $14,995 for a two-socket Xeon 6990E+, and they set the EPYC 9996 against [Nvidia’s Vera CPU in a SPEC comparison](https://shattered.io/nvidia-vera-cpu-vs-amd-epyc-9996-vs-intel-xeon-6-spec-price-2026/) that AMD built itself.

Three facts hold across AMD’s launch materials. Helion runs on EPYC 9006 server CPUs. It ships in liquid-cooled, air-cooled and GPU-assisted forms. And it targets agent workloads, which lean on CPU cores, rather than the GPU-heavy training that the separate Helios rack serves. Everything past that, including customer lists, rack power and cost, remains unpublished as of November 16, 2026.

Don't miss new tech stories on Google

Add Shattered once in the Google app and our stories show up in your news suggestions.

[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)

## What AMD Launched on November 10 for Helion and EPYC 9006

AMD packed several announcements into one week. The company declared general availability of the 6th Gen AMD EPYC server processors on November 10, built on the Zen 6 architecture and the SP7 platform. Server makers were already shipping SP7 systems that day. Supermicro said on November 11 that its H15-generation servers on EPYC 9006 CPUs were shipping, and it widened its AMD collaboration to cover the Helios and Helion platforms. AMD also launched an embedded EPYC 9006 line on SP7 aimed at networking gear, firewalls and storage controllers.

Cloud availability lags the hardware. AMD said Microsoft Azure and Oracle Cloud Infrastructure instances on the new chips were coming soon, with no dates attached. That wording matters for anyone planning capacity, because a coming-soon label can mean weeks or quarters, and AMD has not said which.

Helion is the piece that changes what a rack looks like. AMD introduced it on October 12 as an open blueprint for agentic AI infrastructure, based on an Open Compute Project specification and powered by AMD compute. At that point AMD said the configurations were under development with ecosystem partners. The November 10 materials turned the concept into three named deployment options, and AMD’s Helion page now describes the platform as powered by EPYC 9006 Series server CPUs.

Naming trips people up, so it is worth pinning down. Venice is the codename. AMD’s launch materials call the product 6th Gen AMD EPYC, and the model family is EPYC 9006. Any article that still describes EPYC 9006 as upcoming went stale on November 10. The manufacturing story goes back to May 21, when AMD announced a production ramp for Venice on TSMC’s 2 nm process. The [earlier EPYC 9006 Venice coverage](https://shattered.io/amd-epyc-9006-venice-zen-6-256-cores-pcie-gen6-2026/) has the pre-launch spec picture.

## Three Helion Variants and a Price List That Does Not Exist

AMD splits Helion into Liquid, Air and Air Inference. The names describe cooling and GPU content, and the differences decide which data centers can take which rack. The table below lists only what AMD’s materials state, with unknowns marked as such.

| Variant | Cooling | Role AMD describes | GPU content | List price | 
|---|---|---|---|---|
| Helion Liquid | Liquid | CPU-led agent execution, designed to run alongside Helios in large deployments | None (CPU-only) | Not published | 
| Helion Air | Air | Adds agent capacity inside existing air-cooled data centers | Not stated | Not published | 
| Helion Air Inference | Not stated | CPU agent execution plus local GPU inference in one rack | Local GPU inference | Not published | 
| Platform CPU | n/a | 6th Gen AMD EPYC 9006 Series server CPUs on SP7 | n/a | Not published | 

### Helion Liquid: density first

Liquid is the high-density option. AMD positions it beside Helios in large agentic deployments, where the goal is the maximum number of agents per rack. Cold plates or similar liquid loops let a rack carry far more CPU power than air allows, and that is the only way the highest thread counts in AMD’s claims make physical sense. Operators need liquid-ready halls to use it.

### Helion Air: capacity without a rebuild

Air targets the installed base. Most enterprise and colocation space still runs air-cooled racks with fixed power envelopes, and AMD pitches Helion Air as a way to add agent capacity there without a cooling retrofit. The trade-off is density. Air-cooled racks hit a power ceiling long before they reach the thread counts AMD advertises for the top configurations.

