AMD used its Advancing AI 2026 event to put a number on the table that NVIDIA hadn't yet answered: a 72-GPU Helios rack built around the new MI455X accelerator, matched directly against NVIDIA's own Vera Rubin NVL72. Neither rack is rentable by the hour anywhere today, which changes what "better" means here. It's less a spec fight and more a question of which open-standard interconnect bet, and which pre-order queue, you want to be in before either system ships.
TL;DR: AMD MI455X vs NVIDIA Vera Rubin, Which Should You Plan Around?
Neither rack rents by the hour yet (Oct 2026), so the call is reservation timing and interconnect bet, not which one wins on paper.
| Spec | AMD MI455X (Helios) | NVIDIA Rubin (Vera Rubin NVL72) |
|---|---|---|
| Memory per GPU | Not yet published at GPU level | 288GB HBM4 at 22TB/s (Oct 2026) |
| Inference compute | AMD claims 10-15% more tokens/sec/GPU vs Rubin, Kimi K2 Thinking test | 50 PFLOPS NVFP4 (Oct 2026) |
| Scale-up interconnect | UALink, an open standard | NVLink 6, 3.6TB/s per GPU, closed |
| Rack price (3rd-party est., Jan 2026) | $2.4M-$5.5M | $3M-$7M |
| Reserve on Spheron today | No listing | Pre-order only, ships H2 2026 |
Spheron rents B200 on-demand now, the closest live alternative while both racks stay pre-order or unlisted.
What AMD Announced at Advancing AI 2026: Helios and MI455X
AMD's Advancing AI event positioned Helios as the company's direct rack-scale answer to NVIDIA's NVL72 design, built around the new MI455X accelerator. The framing was deliberate: not a chip spec sheet but a full rack with its own interconnect fabric, memory pool, and networking stack, the same unit of comparison NVIDIA has used since GB200 NVL72. StorageReview's reporting on the launch frames Helios explicitly as AMD's answer to Vera Rubin, not to the shipping B300: AMD is trying to land ahead of NVIDIA's next platform, not catch up to its current one.
That's a shift from the MI300X era, when AMD's pitch was mostly "cheaper than H100 with good enough throughput." Helios bets that the rack, not the chip, is the real unit of competition, and that an open interconnect standard can take share from NVLink the way Ethernet eventually did against proprietary fabrics. Our AMD Helios rack-scale guide covers the deployment side, including the UALink and Ultra Ethernet stack.
Spec for Spec: HBM4 Capacity, Bandwidth, and FP4/FP8 Compute
Both platforms move to HBM4 for the first time, and both lean on FP4 as the default inference format. Here's what each side has published, with the caveat that neither vendor has shipped silicon to a cloud customer, so every number below is a preliminary vendor spec, not an independent benchmark.
| Metric | AMD MI455X (Helios) | NVIDIA Rubin GPU (Vera Rubin NVL72) |
|---|---|---|
| Memory type | HBM4 | HBM4 |
| Memory per GPU | Not yet independently confirmed at GPU level | 288GB |
| Memory bandwidth per GPU | Not yet independently confirmed at GPU level | Up to 22TB/s |
| Inference compute | AMD claims 10-15% higher tokens/sec per GPU than Rubin, per its own Kimi K2 Thinking modeling | 50 PFLOPS NVFP4 |
| Training compute (dense) | Not published by AMD at GPU granularity | 35 PFLOPS NVFP4 |
| Scale-up interconnect | UALink (open standard) | NVLink 6, 3.6TB/s per GPU |
| Rack GPU count | 72 MI455X per Helios rack | 72 Rubin GPUs per NVL72 rack |
| Rack price | Not published; third-party estimates range $2.4M-$5.5M | Not published; third-party estimates range $3M-$7M |
The gaps in this table say as much as the numbers. NVIDIA published a GPU-level spec sheet for Vera Rubin with per-GPU figures for memory, bandwidth, and compute. AMD's public materials lean on rack-level and relative-performance claims rather than a matching per-GPU datasheet, which is why the MI455X column has more blanks. That's not evidence on which chip is faster, it's evidence AMD launched with a benchmark argument instead of a spec sheet, and NVIDIA did the opposite. Our Vera Rubin NVL72 guide has NVIDIA's full published spec table, including NVLink 6 rack-aggregate bandwidth and training throughput at the full 72-GPU scale.
