AMD Helios (72x MI455X, 31 TB HBM4, ~$5-5.5M) vs Nvidia Vera Rubin NVL72 (3.6 EF FP4, ~$3.5-4M), compared spec by spec and dollar by dollar — with the supply-chain read for $AMD, $NVDA, $MU, $TSM.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
AMD Helios is AMD's ($AMD) first rack-scale AI system: 72 Instinct MI455X GPUs, 31 TB of HBM4 memory, 2.9 exaflops of FP4 compute, and a reported price of $5–5.5 million per rack, roughly 40% above Nvidia's ($NVDA) competing Vera Rubin NVL72. It launched July 20, 2026 with Microsoft ($MSFT), Meta ($META), OpenAI, Anthropic, and Oracle ($ORCL) committed to deployments.
For two product generations AMD sold accelerators against Nvidia's accelerators and kept losing the datacenter on everything around the chip: the interconnect, the networking, the software, the rack. Helios changes the unit of competition. Hyperscalers no longer buy GPUs and assemble systems; they buy a liquid-cooled rack that behaves like one giant GPU. Nvidia has owned that category since the GB200 NVL72. Helios is the first credible second source, and the two machines disagree about what an AI rack is for.
👉 The memory angle first: a single Helios rack carries 31 TB of HBM4. At the gigawatt scale AMD's customers have committed to, that is one of the largest HBM order books ever placed — track the suppliers in the Memory Supercycle basket: $MU, SK Hynix ($000660.KS), Samsung ($005930.KS).
The launch came in two acts. On July 20, AMD unveiled Helios with Microsoft as the first named customer — Azure will deploy Helios racks "at scale" for frontier-model inference. Two days later, at its Advancing AI keynote, AMD added a deal with Anthropic: up to 2 gigawatts of MI450-series GPUs in Helios racks, the first gigawatt arriving in the first half of 2027, alongside an AMD equity investment in Anthropic of up to $5 billion.
Stack those on the commitments already public and the demand book looks like this:
| Customer | Commitment | First deployment |
|---|---|---|
| OpenAI | 6 GW of Instinct GPUs (with warrants for up to 160M AMD shares) | 1 GW of MI450, H2 2026 |
| Meta ($META) | Up to 6 GW, first gigawatt on Helios racks | H2 2026 |
| Anthropic | Up to 2 GW in Helios racks + up to $5B AMD investment | H1 2027 |
| Microsoft ($MSFT) | Azure Helios deployment "at scale" (size undisclosed) | — |
| Oracle ($ORCL) | 50,000 MI450-series GPUs (announced Oct 2025) | Q3 2026 |
Call it 14+ gigawatts of committed AMD accelerator demand. Whatever you think of the specs below, the fight is funded.
A naming note before the table: Nvidia originally called its Rubin rack "NVL144", counting the 144 reticle-sized compute dies. In December 2025 it renamed the same machine VR200 NVL72, counting the 72 GPU packages. Same rack, either name in coverage.
| AMD Helios | Nvidia Vera Rubin NVL72 | |
|---|---|---|
| GPUs | 72× Instinct MI455X (CDNA 5) | 72× Rubin packages (144 dies) |
| CPUs | 18× EPYC "Venice" (Zen 6) | 36× Vera (88 custom Arm cores each) |
| FP4 compute | 2.9 EF | 3.6 EF |
| FP8 compute | 1.4 EF | 1.2 EF |
| HBM4 per GPU | 432 GB | 288 GB |
| HBM4 per rack | 31 TB | 20.7 TB |
| HBM bandwidth per GPU | 19.6 TB/s | up to 22 TB/s |
| Total fast memory | 31 TB HBM4 (+ Venice DDR5) | ~75 TB (HBM4 + 54 TB CPU LPDDR5X) |
| Scale-up fabric | UALink over Ethernet, 260 TB/s | NVLink 6, 260 TB/s (3.6 TB/s per GPU) |
| Scale-out | 43 TB/s, Ultra Ethernet via Pensando NICs | ~28.8 TB/s, ConnectX-9 (1.6 Tb/s per GPU) |
| Rack | OCP Open Rack Wide (double-wide), ~7,000 lbs | Oberon-derived, NVL72 form factor |
| Power (reported) | ~225–245 kW | ~190–230 kW |
| Reported price | $5–5.5M | $3.5–4M |
| Volume shipments | First racks by end of 2026, volume Q2 2027 | Late 2026 (samples already at customers) |
Both FP4 numbers are vendor peak figures; sparsity and datatype conventions differ, so treat the ratio, not the absolutes, as the signal.
