AMD Helios: A Full Guide to the MI455X Rack, HBM4, UALink over Ethernet, Customers and Limits

AMD launched Helios on July 23, 2026 with a direct challenge to NVIDIA: a 72-GPU rack that it says delivers up to 15% more AI compute, 50% more HBM capacity and 50% more scale-out bandwidth than Vera Rubin NVL72. Three months later, AMD says initial Helios systems are installed and running at OpenAI, Meta, Microsoft and Oracle.

That makes Helios the first serious open-standards alternative to NVIDIA's NVL72 racks to reach customer sites. It also makes it one of the most hyped products in AI infrastructure. So what's actually inside the rack, which numbers are AMD's own projections, and what should a buyer believe?

This full guide covers the MI455X GPU, the Helios rack layout, UALink over Ethernet, power and cooling, the customer deals, how Helios compares with Vera Rubin NVL72, and the honest limits, all as of October 31, 2026. For the previous AMD generation, see Inside Deep Tech's AMD Instinct MI350 full guide.

Last updated: October 31, 2026.

Key Takeaways

• AMD Helios is a rack-scale reference design with 72 MI455X GPUs, delivering 2.9 exaflops of FP4, 1.4 exaflops of FP8 and 31TB of HBM4 per rack.
• Each MI455X carries 432GB of HBM4 at up to 23.3 TB/s and 40.3 petaflops of MXFP4, built on CDNA 5 with TSMC 2nm and 3nm dies.
• The 72 GPUs talk over UALink over Ethernet (UALoE) at 260 TB/s of scale-up bandwidth, switched by merchant Ethernet silicon rather than a proprietary fabric.
• Committed demand is large: 6 gigawatts from OpenAI, 6 gigawatts from Meta and up to 2 gigawatts from Anthropic, each spread across multiple years and generations.
• Honest limit: the headline performance and tokens-per-dollar claims are AMD-modeled projections, and AMD hasn't disclosed how many racks are installed or serving production traffic.

What AMD Helios actually is

Helios isn't a chip and, strictly, isn't even a product AMD sells. AMD describes it as a reference design, not a product for sale: a blueprint that OEM and ODM partners use to build their own branded racks. It's also AMD's first rack-scale AI design, built on Meta's Open Rack Wide (ORW) standard that Meta submitted to the Open Compute Project.

The rack combines four AMD silicon families. MI455X GPUs do the AI math, 6th Gen EPYC "Venice" CPUs act as hosts, Pensando "Vulcano" AI NICs handle scale-out networking, and Pensando "Salina" DPUs handle front-end networking, storage and security. ROCm software ties it together, with AMD claiming Day-0 support for PyTorch, TensorFlow, JAX, ONNX Runtime, vLLM and Triton.

AMD first showed Helios at the OCP Global Summit in October 2025, where the design was tied to the "MI450 Series" name. The shipping rack uses the MI455X, which AMD lists as part of the MI450 Series in its customer releases, including the Anthropic announcement. That naming overlap explains why the OpenAI and Oracle deals say MI450 while the spec sheets say MI455X.

The MI455X GPU: CDNA 5, 2nm and 432GB of HBM4

The MI455X is the engine. AMD's spec page lists 320 billion transistors, 256 work group processors, a 2,400 MHz peak clock and a 192MB L2 cache. The compute dies use TSMC 2nm and the supporting dies use 3nm, and the module is direct liquid cooled only.

Memory is the headline. Each MI455X carries 12 stacks of HBM4 for 432GB at up to 23.3 TB/s. By AMD's comparison on the same page, that's 50% more capacity than the 288GB on each Rubin GPU. Inside Deep Tech's HBM full guide explains why HBM4 supply, not GPU dies, now sets how fast these racks ship.

StorageReview's architecture deep dive adds detail AMD's spec page skips. CDNA 5 drops the MI355X's 256MB Infinity Cache and instead uses two 96MB L2 caches, one per Fabric and Cache Die, with HBM4 rising from eight 1,024-bit stacks to twelve 2,048-bit stacks. Compute dies are hybrid-bonded on top of those base dies.

Source: AMD MI455X spec page; AMD MI355X spec page; Phoronix launch report for the up to 4x MXFP4 and MXFP8 uplift over MI355X.

On AMD's own inference test, the jump is larger than the FLOPS. AMD says MI455X delivers up to 34x higher token throughput at high interactivity and up to 18x lower token cost than MI355X on DeepSeek-V4-Flash. Those are AMD lab numbers at chosen operating points, so read them as a ceiling, not an average.

Inside the Helios rack: 18 compute trays, six switch trays

Each Helios rack holds 18 compute trays with four GPUs each, for 72 GPUs. Per StorageReview, the trays sit in two banks of nine with six switch trays between them, inside an ORW cabinet 1.2 meters wide and 1.3 meters deep. Every GPU-to-switch link is copper through four blind-mate cable cartridges at the rear, so trays slide out without hand recabling.