Rack prices for both remain third-party estimates rather than vendor list prices. Reporting on Helios puts rack cost anywhere from $2.4-3.2 million up to roughly $5-5.5 million, and estimates for Vera Rubin NVL72 range from $3-4 million to $5-7 million, according to Introl's CES 2026 coverage. A 2x spread inside a single range tells you how early this market still is; treat any specific rack price you see elsewhere as a guess dressed up as fact.
Rubin vs Blackwell: Why Vera Rubin Isn't the GPU in Market Today
The most common confusion in this comparison isn't AMD vs NVIDIA, it's Rubin vs Blackwell. A lot of search traffic for "Vera Rubin" is actually looking for the B200 or B300 shipping right now, because the generational jump hasn't fully registered. It's worth being precise about what changed.
Blackwell (B200/B300) is the generation currently renting on every major GPU cloud, built on HBM3e memory and NVLink 5. Vera Rubin is the next architectural generation after Blackwell, not a faster version of it: it moves to HBM4, pairs each Rubin GPU with a new Vera CPU instead of Grace, and NVIDIA's own preliminary numbers put a single Rubin GPU at 50 PFLOPS of NVFP4 inference compute and 35 PFLOPS of dense NVFP4 training compute, figures not yet matched against an independent B200 benchmark. Our Rubin vs Blackwell vs Hopper guide walks the full three-generation comparison if you need Hopper in the picture too.
The part that matters for a buying decision: Blackwell is a known quantity with real fleet data behind it, Rubin is a vendor spec sheet. Nobody outside NVIDIA and its first cloud partners has run a production workload on a Rubin GPU yet. Any claim you read about Rubin's real-world throughput, including AMD's own numbers below, is a model built on NVIDIA's published specs, not a measurement taken on hardware.
Open UALink vs NVLink 6: Two Different Scale-Out Bets
The interconnect choice has little to do with raw chip speed and everything to do with who controls the next five years of rack design. Helios scales its 72 MI455X GPUs with UALink, an open standard whose backing consortium includes AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft. Vera Rubin NVL72 scales its 72 Rubin GPUs with NVLink 6, which runs at 3.6TB/s per GPU and remains entirely NVIDIA's closed fabric, per NVIDIA's own spec page. An open standard means more vendors can eventually build switch silicon for it, the same argument that let Ethernet outlast every proprietary data-center fabric before it. A closed fabric means NVIDIA controls the entire upgrade path and keeps switch margin in-house, which is also why NVLink has stayed ahead on raw bandwidth generation after generation.
What neither vendor's marketing tells you is how much the size of that scale-up domain matters in production, and that's something current-generation Blackwell hardware can speak to directly. As of October 2026, splitting an 8xB300 node into two 4-GPU copies gave lower per-user speed than running one full 8-GPU copy on 8xB200: on GLM-5.3, that was 254 tokens per second per user on the split B300 configuration against 324 tokens per second on the single B200 domain. The trade wasn't a straight loss: the B300 node served twice the concurrent agent sessions, 64 versus 32, because splitting the domain let it run two independent copies of the model instead of one. That tension is exactly what a bigger scale-up domain is supposed to solve, and it's what both UALink and NVLink 6 are being built to push further out before either Helios or Vera Rubin reaches a customer's rack.
If you're planning around either platform, that data point is the one to hold onto: domain size isn't an abstract interconnect spec, it's a direct trade between single-user latency and how many concurrent sessions a rack can serve. A 72-GPU UALink domain and a 72-GPU NVLink 6 domain are both attempts to make that trade-off disappear at larger scale; neither vendor has published fleet data proving it does yet.
AMD's Own Benchmark Claim: Kimi K2 Thinking Tokens per Dollar
AMD's headline performance claim for this launch is specific and worth examining on its own terms. According to StorageReview's coverage of the Helios and MI455X launch, AMD's internal benchmarks claim 10-15% more tokens per second per GPU on the Kimi K2 Thinking model, and up to 30% more tokens per dollar, against NVIDIA's published Vera Rubin NVL72 figures.
Read that claim for what it is: a modeled comparison against a competitor's published spec sheet, not an independent benchmark run on both systems side by side. That's a normal way for a chipmaker to open a launch, NVIDIA has done the same with its own cross-generation comparisons, but it means the 10-15% and 30% figures are AMD's best-case projection against NVIDIA's best-case projection. Until both systems are in independent hands running the same model on the same framework, treat the comparison as directional. If you're building a 2027 capacity plan on either vendor's launch-day benchmark slide, build in enough margin to be wrong by double digits either way.
