AMD Shows Helios AI: A Double-Wide Rack Built for a Record Number of Accelerators

AMD has shown a concept for Helios AI, a turnkey server AI infrastructure platform. It is a non-standard double-wide rack, designed to fit the maximum number of accelerators per unit of data center floor space. ServerNews reported the news. According to the platform's previously published specs, a single rack runs 72 Instinct MI450-series accelerators, with a combined HBM4 capacity of about 31 TB. The claimed peak performance reaches 2.9 exaflops in FP4.

The report itself gives few details: this is a concept, not a production product with a price tag and delivery dates. The figures below therefore come from earlier AMD announcements, and they should be checked against the final specification.

Why a rack needs a second width

The standard 19-inch rack has long been running into the limits of physics. Modern AI accelerators need heavy power delivery, thick buses, liquid cooling, and dense network switching, and all of it has to fit into the same enclosure. A double-wide design gives more room for trays, liquid cooling piping, and cable runs. Such a rack is also easier to service.

The main reason lies elsewhere. The more accelerators sit in a single fast-interconnect domain, the better training and inference of large models scale. Moving data between racks costs more, both in latency and in energy. The engineers' task is simple: fit as many accelerators as possible into one rack and link them so they work as one large computer.

What we know about the internals

Based on AMD's data from earlier presentations, Helios is built from the following components:

• 72 Instinct MI450-class accelerators on a new-generation CDNA architecture;
• about 31 TB of HBM4 memory across the rack;
• up to 2.9 exaflops in FP4 and about 1.4 exaflops in FP8;
• next-generation EPYC processors and Pensando network adapters;
• accelerator-to-accelerator interconnect based on the open UALink standard, tunneled over Ethernet;
• liquid cooling as a baseline requirement.

Openness matters separately. AMD is betting on standards rather than a closed proprietary bus. UALink and Ethernet let customers mix hardware from different vendors and avoid locking into a single ecosystem. For hyperscalers, this is a selling point: they do not like depending on a single supplier.

The market: the rack as the unit of sale

This is part of a broader change in model. Previously, the manufacturer sold accelerators and partners built the servers. Now AMD, like NVIDIA with its NVL72 racks, sells the entire system: compute, memory, networking, cooling, and a reference design. For the customer, this shortens the path from order to a running cluster. For AMD, it is a chance to take share from the leader at the level where NVIDIA is strongest: system integration.

Failure here would be costly. Large customers such as OpenAI and Oracle, as previously reported, are planning for thousands of Instinct accelerators. Whether they get the promised capacity timelines depends on how mature the rack turns out to be.

What to do about it

If you are planning an AI cluster for 2027, look at three things. First, whether your data center is ready for a double-wide rack and liquid cooling. This often comes down to floor loading, piping, and electrical power per rack. Second, the maturity of the ROCm software stack for your frameworks. Without it, even the best hardware sits idle. Third, real prices and delivery dates, which the concept does not yet include.

For most companies, Helios will be a matter of renting from cloud providers rather than buying. That is also where to expect the first independent benchmarks for training and inference. They will show what the claimed density is worth.