Where each machine wins:
Divide the reported mid-range prices ($5.25M Helios, $3.75M Rubin) by what each rack contains and the strategies separate cleanly:
| Per-unit cost (mid estimates) | AMD Helios | Vera Rubin NVL72 | Cheaper |
|---|---|---|---|
| Per PF of FP4 | ~$1,810 | ~$1,040 | Nvidia, by ~42% |
| Per PF of FP8 | ~$3,750 | ~$3,125 | Nvidia, by ~17% |
| Per GB of HBM4 | ~$169 | ~$181 | AMD, by ~7% |
| Per TB/s of scale-out | ~$122k | ~$130k | AMD, by ~6% |
Nvidia sells FLOPs cheaper. AMD sells bytes cheaper — despite charging 40% more per rack. That is the entire strategic argument in one table.
It matters because the two dominant AI workloads bind on different resources. Pre-training is FLOPs-bound, and Rubin is the better training machine per dollar. But inference on long-context and agentic models is memory-bound: the KV cache for a frontier model serving million-token contexts eats HBM capacity, and decode throughput tracks memory bandwidth and capacity, not peak FLOPs. We made the general version of this argument in the Rubin roadmap post — decode runs at under 20% of peak FLOPs on current hardware; bytes are the binding constraint. Helios is a bet that the buyers of 2027 are provisioning for exactly that workload. It's telling that Microsoft's stated use case for Helios on Azure is frontier-model inference.
The components decide which suppliers get paid no matter who wins the rack fight.
HBM4 is the common denominator. 31 TB per Helios rack, 20.7 TB per Rubin rack, and every gigawatt of either is tens of thousands of HBM4 stacks from exactly three companies: Micron ($MU), SK Hynix ($000660.KS), and Samsung ($005930.KS). AMD's memory-heavy configuration makes it the more HBM-intensive design per rack — good news for a memory industry already sold out through 2026. Background: our Micron deep dive and the High Bandwidth Memory trend.
TSMC ($TSM) prints both sides. MI455X and Rubin are both 3nm-class chiplet designs on CoWoS-style advanced packaging — the same Advanced Semiconductor Packaging bottleneck, the same HBM / Packaging basket of bonders, testers, and OSATs behind it.
Networking is where the philosophies split. Nvidia keeps the fabric in-house: NVLink 6 switches, ConnectX-9 NICs, BlueField-4 DPUs, Spectrum-X co-packaged-optics Ethernet — interconnect dollars stay inside $NVDA (the CPO piece flows to the Photonics / CPO basket). AMD routes everything through open Ethernet: UALink tunneled over Ethernet for scale-up, Ultra Ethernet through its own Pensando "Vulcano" 800G NICs for scale-out. Every Helios rack is a design win for the merchant Ethernet ecosystem — switch silicon, retimers, cabling — tracked in the Networking / Retimers basket: $AVGO, $MRVL, $ALAB, $CRDO.
Power and cooling don't care who wins. Both racks are liquid-cooled, 800 VDC, and pull 200–250 kW — five to six times a Hopper-era rack. That is the demand driver behind the 800 VDC Architecture basket ($VICR, Delta $2308.TW, $MPWR) and the Cooling basket ($VRT, $MOD, $NVT). Supermicro ($SMCI) is already building Helios-based systems; the Taiwanese ODM complex (Foxconn $2317.TW, Quanta $2382.TW) assembles Nvidia's racks.
