HPE has its first AMD Helios rack order and OpenAI is testing Helios in the lab, so AMD GPUs are back in every AI infrastructure conversation. Helios is a 72-GPU liquid-cooled rack for hyperscale buyers; for everyone else the real choice is an 8-GPU MI355X, MI350X or MI325X server. Here is how the parts and platforms line up.
AMD GPUs have been the most persistent thread in the AI infrastructure conversation this week. HPE announced its first order for the AMD Helios rack, a $1.2 billion deal with a US cloud provider, and raised its networking outlook on the back of it. AMD has been showing Helios systems running in OpenAI's lab, and it is recirculating a production inference case study built on MI325X servers at DigitalOcean. Hosting providers and enterprise teams we speak to are asking the obvious follow-up: if AMD is now a credible second source to NVIDIA, what can we actually buy, and does it have to be a rack-scale system? For almost everyone below hyperscale the answer is no. Helios is a 72-GPU, liquid-cooled, double-wide rack. The buyable unit today is an 8-GPU server built on AMD's Universal Base Board, and there are three current Instinct parts to choose between.
What is driving the conversation
- HPE's Helios order (announced 30 September 2026) is the first for the AMD Helios AI Rack by HPE. Each rack carries 72 Instinct MI455X GPUs with EPYC "Venice" CPUs and Pensando Vulcano AI NICs, linked by six HPE Juniper Networking QFX5252 scale-up Ethernet switch trays running UALink over Ethernet, with direct liquid cooling and HPE deployment services.
- AMD says OpenAI has had Helios systems for several months and expects to bring Helios online through multiple deployment partners from the second half of 2026, ramping through 2027, as the first phase of the 6 GW deployment the two companies agreed in October 2025.
- The inference case study AMD is promoting dates from January 2026: on 8-GPU MI325X servers at DigitalOcean, a Qwen3-235B FP8 workload reached roughly twice the request throughput per server of the customer's previous generic setup, under fixed latency targets. The gain came from software tuning (parallelism layout and serving configuration) as much as from the hardware.
The three Instinct parts you can buy as servers today
All three are OAM modules on an 8-GPU UBB 2.0 baseboard with seven Infinity Fabric links per GPU and a PCIe Gen5 x16 host link. MI325X is drop-in compatible with the MI300X platform. MI350X and MI355X share the same silicon and memory; the difference is power, and therefore sustained performance and cooling.
| GPU | Architecture | Memory | Peak bandwidth | Typical board power | 8-GPU HBM total |
|---|---|---|---|---|---|
| MI300X | CDNA 3 | 192 GB HBM3 | 5.3 TB/s | 750 W | 1.5 TB |
| MI325X | CDNA 3 | 256 GB HBM3E | 6 TB/s | 1000 W | 2 TB |
| MI350X | CDNA 4 | 288 GB HBM3E | 8 TB/s | 1000 W | 2.3 TB |
| MI355X | CDNA 4 | 288 GB HBM3E | 8 TB/s | 1400 W | 2.3 TB |
| MI455X (Helios only) | CDNA 5 | 432 GB HBM4 | n/a | n/a | 31 TB per 72-GPU rack |
- MI355X peak figures from AMD: 10.1 PFLOPS MXFP4 and MXFP6, 5 PFLOPS OCP-FP8 dense, 2.5 PFLOPS FP16 dense, 78.6 TFLOPS FP64, 256 MB Infinity Cache.
- The CDNA 4 parts add MXFP4 and MXFP6 support, which matters for low-precision inference on current open models. CDNA 3 parts (MI300X, MI325X) top out at FP8.
- AMD sells the MI455X only inside the Helios rack, so it is not an option for a single-server purchase.
8-GPU AMD Instinct platforms by vendor
| Vendor model | Height | GPU options | Cooling | CPUs and memory | Power supplies |
|---|---|---|---|---|---|
| HPE ProLiant Compute XD685 | 5U (DLC) or 6U (air) | 8x MI355X, or NVIDIA B300, B200, H200 | MI355X configurations are direct liquid cooled; HPE lists air cooling for H200 only | 2x EPYC 9005, 24 DDR5-6400 RDIMM | Per HPE QuickSpecs |
| Dell PowerEdge XE9785 | 10U | 8x MI355X 288GB 1400W OAM, or 8x NVIDIA HGX B300 | Air cooled | 2x EPYC 9005 up to 192 cores each, 24 DDR5 RDIMM up to 6 TB | 12x 3200 W Titanium |
| Supermicro AS-8126GS-TNMR | 8U | 8x MI325X or MI350X | Air cooled | 2x EPYC 9005/9004 up to 500 W, 24 DIMMs up to 6 TB | 6x 5250 W Titanium (3+3) |
| Supermicro AS-4126GS-NMR-LCC | 4U | 8x MI325X or MI355X | Direct to chip liquid cooling, onsite service required | 2x EPYC 9005/9004 up to 500 W, 24 DIMMs up to 6 TB | 4x 6600 W Titanium (2+2) |
Specifications are from HPE, Dell, Supermicro and AMD product pages and QuickSpecs. Exact GPU kit and option part numbers vary by region and configuration, so we confirm them on every quote rather than publish a generic list.
What to check before you choose
- Rack power. Eight MI355X GPUs alone are rated at 11.2 kW of board power before CPUs, memory, NICs and fans; eight MI350X or MI325X are 8 kW. Most UK colocation footprints are still provisioned at 6-12 kW per rack, so one MI355X server can fill a rack on its own.
- Cooling. If the hall has no liquid loop, the realistic options are the air-cooled Dell XE9785 with MI355X or an air-cooled MI350X or MI325X system. HPE's XD685 with MI355X and Supermicro's 4U MI355X system both need direct liquid cooling.
- Memory per model. 288 GB per GPU on MI350X and MI355X means 2.3 TB of HBM per 8-GPU server, enough to hold many large open-weight models on one node at FP8 without spilling across servers.
- Software. Check that your serving stack and model versions are validated on ROCm before ordering. The published throughput gains on Instinct came from tuning, not from swapping the GPU alone.
- Upgrade path. MI325X drops into MI300X platforms, so existing MI300X estates can step up memory without a new chassis design.
The ETON view
Helios is real, but it is a hyperscale and neocloud product: 72 GPUs, a double-wide Open Rack Wide frame, liquid cooling and a first order worth $1.2 billion. For a UK hosting provider, AI startup or enterprise inference project the question is not Helios versus NVIDIA's rack systems, it is which 8-GPU server fits the power and cooling you already have.
Our view on the trade-offs. MI355X gives the most performance per server but needs a room that can deliver well over 11 kW to a single box, and in most UK halls that means liquid cooling or a dedicated high-density row. MI350X and MI325X are the practical air-cooled choices and the easiest to place in existing colocation. MI300X and MI325X platforms are also the parts most likely to appear on the secondary market first as early AMD adopters refresh, which is where the best cost per token for steady inference will come from over the next year.
On availability, GPU servers are allocated rather than shelf stock across every vendor right now, and neither HPE nor its first Helios customer has published a rack count or delivery schedule. Being vendor neutral, we quote the same AMD configuration across HPE, Dell and Supermicro and tell you which one can actually ship, alongside the NVIDIA HGX equivalent so the comparison is like for like. When you refresh, we also buy back the outgoing GPU servers, which helps fund the next step.
Related infrastructure
Category: AI & GPU · Vendor: AMD, HPE, Dell, Supermicro · Technology: AMD Instinct MI355X, MI350X, MI325X, MI455X Helios · Last verified: 03 Oct 2026 · ~6 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
