AMD Helios has its first disclosed order through Hewlett Packard Enterprise: cloud provider Vultr has ordered $1.2 billion of HPE-built rack systems for its United States data centres. HPE announced the order on September 30, 2026. The news is less a final verdict on whose AI chips win than a commercial test of whether a full rack of accelerators, Ethernet switching, cooling and software can be delivered and operated as one dependable cloud product.
AMD Helios is Advanced Micro Devices’ rack-scale AI design. HPE is the systems and networking supplier for this order, while Vultr is the privately held cloud company planning to deploy the equipment at its US sites. Reuters, Bloomberg and Dow Jones Newswires each reported the order. Their accounts support the transaction and its basic structure; the detailed component specifications below come from the companies and should be read as product claims, not independently measured performance.
AMD Helios turns a chip plan into a named order
The distinction between a product roadmap, a customer selection and a purchase order matters. AMD and HPE described their Helios collaboration in December 2025. AMD later listed Vultr among companies planning Helios deployments in July 2026. The September 30 announcement adds a named dollar amount and identifies HPE as the supplier of the first Helios systems it has sold. That is a stronger commercial signal than a partner logo, but it is still an order, not evidence that all equipment has shipped or entered service.
HPE describes the order as spanning Vultr’s US cloud data-centre locations. It does not say how many racks Vultr bought or disclose prices per rack, deployment dates, purchase milestones or the mix between hardware, software and services. Dividing $1.2 billion by an assumed rack price would therefore manufacture a rack count. Even if one knew the equipment price, the announcement does not say whether the total includes networking, cooling integration, service contracts or later phases.
What does the HPE–Vultr order actually prove? It proves that HPE and Vultr have publicly announced a $1.2 billion Helios order and that HPE has its first named customer for its implementation of AMD’s rack design. It does not prove delivered GPUs, available cloud instances, profitable operation or a measured performance advantage over another AI system. Those tests require shipment updates, live service availability and customer workload evidence.
What is inside the AMD Helios rack?
According to HPE’s product description, a Helios rack combines 72 AMD Instinct MI455X GPUs with AMD EPYC “Venice” central processors, AMD Pensando “Vulcano” network interface cards and AMD ROCm software. HPE says its version adds six HPE Juniper Networking QFX5252 scale-up Ethernet switch trays per rack. It also points to direct liquid cooling and deployment services. AMD’s Helios product page separately describes the 72-GPU design and its standards-based network approach.
Those parts play different roles. GPUs carry out the parallel calculations used in model training and inference. CPUs coordinate the system and handle work unsuited to accelerators. Networking moves information among GPUs and between racks. Cooling removes heat from densely packed components. Software makes the hardware usable for developers. The entire rack must work together; a bottleneck in any one layer can undermine the economics of the rest.
“Scale-up” networking links accelerators within a tightly coupled system, whereas “scale-out” links clusters and sites. HPE presents its Ethernet switch trays and open UALink-over-Ethernet approach as an alternative to more closed rack fabrics. The word “open” describes the proposed standards and interoperability route. It does not by itself show that an application can switch suppliers without code changes, retraining or operational cost. HPE and AMD have not published an independent production benchmark for this specific Vultr installation.
The headline figure is the order value, not the rack specification. HPE’s statement lists the components but does not establish how many of the announced systems are currently in Vultr’s facilities. Readers should also avoid treating AMD’s advertised Helios throughput or energy-efficiency figures as results Vultr has achieved. Real-world performance depends on model size, software maturity, memory movement, network congestion, cooling conditions and utilisation.
| Item | What is disclosed | What remains undisclosed |
|---|---|---|
| Commercial commitment | $1.2 billion Vultr order, according to HPE | Payment and revenue-recognition schedule |
| Rack design | 72 AMD MI455X GPUs and six HPE Juniper scale-up switch trays per rack, according to HPE | Number of ordered or installed racks |
| Deployment geography | Vultr US cloud data centres, according to HPE | Specific locations and activation dates |
| Workloads | Intended for AI training and inference, according to HPE and Vultr | Live customer workloads or measured efficiency |
Vultr already had an HPE-and-Nvidia plan
A significant nuance is easy to miss in a “chip rivalry” headline. On June 17, HPE announced that Vultr had selected HPE and Nvidia systems for cloud-scale AI data centres. That earlier plan named Nvidia GB300 NVL72 by HPE equipment and Spectrum-X networking. The September AMD order therefore sits beside a public Nvidia selection. Neither announcement says the newer programme replaces the older one.
