AMD has rolled out a rack-scale platform called Helios and is taking direct aim at the only real alternative on the market, Nvidia's Vera Rubin NVL72 system. According to TechRadar's reporting, AMD framed Helios around five head-to-head comparisons and walked away claiming wins on most of them, though some of those wins hinge on how you count GPUs and which arithmetic format you use.
For anyone running WordPress shops, packaging cPanel hosting, or moving clients onto bigger infrastructure, this is mostly background noise today and could turn into real noise by the end of the year.
At a glance:
- Helios pairs 72 MI455X accelerators with 18 EPYC Venice CPUs in an Open Rack Wide chassis.
- AMD claims 31TB of HBM4 per rack against roughly 20.7TB on Nvidia's Vera Rubin NVL72.
- On per-GPU FP4 compute, AMD claims a 15% lead; on full rack numbers, Nvidia still publishes higher.
- AMD also claims 30% more tokens per dollar versus the rival it declined to name.
- First Helios deployments are slated for Q4; Vera Rubin has been in full production since Q1.
- Inter-rack links rely on merchant Broadcom switch silicon, with UALink carried over Ethernet inside the chassis and Ultra Ethernet used to reach the next rack.
What's actually different inside the rack
The single most concrete AMD claim is the memory figure. 31TB of HBM4 against about 20.7TB on the Nvidia side is a real gap, and it is all HBM on AMD's box. Nvidia pads its number out to roughly 75TB per rack only when you add 54TB of LPDDR5X to the HBM, but that slower memory is not the same tool for the job.
Per accelerator, AMD lists 432GB of HBM and 23.3TB/s of bandwidth. Nvidia matches on scale-up fabric at 3.6TB/s per accelerator and 260TB/s per rack. On raw fabric speed, the two systems are even. The difference is where memory sits and how open the design is.
Where the claims get slippery
The FP4 fight is the one to watch with a skeptical eye. AMD says it beats Nvidia by 15% per GPU on FP4, but its own 72-GPU rack lands at 2.9 exaflops against Nvidia's published 3.6 exaflops. AMD does not spell out the unit of comparison, and the two sides use different numeric formats, MXFP4 on AMD's side and NVFP4 on Nvidia's. Against individual dies, AMD's lead widens. Against two-die packages, AMD trails.
Both vendors agree real-world FP4 rarely hits peak. AMD itself told its audience that measured throughput lands at roughly half the rated peak because of memory movement, kernel efficiency and scheduling overhead. So treat the press slide and the production cluster as two different animals.
Our analysis
Helios is the first time AMD has put up a credible full-rack answer to Nvidia since the AI build-out started. The hardware list is serious, and the openness angle is the real story. Helios publishes its fabric specs in the open, which lets hyperscalers and large builders customize variants. Nvidia's rack does not offer that latitude.
The catch is supply. Vera Rubin is already shipping in volume with partner availability this half. Helios is not due in customer racks until Q4. On top of that, HBM4 supply is tight, and a lot of it is already spoken for by the largest buyers. AMD could end up with a strong platform and not enough of it to matter in 2025.
Who wins on paper: AMD's architects. Who wins on volume this year: Nvidia, almost certainly. Who benefits regardless: buyers who want a second source and open fabrics, because AMD's existence keeps pressure on Nvidia's pricing and roadmap.
Worked example for a hosting reseller
Say you run a small agency on reseller hosting with cPanel and you have two AI-adjacent clients eyeing GPU-backed inference. You do not buy racks yourself, but your upstream provider might. If your provider leans toward Helios, ask three things before signing a longer contract: what memory tier your workloads will actually sit in (HBM4 or LPDDR5X), what the measured FP4 throughput looks like on the kernel your client uses, and whether the deployment date is a real Q4 or a roadmap slide. If your provider sticks with Vera Rubin, ask about HBM4 allocation and whether your client's jobs will land on a die or a package, since that changes the price-per-token math. Either way, treat any 30% tokens-per-dollar claim as a ceiling, not a guarantee.
What to watch next
Watch HBM4 allocation lists. If AMD gets a real share outside the top hyperscalers, reseller-facing clouds will have a second option by mid-2026 and pricing will follow. If HBM4 stays locked up, Helios stays a press-release winner for another year. Bottom line: AMD has earned the right to be on the shortlist, but resellers should plan around deployment dates, not slide decks.
Key takeaways
- Helios wins on HBM4 capacity per accelerator and on openness; Nvidia wins on shipped volume and full-rack FP4 numbers today.
- FP4 comparisons between the two vendors use different number formats and different die-versus-package counting, so treat percentages with caution.
- Real-world FP4 throughput is roughly half the peak rating on both sides, which matters more than the headline exaflop figure.
- Resellers should ask upstream providers about memory tier, measured throughput on real kernels, and actual deployment timing before committing.
Frequently asked questions
When will AMD Helios racks actually ship to customers?
According to TechRadar, first Helios customer shipments are lined up for Q4, while Nvidia's Vera Rubin has been rolling off production lines since the first quarter.
Is AMD's 15% per-GPU FP4 lead a fair comparison?
Not fully. AMD's lead is per GPU while Nvidia publishes higher per-rack numbers, the arithmetic formats differ (MXFP4 vs NVFP4), and AMD has not stated whether it is counting Nvidia's dies or two-die packages.
Why does HBM4 capacity matter more than total memory?
HBM is the tier models need to stay resident in for fast inference. Nvidia's larger 75TB figure includes 54TB of slower LPDDR5X, which is not a substitute for HBM on heavy workloads.
Does the open fabric story change anything for small resellers?
Indirectly. If hyperscalers adopt Helios because the fabric specs are public, reseller-facing clouds get a second GPU source, which usually pressures price and roadmap concessions from Nvidia.
Should resellers plan around Helios now?
Not yet. Plan around confirmed deployment dates and measured kernel performance, not the launch slides.