A compute tray pairs four MI455X modules with one Venice CPU. AMD says Venice offers up to 256 Zen 6 cores and 1.6 TB/s of memory bandwidth. StorageReview reports the reference tray uses a 96-core Venice part that boosts to 5GHz with 1TB of DDR5, a 1:4 CPU-to-GPU ratio, though customers can fit any Venice SKU in the standard SP7 socket.

At rack level, AMD quotes these figures:

Source: AMD Helios product page; AMD Advancing AI 2026 Helios update for the 1.7 PB/s aggregate bandwidth and tray count.

The Phoronix launch report adds two totals that put the scale in perspective: 4,600 CPU cores and 18,000 GPU compute units per rack. Each tray weighs around 170 pounds, which matters for anyone planning floor loading and service carts.

UALink over Ethernet: how 72 GPUs share memory

The scale-up fabric is the most interesting design choice in Helios. Instead of a proprietary switch like NVLink, AMD uses UALink memory semantics carried over Ethernet, called UALoE. Each MI455X has 36 UALoE links at 400 Gb/s each, for 3.6 TB/s of bidirectional bandwidth per GPU.

Those links land on ordinary switch silicon. StorageReview reports each of the six switch trays holds two Broadcom Tomahawk 6 ASICs with 512 lanes of 200G, and each switch terminates 216 links. AMD's own page describes two 512-lane 200G UALoE switch ASICs per switch tray at 21.6 TB/s each. With a single switching tier, every GPU is one fixed-latency hop from every other GPU.

This matters for the UALink story. Native UALink switch silicon isn't what ships in Helios. The rack runs UALink's protocol on Ethernet hardware that hyperscalers already buy at volume, which lowers supply risk but leaves native UALink switches for later. Inside Deep Tech's UALink full guide covers the consortium, the 2.0 spec and who's building native parts.

Scale-out runs over standard Ethernet aligned with the Ultra Ethernet Consortium. StorageReview counts three Vulcano 800G NICs per GPU, 12 NICs per tray and 2,400 Gb/s of scale-out bandwidth per GPU, attached directly to the GPUs so rack-to-rack traffic never crosses the host CPU.

Power, cooling and facility planning

AMD hasn't put a rack power number on its Helios product page. AMD's launch blog says power and thermal details are available to qualified customers and partners under NDA. The best public figures come from partners and reviewers.

• Rack draw. StorageReview reports 225 to 245 kW per rack depending on workload, fed over a 50V liquid-cooled busbar with about 385 liters of coolant per minute from the facility loop.
• Facility blueprint. Schneider Electric and AMD released a reference design for AI racks up to 246 kW and AI clusters up to 10.4 MW of IT load, modeling an annualized PUE of about 1.12 at full load in the greenfield case.
• Cooling split. The Schneider design still pairs the liquid loop with rear-door heat exchangers and fan walls for the air-side heat, so a Helios hall is a hybrid liquid and air design, not a pure liquid room.

Those numbers put Helios in the same facility class as NVIDIA's NVL72 racks. Operators planning a Helios deployment need direct-to-chip liquid cooling, coolant distribution units and floor loading rated for heavy double-wide racks. Inside Deep Tech's data center liquid cooling full guide covers the CDU and facility-water choices in detail.

Who is buying Helios: OpenAI, Meta, Oracle, Microsoft and Anthropic

Helios arrived with more committed demand than any previous AMD data center GPU. The deals are large, multi-year and, in two cases, paired with warrants or equity investments.

Source: AMD and OpenAI release; Oracle and AMD release; AMD and Meta release; AMD and Microsoft release and Microsoft blog; AMD and Anthropic release.

AMD's CFO said the OpenAI deal alone was expected to bring tens of billions of dollars in revenue to AMD. Both the OpenAI and Meta warrants vest in tranches, starting with the first gigawatt, and also depend on AMD's share price and on each customer hitting technical and commercial milestones.

The latest milestone came on October 27, 2026. AMD said initial Helios systems are installed and being tested by Meta on the path to broader deployment, that Meta will deploy a custom Instinct GPU optimized for its workloads, and that AMD will invest $1 billion in Meta's equity offering.

AMD Helios vs NVIDIA Vera Rubin NVL72

Both racks put 72 GPUs in one scale-up domain with HBM4 and liquid cooling, so the comparison is unusually direct. The vendors' public numbers line up like this.

Source: AMD Helios product page and AMD Helios update; NVIDIA Vera Rubin NVL72 page.

The table explains AMD's "15% more compute" claim. AMD compares its 2.9 EF of dense FP4 with NVIDIA's 2,520 petaflops of NVFP4 training compute, not with NVIDIA's 3,600 petaflop inference figure, which NVIDIA quotes with its own Transformer Engine assumptions. Both choices are defensible, and both are marketing.

Memory is where Helios clearly leads on paper: 31TB versus 20.7TB of HBM4 per rack, a 50% gap that matters for very large mixture-of-experts models and long-context KV caches. AMD's own update also claims only 6% more HBM bandwidth than the competing rack, which suggests AMD is comparing against a higher NVIDIA bandwidth figure than the 1,400 TB/s on NVIDIA's current page.