### Helion Air Inference: one rack for both jobs

Air Inference combines CPU-based agent execution with local GPU inference. The idea is that an agent plans, calls tools and parses results on CPU cores, then hands model calls to GPUs sitting a few inches away rather than across a data center network. One report describes the enterprise design as pairing EPYC CPU nodes with Instinct MI350P GPU nodes and general-purpose compute, though AMD’s own pages do not confirm the GPU model. One caution applies. AMD associates MI450 with Helios, and nothing in the materials supports describing MI450 as the GPU inside Helion generally.

On price, the picture is blank across the board. No official AMD price exists for Helion Liquid, Helion Air, Helion Air Inference or an assembled rack. AMD has not confirmed a numbered release version either. Procurement teams will deal with OEMs and ODMs, and quotes will follow configuration, memory and cooling choices.

## Helion vs Helios: Two Racks for Two Jobs

AMD sells two rack families, and confusing them leads to bad purchasing assumptions. Helios is the GPU-centered, rack-scale system for frontier training and large-scale inference. Helion is the CPU-led blueprint for the agents that run on top of those models. AMD says Helion Liquid is designed to operate side by side with Helios, which tells you the company sees the two as parts of one data center rather than rivals.

| Item | AMD Helios | AMD Helion | 
|---|---|---|
| Main job | Large-scale AI training and inference | Agentic AI infrastructure | 
| Core hardware | Instinct MI455X or MI450X GPUs, 6th Gen EPYC, Pensando networking | EPYC 9006 server CPUs, optional local GPU in Air Inference | 
| Headline AMD figure | Up to 2.9 exaFLOPS FP4, 1.4 exaFLOPS FP8, 31 TB HBM4 | Up to 73,728 threads per rack in some configurations | 
| Memory | Up to 432 GB HBM4 at 19.6 TB/s per GPU | DDR5 (AMD’s agents-per-rack model uses 16 DDR5-8000 RDIMMs per socket) | 
| Rack standard | Built on Meta’s 2025 OCP Open Rack for AI design | Based on an Open Compute Project specification | 
| Status | Initial systems running customer workloads by November 3 | Configurations planned for November 2026 availability | 
| Price | Not published | Not published | 

The practical split shows up in the memory row. Helios leans on HBM4, which is scarce and expensive, while a Helion rack leans on commodity DDR5 channels. AMD’s materials cite 43 TB/s of scale-out bandwidth for Helios over Ethernet-based Pensando networking, a figure AMD says beats the competition by more than 50 percent. That is a vendor claim, and the materials reviewed show no independent test of it. For the original rack launch details, see [AMD’s Helios rack debut coverage](https://shattered.io/amd-helios-rack-72-gpus-cpu-only-claims-2026/).

Openness is the other thread. Helios builds on Meta’s 2025 OCP Open Rack for AI design, and Helion follows an Open Compute Project specification. That choice lets multiple vendors build compatible racks, which is the opposite of a sealed, single-vendor appliance. Three chipmakers pitched open AI racks at [the OCP 2026 showcase](https://shattered.io/ocp-2026-nvidia-amd-intel-open-ai-racks-2026/), and AMD’s pairing of Helios and Helion fits that pattern.

## The 73,728-Thread Claim Against AMD’s Own Footnotes

AMD says some high-density Helion configurations reach up to 73,728 threads per rack. [TrendForce also carried the figure](https://www.trendforce.com/news/2026/11/11/news-amd-epyc-venice-cpus-begin-shipping-on-tsmc-2nm-verano-reportedly-to-adopt-n2p-with-a14-planned-for-2028/) in its coverage of the Venice shipment start. AMD gives no per-rack breakdown, but its own footnotes allow a sanity check.