AMD's pitch to hyperscalers isn't only specs — it's structure. Helios's scale-up fabric is UALink (a consortium standard) running over Ethernet, its scale-out is Ultra Ethernet, and its double-wide Open Rack Wide chassis was contributed to the Open Compute Project. A hyperscaler that deploys Helios keeps its switch vendors, its rack tooling, and its exit options. A hyperscaler that deploys NVL72 buys Nvidia's fabric, Nvidia's NICs, and Nvidia's rack, and its next rack is whatever Nvidia ships next.
That is precisely why Microsoft framed its announcement as a full-stack partnership across EPYC, Pensando, and Helios, and why every large buyer has now committed capital to a second source. The counterargument is equally simple: consortium standards move at consortium speed, Nvidia's vertical stack demonstrably works today, and the moat was never really the fabric — it's CUDA. ROCm has closed much of the training gap on paper, but "on paper" is the operative phrase until a gigawatt of Helios runs production frontier workloads.
For AMD: the shipping gap. Rubin volume lands in late 2026; Helios volume lands Q2 2027. Nvidia gets one to two quarters of uncontested Rubin deployments, and then Rubin Ultra NVL576 (576 dies, 600 kW Kyber racks, a claimed 15 EF of FP4) arrives in H2 2027, right on top of Helios's ramp. AMD's 432 GB memory lead over Rubin evaporates against Rubin Ultra's 1 TB-per-package target. Add ROCm execution risk and the physical reality that a 7,000-lb double-wide rack forces datacenter floor-plan changes, and the premium price has to earn itself fast.
For Nvidia: the customer base is now funding its own leverage. OpenAI took warrants in AMD, AMD is investing in Anthropic, Meta and Microsoft are deploying both platforms. If Helios inference economics hold up in production, the 75%-gross-margin pricing umbrella that makes Nvidia a $5 trillion company is exactly what a memory-cheaper, Ethernet-native competitor attacks. And the reported fact that customers accepted a 40% Helios price premium suggests the scarcity value of a second source is real.
What is AMD Helios? AMD Helios is a rack-scale AI system that integrates 72 Instinct MI455X GPUs, 18 EPYC "Venice" CPUs, and Pensando networking into a single liquid-cooled, double-wide OCP rack delivering 2.9 exaflops of FP4 compute and 31 TB of HBM4 memory. It is AMD's first product sold as a complete rack rather than as chips, and its first direct competitor to Nvidia's NVL72-class systems.
How much does an AMD Helios rack cost? Reported estimates put Helios at $5–5.5 million per rack, roughly 40% above the $3.5–4 million estimated for Nvidia's Vera Rubin NVL72. Neither company publishes list prices; actual hyperscaler pricing is negotiated and lower.
Is AMD Helios faster than Nvidia Vera Rubin? Depends on the axis. Rubin has ~24% more FP4 inference compute (3.6 vs 2.9 EF) and higher memory bandwidth per GPU. Helios has ~17% more FP8 training compute (1.4 vs 1.2 EF), 50% more HBM4 capacity (31 vs 20.7 TB per rack), and ~50% more scale-out bandwidth. For memory-bound inference workloads, capacity per dollar favors Helios; for FLOPs-bound work, Rubin.
Who is buying AMD Helios? Microsoft (Azure), Meta (first gigawatt of a 6 GW commitment), OpenAI (part of a 6 GW deal), Anthropic (up to 2 GW), Oracle (50,000 MI450-series GPUs), and Tata Consultancy Services ($TCS.NS) have all announced commitments.
When does Helios ship? First racks ship in the second half of 2026, with volume production in Q2 2027. Nvidia's Vera Rubin NVL72 reaches volume in late 2026, giving Nvidia roughly a one-to-two-quarter head start.
Which stocks benefit regardless of who wins? The shared bill of materials: HBM4 suppliers ($MU, SK Hynix, Samsung), TSMC ($TSM) for silicon and packaging, and the power and cooling chain behind 200–250 kW racks ($VRT, $VICR, Delta). This article is research and education, not financial advice — do your own diligence.