Vultr Brings AMD's 72-GPU Helios Rack to the Cloud — Specs, Timeline, and Why It Matters

Per-Rack Specifications: 72 GPUs, 31 TB HBM4, 1.67 PB/s Bandwidth

The AMD Helios rack deployed on Vultr delivers a dense, integrated AI compute platform with the following specifications :

• 72 AMD Instinct MI455X GPUs based on next-generation CDNA architecture
• Up to 31 TB of HBM4 memory per rack (each MI455X carries 432 GB of HBM4)
• Up to 1.67 PB/s memory bandwidth per rack
• 2.9 exaFLOPS (FP4) inference and 1.4 exaFLOPS (FP8) training per rack — all AMD-reported figures not yet independently benchmarked
• 18 AMD EPYC CPUs (sixth-generation "Venice" processors)
• AMD Pensando networking designed for lossless, high-throughput AI clusters
• Scale-up fabric: 260 TB/s UALink over Ethernet
• Scale-out fabric: 43 TB/s UEC-aligned Ethernet (open standards-based)
• ROCm 7 software stack
• Direct liquid cooling (DLC) options

The rack is a 7,000-pound double-wide unit drawing up to 245 kW .

Integration of EPYC 'Venice' CPUs and Pensando Networking

The Helios rack integrates 18 sixth-generation AMD EPYC "Venice" CPUs — one per four GPUs across 18 compute trays — linked via AMD Pensando networking hardware for lossless, high-throughput AI cluster communication . This co-optimized silicon approach, combining GPUs, CPUs, networking, and software into a single integrated system, is AMD's first true rack-scale AI platform and its most direct answer to Nvidia's Vera Rubin NVL72 .

Deployment Timeline: Pre-Orders in Q4 2026, Production from 2027 to 2028

• Pre-orders for reserved capacity opened in Q4 2026
• The MI455X GPU is due in Q4 2026
• Helios deployments on Vultr are planned from 2027 through 2028

Engineering samples and limited-volume Helios production ship in the second half of 2026, with full mass-production ramp and first production AI tokens not expected until Q2 2027, per industry timelines .

Deployment Options: Bare Metal, Virtualized Instances, and Kubernetes

Vultr offers three deployment models for the Helios platform :

• Bare metal — direct access to physical GPU servers
• Virtualized instances — composable cloud infrastructure via the Vultr Console and API
• Kubernetes — via the free Vultr Kubernetes Engine (VKE) control plane for repeatable, cloud-native AI deployment and scaling

Customers can also use Vultr Clusters (self-service) with a Slurm or Kubernetes workload scheduler and an observability suite for GPU health monitoring . Vultr is positioning the new systems for model training, inference, fine-tuning, and agentic AI as businesses move from pilot projects to production .

Broader Cloud Momentum for AMD Helios

Microsoft Azure

On July 20, 2026 — three days before Vultr's announcement — Microsoft announced it will deploy the AMD Helios rackscale solution at scale on Azure for frontier-model AI inference, customer applications, and Microsoft's own AI services . Volume deployments are planned for the second half of 2026 . The expanded partnership covers Instinct GPUs, EPYC CPUs, Pensando networking, and ROCm software . Microsoft's official blog noted three new Azure VM families: HDv2, HXv2, and ND MI455X v7 .

Other Named Customers

• Anthropic announced a strategic partnership to deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs in Helios rack-scale systems
• Meta and OpenAI are also named as Helios customers
• TensorWave will operate an all-AMD AI cloud built on Helios racks
• Alignment Engine plans Helios deployments at an Ohio data center
• Schneider Electric co-developed a validated reference design for Helios deployment

HPE

The available evidence does not directly confirm a December 2025 HPE agreement or its 2026 availability timeline. No specific source for that claim was found.

Competitive Context: AMD's Rack-Scale Strategy vs. Nvidia

AMD Helios directly competes with Nvidia's Vera Rubin NVL72 . Key differentiating elements:

• Memory leadership: 31 TB HBM4 per rack (72 GPUs × 432 GB each) versus Nvidia's Rubin at roughly 20 TB — a ~50% capacity advantage per GPU . Per-rack memory bandwidth is roughly 1.67 PB/s .
• Paper specs vs. reality: The Register notes that Helios is bigger and faster than Vera Rubin "on paper" but that AMD has yet to prove performance in production deployments .
• UALink over Ethernet (260 TB/s scale-up): AMD's open-standard alternative to Nvidia's proprietary NVLink. UALink is built as an Ethernet-based, open industry standard .
• UEC-aligned Ethernet (43 TB/s scale-out): Helios uses Ultra Ethernet Consortium (UEC)-aligned Ethernet for open scale-out networking, avoiding proprietary InfiniBand .
• Open standards positioning: AMD is positioning Helios as a more open, standards-based alternative to Nvidia's vertically integrated rack-scale systems, aiming to appeal to hyperscalers and cloud providers seeking vendor diversity .
• Training gap acknowledged: Multiple reports note that while Helios leads in memory and inference specs, its training performance trails Nvidia's offerings, and all figures are vendor-reported and not yet independently benchmarked .
• Pricing: The Helios rack is listed at approximately $5.25 million per unit .