The EPYC 9996 has 256 cores and 512 threads with SMT enabled. Divide 73,728 by 512 and you get 144 processors. At 256 cores each, that equals 36,864 cores. This is arithmetic on AMD’s numbers, not an AMD statement, and it assumes the top configuration uses the 256-core part. If AMD fills the rack with smaller chips, the processor count rises accordingly.

Now compare that with AMD’s air-cooled reference. One footnote models agents per rack using EPYC 9996 servers in a 35 kW rack, with a maximum of 20 servers. If those are two-socket servers, the rack holds 40 processors and 20,480 threads. The 73,728-thread configuration carries 3.6 times that, which no 35 kW air-cooled rack can support. It also explains why Helion Liquid exists. A reported 400 W to 600 W range for 256-core SP7 parts puts 144 processors at roughly 57.6 kW to 86.4 kW for the CPUs alone, before memory, networking and fans.

One more number needs care. A report on the CPU-only hyperscale design says it supports up to 32,768 CPU cores. At two threads per core, that equals 65,536 threads, not 73,728. The two figures probably describe different configurations, and AMD has not reconciled them in public. Until it does, treat 73,728 as a ceiling for the densest build and 32,768 cores as a separate, unverified design point.

## EPYC 9006 Venice: What Sits Inside the Rack

The processor generation matters as much as the rack. EPYC 9006 uses Zen 6 on SP7, and AMD’s EPYC 9996 tops the stack at 256 cores and 512 threads. AMD says that count exceeds any other publicly disclosed single-socket CPU, based on a July 2026 comparison. For the Helion buyer, core count per socket sets how many agent sandboxes a rack can host before it runs out of power or memory.

AMD’s embedded 9006 materials, which share the SP7 generation, give a sense of the I/O budget. They list up to 256 cores, 16 DDR5 memory channels, up to 96 PCIe Gen 6 lanes and CXL 3.1, with up to 2.7 times the memory bandwidth and 1.5 times the I/O bandwidth of the prior generation. Server SKUs may differ in detail, so check the OEM spec sheet for the exact part. Still, the direction is clear. AMD is feeding more cores with more memory bandwidth, which is what agent fleets with many small, memory-hungry processes need.

AMD also put the CPU in a supporting role for its accelerators. The company describes EPYC 9006 platforms as providing CPU capacity for agents and for feeding GPUs. That dual role is the economic case for Helion Air Inference. A GPU that waits on a starved CPU wastes expensive silicon, so pairing the two in one rack aims to keep the accelerator busy.

## AMD’s Scorecard: EPYC 9996 vs Xeon 6, Vera and Arm AGI

AMD’s launch footnotes include a six-row comparison of two-socket server configurations. The numbers come from AMD, which chose the competitors, the benchmark and the pricing date. For Nvidia’s Vera and Arm’s AGI CPU, AMD supplies its own estimates because no list price or matching result exists. Read the table as AMD’s argument, not as an audit.

| Two-socket configuration | Total cores | Listed power | Listed price | AMD-listed score | 
|---|---|---|---|---|
| AMD EPYC 9996 | 512 | 600 W | $14,904 | 5,470 | 
| AMD EPYC 9965 | 384 | 500 W | $11,988 | 3,260 | 
| Intel Xeon 6990E+ | 576 | 450 W | $14,995 | 3,050 | 
| Intel Xeon 6980P | 256 | 500 W | $13,955 | 2,510 | 
| Nvidia Vera | 176 | 450 W | Not published | 1,459 (AMD estimate) | 
| Arm AGI | 272 | 300 W | Not published | 1,944 (AMD estimate) | 

The arithmetic is striking. AMD’s 5,470 for the EPYC 9996 is about 3.7 times its 1,459 estimate for Vera and about 2.2 times its 2,510 for the Xeon 6980P. Dollars per score point tell a similar story. The EPYC 9996 works out to roughly $2.72, the EPYC 9965 to $3.68, the Xeon 6990E+ to $4.92 and the Xeon 6980P to $5.56. Those ratios use AMD’s prices dated November 5, 2026, and they cover the processors only, not full servers.