Where You Can Actually Reserve Either One Today
This is the question every spec comparison skips, and the one that actually determines what you do this quarter. Checked against Spheron's own pricing page, here's where things stand:
- Vera Rubin: listed as pre-order only, with NVIDIA targeting H2 2026 for shipment. You can register interest now; there's no confirmed hourly rate.
- MI455X / Helios: not listed at all. No pre-order page, no reservation queue, nothing to put money against yet.
- B200: rentable today at $11.12/hr on-demand and $5.25/hr spot as of 17 Sep 2026.
- B300: rentable on Spheron's B300 page at $5.40/hr on-demand and $5.84/hr spot as of 17 Sep 2026.
- MI300X: the closest AMD option available now, on-demand at $3.59/hr as of 17 Sep 2026, running the same ROCm stack MI455X will eventually use.
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 17 Sep 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing for live rates.
Neither MI300X nor B300 substitutes for the HBM4 capacity or bandwidth this comparison is actually about; they're a practical stopgap, not a preview of either next-gen rack's real performance. If your workload genuinely needs Rubin-class memory bandwidth or MI455X-class capacity today, there isn't a rental path for either one yet. Our Vera Rubin cloud availability and cost-per-token guide tracks the Rubin rollout timeline if you want to watch that side specifically.
Which One Should Shape Your 2027 Infrastructure Plan
Neither Helios nor Vera Rubin NVL72 is a decision you can make with money today, so the real planning question is which bet to lean toward for next year. If your organization already runs CUDA and can't justify a software migration, Vera Rubin is the lower-friction path: the same programming model you already use, backed by a published, if preliminary, spec sheet. If you're already running ROCm in production, or building new infrastructure and want to avoid locking into a single vendor's switch silicon for the next refresh cycle, Helios and UALink are the more defensible long-term bet, with the caveat that AMD's performance claims aren't independently verified yet. Our ROCm vs CUDA on GPU cloud guide goes deeper on that software-stack trade-off.
Either way, don't let the launch-day spec war delay a decision you can act on this quarter. B200 and B300 are real, rentable Blackwell-generation capacity today, and the domain-size trade-offs measured on that hardware are the closest real preview either camp has of what their next-gen racks will have to solve. Our Vera Rubin NVL4 vs NVL72 decision guide is a useful next read once Rubin capacity opens up, for sizing the right form factor instead of defaulting to the full rack.
Rubin and Helios are both still pre-order hardware, but Blackwell-generation capacity isn't. Check MI300X availability for an on-demand AMD path today, or explore the full GPU rental catalog to reserve capacity while you wait on either next-gen rack.
Frequently Asked Questions
Blackwell (B200/B300) is the generation currently renting on GPU clouds, built on HBM3e memory and NVLink 5. Vera Rubin is the next architectural generation after Blackwell, not a faster Blackwell variant: it moves to HBM4 memory running up to 22TB/s per GPU, scales up with NVLink 6 at 3.6TB/s per GPU, and NVIDIA rates a single Rubin GPU at 50 PFLOPS of NVFP4 inference compute. Rubin is also a full platform change, pairing each Rubin GPU with a new Vera CPU where Blackwell racks pair with Grace.
NVIDIA hasn't published a direct, apples-to-apples PFLOPS comparison between Rubin and B200. What it has published is Rubin's own preliminary rating: 50 PFLOPS of NVFP4 inference compute per GPU and 35 PFLOPS of dense NVFP4 training compute per GPU. None of that is independently benchmarked yet, since Vera Rubin NVL72 has not shipped to any cloud customer, so treat it as NVIDIA's own preliminary spec rather than a measured result.
UALink is an open scale-up interconnect standard backed by a consortium that includes AMD, Broadcom, Cisco, Google, HPE, Intel, Meta, and Microsoft, and it is what ties the 72 MI455X GPUs inside a Helios rack together. NVLink is NVIDIA's proprietary interconnect, now in its sixth generation at 3.6TB/s per GPU inside Vera Rubin NVL72. The practical difference is who can build switches for it: UALink can run over standard Ethernet-class physical layers, while NVLink switch hardware comes from NVIDIA alone.
No, not by the hour. Both are newly announced rack-scale systems without general cloud availability. On Spheron's own pricing page, Vera Rubin is listed as pre-order only with shipment expected in H2 2026, and MI455X or Helios do not appear on the pricing page at all. The closest available alternatives today are NVIDIA B200 and B300 for Blackwell-generation capacity, and AMD MI300X for an on-demand ROCm path.