On economics, AMD claims Helios delivers up to 30% more tokens per dollar than the competitive solution, based on AMD estimates with projected pricing. NVIDIA's side of the story, including Rubin's software lead, is in Inside Deep Tech's Vera Rubin NVL72 full guide.

NVIDIA Vera Rubin NVL72: A Full Guide

NVIDIA's 72-GPU Rubin rack: the two-die Rubin GPU, HBM4, the Vera CPU, NVLink 6, facility planning and honest limits.

Inside Deep Tech

The spec drift buyers should notice

Helios's numbers have moved since the 2025 preview, and AMD's own pages don't all agree yet.

• Memory bandwidth went up. The October 2025 OCP blog listed 19.6 TB/s per GPU and 1.4 PB/s per rack. The July 2026 launch lists 23.3 TB/s per GPU and 1.7 PB/s per rack.
• Some AMD pages still show the old figure. The compute-tray section of the Helios product page still describes 19.6 TB/s per GPU, while the same page's spec list says 23.3 TB/s.
• Rack bandwidth rounding differs. AMD's MI400 Series page gives 1.67 PB/s, which AMD rounds to 1.7 PB/s elsewhere.

None of this is unusual for a product that moved from preview to launch in nine months. But it's a reason to put the exact figures in a purchase contract, not to rely on a web page.

Who should buy Helios, and who should wait

Consider it now if you run large-model inference where memory capacity is the constraint, such as large mixture-of-experts models, long-context agents or many concurrent model replicas, and your team already runs PyTorch or vLLM on ROCm. The extra HBM per rack is the clearest advantage, and Azure's planned ND MI455X v7 VMs offer a path to test it without buying racks.

Wait if your stack depends on CUDA-only kernels, custom NVLink-tuned collectives or NVIDIA's inference tooling. Porting cost can erase a tokens-per-dollar edge. Also wait if your facility can't support liquid-cooled racks near 246 kW.

Watch it if you're an enterprise or sovereign buyer. AMD's HPC-focused MI430X, with 288 TFLOPS of FP64, is expected in 2027, and Helios-based systems through OEMs such as HPE and partners such as TCS should widen access beyond hyperscalers.

Honest limits

• Performance claims are modeled. The tokens-per-dollar and throughput comparisons with Vera Rubin are AMD Performance Labs estimates using projected pricing, run on pre-production or reference hardware.
• Deployment scale is undisclosed. AMD says systems are installed and running at four customers, but hasn't said how many, where, or whether they serve production traffic.
• Power is under NDA. AMD says actual power and thermal data are shared only with qualified customers; the 225 to 245 kW range comes from StorageReview, not an AMD spec sheet.
• Gigawatt deals are long-dated. The OpenAI and Meta agreements span multiple GPU generations, and warrant vesting depends on milestones on both sides.
• Software is the swing factor. ROCm has improved, but independent, third-party benchmarks of Helios at rack scale aren't public yet.

Inside Deep Tech's take

Inside Deep Tech's take: Helios is the most credible challenge to NVIDIA's rack business so far, because AMD matched NVIDIA's unit of sale and then chose open parts where NVIDIA uses proprietary ones. The memory lead is real on paper, and the customer list, with OpenAI, Meta, Oracle and Microsoft systems installed, is stronger than any previous AMD launch.

The open questions are execution questions. Can AMD and its ODMs ramp a rack with this many connectors and this much liquid plumbing at hyperscale volume? Can ROCm close the software gap fast enough that buyers see the modeled economics in real fleets? The next two quarters will answer more than any spec sheet.

Helios competes on packaging and memory as much as on GPUs. For the packaging line behind both racks, see the TSMC CoWoS guide, and for the NVIDIA CPU memory Helios is measured against, the SOCAMM guide.

What to watch next

• AMD's Q3 2026 earnings on Tuesday, November 3, 2026: any Helios revenue, rack shipment counts or customer production dates would be the first hard data since launch.
• Meta's broader deployment after the initial systems now in test, per AMD's October 27 update.
• Oracle's 50,000-GPU OCI supercluster, which was scheduled to start in calendar Q3 2026, and when it becomes publicly available to customers.
• MI500 in 2027 and MI600 in 2028, which AMD reaffirmed at launch, and whether Helios's ORW rack carries forward to them.

FAQ

The bottom line

AMD Helios gives AI buyers a real second option at rack scale: 72 GPUs, more HBM4 than Vera Rubin NVL72, an Ethernet-based scale-up fabric and gigawatt commitments from OpenAI, Meta and Anthropic. The specs are strong, the economics are still AMD-modeled, and the deployment data is thin. Treat Helios as a serious 2027 capacity option and test it on real workloads before betting a fleet on it.

For the previous generation and how AMD got here, read Inside Deep Tech's AMD Instinct MI350 full guide, and for the scale-up standard behind the fabric, the UALink full guide.