Expect pushback. A vendor-built estimate for a competitor’s part with no published price invites dispute, and the dispute is already underway. The [Vera claim versus AMD’s rebuttal](https://shattered.io/nvidia-vera-cpu-1-8x-claim-vs-amd-2-24x-epyc-venice-2026/) shows how quickly each side reframes the same workloads. Independent SPEC submissions and agent-specific benchmarks will settle the argument, and neither exists for Helion racks today.

## Why Agentic AI Pulls Data Centers Back Toward CPUs

A chatbot answer is mostly one big GPU computation. An agent works differently. It plans, calls a tool, waits for a result, parses it, writes code, runs that code in a sandbox, checks the output and loops. The model call is one step among many, and the rest happens on general-purpose cores. When thousands of agents run at once, those cores become the constraint.

AMD built its Helion pitch on that shift. Its agents-per-rack footnote, tied to a 35 kW rack, frames CPU capacity as the scarce input for agent fleets. The framing suits AMD, which sells both sides of the equation, but the underlying logic holds up. Sandboxed code execution, retrieval, browser automation and orchestration all burn CPU time and memory rather than tensor throughput.

Competitors read the same trend. Arm’s own AGI CPU shows up in AMD’s comparison table, and [Arm’s per-rack claim against the older EPYC 9755](https://shattered.io/arm-agi-cpu-beats-amd-epyc-9755-1-99x-rack-2026/) shows the fight over agent racks started before Venice shipped. Demand signals point the same way. Analysts have forecast growth in server CPU shipments tied to agent workloads, which the [TrendForce shipment outlook](https://shattered.io/server-cpu-shipments-31-percent-2026-trendforce-agentic-ai/) lays out.

The implication for buyers is a change in how to size a cluster. Teams used to count GPUs and treat CPUs as an afterthought. Agent platforms invert that ratio for some workloads. A fleet running browser agents or code interpreters may need hundreds of cores per GPU, and a rack with the wrong ratio leaves one resource idle.

## Who Is Running Helios, and Who Will Run Helion

On November 3, AMD said Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure were running models or workloads on initial Helios systems. That list belongs to Helios, not Helion. In the materials reviewed, AMD has not named Helion customers, so any claim that a specific lab or cloud runs Helion racks today lacks support.

The channel is filling in. HPE said Vultr would deploy the first AMD Helios by HPE systems, a deal covered in [HPE’s Helios agreement with Vultr](https://shattered.io/hpe-1-2-billion-amd-helios-vultr-deal-2026/). Azure lists an ND MI455X v7 instance based on Helios. Oracle’s side runs deeper. AMD reportedly said on October 27 that 6th Gen EPYC would power Oracle Cloud Infrastructure Compute E7 instances, with AMD Pensando technology underpinning Oracle’s Acceleron converged NICs, and the [Oracle Helios plan](https://shattered.io/amd-helios-mi450-oracle-50000-gpus-2026/) sits alongside that.

Google adds a wrinkle. AMD reportedly said on November 9 that its fifth-generation EPYC processors power the host infrastructure for Google Cloud’s TPU 8tx, which would show EPYC winning host-CPU slots even beside a rival accelerator. That claim has a single source, so treat it with care. It still hints at where Helion-style CPU capacity could land first, which is inside hyperscalers that already buy EPYC at volume.

Enterprise buyers who want to evaluate Helion should press OEMs on a short list of questions:

- Which Helion variant does the quote cover, and does it include liquid cooling hardware?
- What rack power and thread count does the quoted configuration reach, with SMT on?
- Which EPYC 9006 SKUs, memory capacity and networking come in the bill of materials?
- When do cloud instances on the same silicon reach general availability?
- Does the vendor offer a benchmark on agent workloads rather than only SPEC-style scores?

