A $1.2 billion order makes AMD Helios a concrete commercial undertaking for HPE, but not yet installed cloud capacity. HPE announced on September 30, 2026 that Vultr would deploy its integrated Helios systems in United States data centers. It is HPE’s first order for that platform, a narrower and more useful claim than the first customer for every Helios implementation.
Vultr is buying a complete alternative rack-scale computing environment: accelerators, CPUs, network fabric, cooling, software and deployment support. The purchase will test whether several suppliers can deliver an integrated competitor to Nvidia without one company controlling the whole stack. A large contract establishes demand; production service will establish the result.
What sits inside the ordered system
Each rack has 72 AMD Instinct MI455X GPUs, EPYC host processors code-named Venice and Pensando Vulcano AI network interfaces. ROCm supplies the GPU programming environment. HPE provides rack engineering, direct liquid cooling, services and six Juniper QFX5252 scale-up switch trays.
Scale-up connects accelerators inside one computing domain so they can work on a shared model. This is different from merely connecting independent servers. HPE’s implementation uses UALink over Ethernet, or UALoE, carrying the traffic through an Ethernet-based fabric. Broadcom contributes switch technology in that design.
The distinction matters for large distributed jobs. Accelerators exchange intermediate results, and waiting for collective communication can idle otherwise fast chips. HPE says its six trays provide high-bandwidth, low-latency links among the GPUs. Real behavior will depend on communication patterns, placement and software, not the presence of an Ethernet label.
AMD lists 31 terabytes of rack-level HBM4, 260 terabytes per second of aggregate scale-up bandwidth, 2.9 exaflops peak FP4 and 1.4 exaflops FP8 computation. HBM places fast memory close to the processors; the low-precision formats increase potential throughput when model quality remains acceptable. These are vendor specifications, not measured application outcomes.
The order does not disclose rack quantities, a complete delivery calendar or the allocation among compute, networking, cooling, software and services. Dividing $1.2 billion by 72 therefore cannot produce a valid per-GPU price. HPE also included the announcement in a September 30 SEC disclosure, but formal disclosure does not fill those commercial gaps.
Juniper turns this into HPE’s integration test
HPE completed its Juniper acquisition in 2025. Helios gives the combined organization an opportunity to sell networking inside an accelerator domain, not only connect ordinary servers outside it. The six switch trays are central to the rack rather than optional background equipment.
At its investor event, HPE described more than $1 billion of Helios-related networking opportunity over the following two years and networking tray orders exceeding $200 million. An opportunity forecast, booked orders and equipment operating for customers are different stages. The company did not identify how much networking revenue belongs to Vultr’s contract.
That explains why the deal matters beyond AMD’s GPU share. HPE wants to capture integration work across power, cooling, firmware, telemetry, installation and support. A customer can own excellent processors yet lose usable capacity to unstable software, congestion or slow fault diagnosis.
Direct liquid cooling supports higher density but creates its own obligations. Facilities need compatible fluid distribution, heat rejection, leak detection and service procedures. Technicians must replace parts safely, and spare inventories need to match the installed design. A repeatable rollout across several buildings requires more than one successful rack demonstration.
Vultr has previously adopted AMD Instinct systems, giving it relevant experience. HPE’s services may reduce installation and operating risk, but the agreement alone does not show that the new GPU, CPU, fabric and facility combination has achieved sustained reliability.
Open interfaces do not remove integration or migration
Nvidia’s comparison platform, Vera Rubin NVL72, also groups 72 GPUs into a rack-scale domain, with 36 Vera CPUs and sixth-generation NVLink. CUDA, networking, libraries and familiar operational tools reinforce its position beyond chip arithmetic.
The AMD–HPE proposal is not unintegrated hardware. AMD supplies compute and ROCm; HPE and Juniper supply the rack, switches, cooling and services, with Broadcom networking technology. OCP physical specifications, UALink and Ethernet provide interfaces that more suppliers can implement.
That offers a different distribution of control, not automatic interchangeability. An open interface does not guarantee that any replacement part works without qualification or that an existing CUDA pipeline becomes portable at no cost. Nor does openness prove better throughput.
ROCm supports major frameworks including PyTorch, TensorFlow and JAX. A production model may nevertheless depend on a CUDA-specific kernel, library or monitoring tool. Engineers can need new kernels, communication tuning, dependency replacement and repeated quality testing. Under a tight deadline, that effort can outweigh a lower hardware cost.
A cloud provider can absorb part of this work. Validated containers, recipes, supported configurations and experienced staff can make Helios accessible without every customer building a distributed-systems team. Customers still need to compare how much work remains for their model, not assume a framework checkbox establishes equivalent operation.
Capacity, contracts and utilization remain unresolved
AMD described Helios shipments ramping in late 2026 and 2027; HPE and Vultr did not provide a detailed rollout schedule for this order. The announcement should not be read as all racks already installed, reservable or exposed through a generally available cloud service.
MI455X needs advanced packaging and HBM4. Complete racks also need CPUs, interfaces, trays, power equipment, coolant infrastructure and prepared sites. One missing component can delay an integrated configuration. Dense systems also concentrate facility requirements that existing data centers may not quickly satisfy.
Contract details are similarly limited. Delivery milestones, cancellation provisions, minimum purchases and service-related shares were not disclosed. A purchase commitment, delivered hardware and a flexible multiyear arrangement distribute risks differently; the public value alone does not resolve which obligations apply.
Vultr says demand for high-performance AI infrastructure exceeds supply. That is its commercial assessment, not an independently audited forecast of Helios utilization. No disclosed customer-by-customer demand or future occupied capacity establishes how quickly this purchase earns a return.
Supporting AMD alongside Nvidia can diversify supply and strengthen negotiation. Larger memory may suit some models, while managed capacity can appeal to customers concerned about vendor concentration. Yet low utilization would leave expensive equipment consuming capital. Reserved customer commitments and repeat orders after operation would be stronger evidence than expectations.
Geography also matters. HPE identified United States locations, not a detailed region list. Customers need actual availability, proximity to data, residency options, disaster-recovery arrangements and service commitments before translating the order into usable capacity.
The benchmark is a complete workload in production
Training stresses collective synchronization; long-context inference and mixture-of-experts models stress memory, routing and scheduling differently. Peak FP4 and FP8 figures cannot settle both. Cross-vendor comparisons also require matching precision, sparsity assumptions, software and output quality.
Useful tests would report training completion, tokens per second, first-token and ongoing latency, model size, context, batching, power and tuning effort under the same constraints. Uptime, maintenance effects, fault recovery and scheduling determine how much nominal performance becomes work customers can depend on.
The next milestones are orderable capacity in named regions, independently reviewable workload results and follow-on adoption. Another Vultr expansion after production would be especially informative: it would connect operating experience with renewed demand. More HPE customers would test whether the architecture appeals beyond one established AMD user.
Nvidia can respond through faster Rubin delivery, software improvements, cloud partnerships and Ethernet products of its own. Helios need not displace it everywhere to matter. It needs to become a dependable choice where memory, availability, supplier diversity or open interfaces justify the transition.
The order creates a credible venue for that test. What changes the market will be a working service with measurable economics—not the contract headline or the largest number on a specification sheet.
Sources
HPE order announcement · HPE SEC disclosure · Networking analysis · Nvidia NVL72 specifications · AMD Helios specifications · MI455X technical reporting