Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

AI & Workstations

AMD Helios vs NVIDIA NVL72: Is the 72-GPU Rack Finally a Two-Horse Race?

For two years, "rack-scale AI" meant one thing: an NVIDIA NVL72. Seventy-two GPUs, one NVLink domain, one supplier. AMD's Helios is the first design built to compete at exactly the same size, and HPE sells both. So is this a real race, or a second horse that has not left the stable?

Our view: Helios is credible on hardware and unproven on everything else. That is still the most competition this market has seen.

Helios vs NVL72 at a glance

Compare AMD
Helios AI Rack by HPE
NVIDIA
NVL72 by HPE
🧠 GPUs per rack 72 × AMD Instinct MI455X 72 × Blackwell Ultra (GB300) or 72 × Rubin (Vera Rubin)
💾 Memory per GPU 432 GB
288 GB
🔗 GPU-to-GPU link inside the rack Open standard
UALink over Ethernet
Proprietary
NVLink (NVIDIA's own)
🧰 Software AMD ROCm NVIDIA CUDA
📏 Rack format OCP Open Rack Wide (double-width) Conventional-width rack
💧 Cooling Liquid
Direct liquid
Liquid
Direct liquid, small air share
📦 In customers' hands Early
First HPE order announced Sept 2026
Shipping
GB200 and GB300 shipping

Sources: HPE AI rack-scale systems, StorageReview on MI455X and Helios, The Register.

Where Helios looks strong

Memory, by a wide margin

The MI455X carries 432 GB of HBM4 per GPU, half as much again as the 288 GB on NVIDIA's B300 and Rubin parts. HPE quotes up to 31 TB of GPU memory per Helios rack. For inference on very large models, where memory is usually the first limit, that is a meaningful advantage.

Open plumbing

Inside an NVL72, the GPUs talk over NVLink, which only NVIDIA makes. Helios uses UALink carried over standard Ethernet, with a purpose-built HPE Juniper switch on Broadcom silicon. The rack itself follows the Open Compute Project's Open Rack Wide specification. In principle, more of the system can come from more than one supplier.

A buyer with real money

On 30 September 2026, HPE announced its first Helios order: a US$1.2 billion deal with cloud provider Vultr. One order is not a market. But it moves Helios from roadmap slide to purchase order.

Where the questions are

AMD's numbers are AMD's numbers

AMD says Helios delivers up to 15% more AI compute and 50% more memory than a Vera Rubin NVL72, and up to 30% more inference tokens per dollar than a competing system. These are vendor claims about systems that are only starting to ship. Independent benchmarks at rack scale are not yet available, and paper specifications do not guarantee real-world performance.

Sources: AMD Helios launch material via EE Times Asia, Data Center Frontier

The software moat has not gone anywhere

Most AI teams build on CUDA. Moving to ROCm is possible and getting easier, but it is still a migration with a cost. We looked at that question in detail in NVIDIA vs AMD for Enterprise AI: Is CUDA Lock-In Becoming a Business Risk?

The rack is twice as wide

This is the detail that gets lost in spec comparisons. Open Rack Wide is a double-width format. A data hall laid out for conventional racks cannot simply slot one in. For hyperscale campuses designed around the format, that is fine. For a colocation customer in Singapore renting standard cage space, it is a conversation to have with the operator before anything else.

So, is it a two-horse race?

If you are… Our read
☁️ A GPU cloud or large inference provider Yes Helios is a real negotiating lever and a real option.
🏢 An enterprise with CUDA-based applications Not yet Software migration outweighs the hardware gain.
🏛️ A research or sovereign AI programme Evaluate Worth a formal evaluation, especially if supplier diversity is a goal.
📍 Hosting in standard colocation space Check first Check the rack format with your operator first.

The honest answer is that competition has arrived at the top of the market, and buyers benefit even if they never order a Helios. A supplier with a credible rival quotes differently from a supplier without one.

Related on SourceIT

Share This Article
Facebook

More from our blog


← Older Post Newer Post →