## Market Impact: Revenue, Supply and the Nvidia Response

The business backdrop explains why AMD pushed Helion now. In its third-quarter results on November 3, AMD reported a record quarter, with Data Center segment revenue up 97 percent year over year, driven by record sales of Instinct GPUs and EPYC CPUs. The company said demand for both product lines was strengthening and that it expected data center growth to accelerate in the fourth quarter as it scaled Helios and Venice production.

Helion gives that growth story a CPU leg. GPU racks draw headlines, but they depend on HBM and advanced packaging, both of which constrain supply. A CPU rack on DDR5 sidesteps part of that squeeze. AMD’s chief executive has said that [AI chip demand outpaces supply](https://shattered.io/amd-ai-chip-demand-outpaces-supply-2028-2026/), so a product line with a different bottleneck has strategic value even before volumes show up.

On the GPU side, AMD used November 12 to sharpen its competitive claim. It reported 21,333 tokens per second per MI455X GPU on DeepSeek-R1 in an MLPerf Inference Offline configuration, and it claimed up to 30 percent higher throughput than Nvidia’s latest published Vera Rubin result. AMD also introduced MI450X on Helios the same day. Both are AMD statements about its own submissions, and Nvidia has not answered them in the materials reviewed.

Pricing remains the missing variable for every market judgment. Without a Helion list price, analysts cannot model margins, and customers cannot compare against a Vera CPU rack or an Arm-based design. AMD’s own CPU price table is the closest proxy, and it covers only processors. The rest of the rack, including liquid loops, memory, networking and power delivery, will determine whether the $2.72-per-score-point advantage survives at system level. More of this ongoing coverage sits in the [hardware section](https://shattered.io/category/hardware/).

## What AMD’s Chief Executive Says About Agentic Racks

AMD’s top executive has framed agent workloads as the demand driver behind Helion. Lisa Su, AMD’s Chair and Chief Executive Officer, is quoted in Rich Miller’s Data Center Richness newsletter saying [“Agentic workloads have really taken off.”](https://datacenterrichness.substack.com/p/agentic-ai-is-hungry-for-hardware) The line is short, but it matches what the Helion product structure implies. AMD built a CPU rack because it sees agent demand as large enough to deserve its own hardware category.

On the portfolio itself, Su told an interviewer on CEOInterviews.AI that [“We believe these are the best server portfolio for agentic racks because we have the ability to really mix and match for the right workload.”](https://ceointerviews.ai/interview/3051249/) Mix and match is the operative phrase. Liquid, Air and Air Inference let one buyer combine CPU-heavy and GPU-assisted racks, and AMD wants that flexibility to be the selling point against single-configuration rivals.

In the same CEOInterviews.AI conversation, Su described the standards choice behind the Helios and Helion pairing: “I would say we made a very deliberate choice with Helios to use the OCP Open Rack Wide standard.” The choice ties both rack families to the Open Compute ecosystem, which should ease third-party builds. It also exposes AMD to a more competitive field, since any OCP-compliant vendor can pitch against it.

Su also stressed how hard the flagship system was to build. In remarks carried by [Stocktwits](https://stocktwits.com/news-articles/markets/equity/amd-ceo-lisa-su-sees-ai-inference-boom-fueling-1-5t-demand-by-2030-about-60-of-the-market-going-forward/cZxBO5ORBoh), she said, “Helios is absolutely the most sophisticated and complex system that we’ve built at AMD together with our partners.” The complexity point cuts both ways. It signals ambition, and it flags execution risk, which is why AMD’s disclosure of working customer deployments on November 3 mattered to investors.

## History: From Helios in July to Helion in October

The Helion story is a short one, and the timeline shows how fast AMD moved from GPU racks to agent racks. Dates below come from AMD’s announcements and filings, with single-source items flagged.

- **May 21, 2026:** AMD announced a production ramp for Venice EPYC processors on TSMC’s 2 nm process.
- **July 23, 2026:** AMD reportedly launched Helios at its Advancing AI event as a rack combining Instinct MI455X GPUs, Venice CPUs, Pensando networking and ROCm software.
- **October 12, 2026:** AMD introduced Helion as an open blueprint for agentic AI, with CPU-only hyperscale and integrated CPU and GPU enterprise designs.
- **October 27, 2026:** AMD reportedly expanded its Oracle relationship around EPYC 9006 and Pensando networking.
- **November 3, 2026:** AMD reported third-quarter results and named five organizations running workloads on initial Helios systems.
- **November 10, 2026:** General availability of 6th Gen EPYC on SP7, and the three named Helion variants.
- **November 11, 2026:** Supermicro began shipping H15 servers on EPYC 9006.
- **November 12, 2026:** AMD introduced MI450X on Helios and published MI455X inference results.

Two features stand out. First, the gap between announcement and availability shrank. Helion went from blueprint to named variants in under a month. Second, AMD tied every step to open standards, which echoes the multi-vendor rack push around OCP. Helios drew on Meta’s 2025 rack design, and Helion followed an OCP specification, so AMD is building its rack business on shared plumbing rather than a proprietary chassis.

## Five Predictions for Helion Through Early 2027

These are forecasts, not facts, and each carries a stated condition that would prove it wrong.

- **OEM names arrive first.** AMD promised November availability, so expect named server makers to announce Helion-based configurations before year end. If no OEM publishes a Helion SKU by the end of the first quarter of 2027, the November promise slipped.
- **No public price list.** Helion sells as a blueprint built by partners, so quotes will stay private. A published list price in the first half of 2027 would contradict this call.
- **Clouds move before enterprises.** Azure and Oracle instances on EPYC 9006 reach general availability ahead of any sizable enterprise Helion rollout, because hyperscalers already buy EPYC in volume.
- **Air Inference becomes the volume seller.** It fits existing air-cooled halls and combines both workloads, so enterprise buyers will pick it more often than Liquid. Liquid dominates only at hyperscale.
- **Benchmark disputes escalate.** Nvidia, Intel or Arm partners will publish their own agent-density numbers, and the Vera estimate in AMD’s table will draw a formal response.

The wildcard is software. Agent platforms still target GPU-first stacks, and moving a fleet onto CPU-heavy racks requires scheduler and sandbox tooling that treats cores as the scaling unit. If the tooling lags, the hardware will wait for it.

## FAQ: AMD Helion Racks

**What is AMD Helion?**

AMD Helion is an open rack-scale blueprint for agentic AI infrastructure, based on an Open Compute Project specification and powered by 6th Gen AMD EPYC 9006 Series server CPUs. AMD introduced it on October 12, 2026, and defined three variants on November 10.

**How many threads can a Helion rack hold?**

AMD says some high-density configurations reach up to 73,728 threads per rack. The company has not published a full per-rack breakdown, so treat the number as a ceiling for the densest build.

**How much does an AMD Helion rack cost?**

AMD has published no price for Helion Liquid, Helion Air, Helion Air Inference or a complete rack. Pricing will come through OEM and ODM quotes that depend on configuration.

**When can you buy Helion?**

AMD said Helion configurations were planned for availability in November 2026. EPYC 9006 server systems have shipped from OEMs and ODMs since November 10, and Supermicro began shipping H15 servers on November 11.

**How does Helion differ from Helios?**

Helios is AMD’s GPU-centered rack for large-scale training and inference. Helion is the CPU-led blueprint for agent workloads, and Helion Liquid is designed to run alongside Helios.

**Is Venice the same as EPYC 9006?**

Yes. Venice is the codename, and AMD’s launch name is 6th Gen AMD EPYC, with the model family called EPYC 9006 Series.

**Does Helion require GPUs?**

No. Helion Liquid is CPU-only, and Helion Air adds agent capacity to air-cooled data centers. Only Helion Air Inference combines CPU execution with local GPU inference in the same rack.

**Which clouds will offer the new EPYC chips?**

AMD said Microsoft Azure and Oracle Cloud Infrastructure instances on 6th Gen EPYC were coming soon. It gave no dates.