This multi-supplier picture is more useful to a cloud customer than a simple winner-and-loser narrative. A provider can offer different accelerator options for workloads with different software, memory, availability and cost requirements. It might also reduce reliance on one supply chain. But Vultr has not disclosed an exact allocation between AMD and Nvidia capacity, and customers have not yet shown which workloads will run on each platform. The practical difference will emerge in published instance availability, pricing and supported tools, not merely in procurement headlines.
It is also important to distinguish HPE’s role from AMD’s. AMD specifies the accelerator architecture and supplies core chips. HPE integrates rack-scale systems, Juniper networking and liquid-cooling expertise. Vultr operates cloud infrastructure and sells access to customers. This creates several separate execution risks: chip supply, rack assembly, facility readiness, software compatibility and actual demand from renters of AI compute. A signed order reduces commercial uncertainty for HPE; it does not remove those later gates.
Why the Juniper connection matters to HPE
HPE’s announcement emphasises that Vultr had worked with Juniper Networks for nearly three years. HPE acquired Juniper in 2025, as both Bloomberg and Dow Jones noted in their September 30 reports. The order is therefore a case study for HPE’s claim that selling compute and networking together can win large AI infrastructure projects.
That claim has a technical reason. A rack of accelerators needs extremely fast communication among chips; otherwise expensive compute can sit idle waiting for data. HPE says its six Juniper switch trays per Helios rack connect all 72 GPUs through standards-based Ethernet. If integration works at the promised scale, it could make HPE more than the assembler of AMD chips. It would make the network design, cooling and operational support part of the value the cloud buyer is purchasing.
HPE also raised its longer-term networking growth outlook at its investor event, Reuters reported. That forecast is management’s projection, not a realised result of the Vultr order. It should not be added to the $1.2 billion order or presented as another customer contract. The order could support the broader strategy, but investors and buyers still need to see actual delivery, margins and sustained service demand.
Lapaas Voice previously explained an HPE arrangement tied to AI networking sales with Oracle and HPE’s third-quarter AI infrastructure results. Those developments show why HPE wants large customers to buy integrated systems rather than isolated components. The Vultr order is a different event: it names a specific Helios platform, buyer and announced value.
What Indian cloud and AI teams should watch
There is no announced Indian Vultr deployment in this deal. HPE explicitly describes US data-centre locations, so it would be wrong to imply immediate India capacity or local pricing. The India relevance is the buying model: organisations evaluating large AI systems must judge a complete stack of accelerators, networking, software and power, not compare GPU specifications alone. India-based developers using international cloud platforms will care about whether Vultr makes the new capacity available in regions and service tiers they can actually use.
The second question is portability. Open rack interfaces may broaden hardware choices, but application portability depends on frameworks, kernels, model support, monitoring tools and the operational skill to tune them. Teams should request workload-specific performance and total-cost evidence before treating an “open” architecture as automatically cheaper. This is an inference from how AI infrastructure is operated, not a claim that HPE or Vultr has failed a test.
The third question is resilience. If Vultr offers both Nvidia and AMD rack platforms, customers could gain alternatives when supply or prices shift. But a multi-vendor fleet also introduces software and support complexity. Buyers need to know which models and libraries are supported on each platform, whether their data can move between them, what service-level commitments apply, and whether performance is comparable for their own workloads. HPE’s announcement answers none of those customer-level questions yet.
For a broader comparison of the platform’s earlier positioning, see Lapaas Voice’s AMD Helios explainer. That article covered the system’s general challenge to Nvidia. The new development is narrower and more concrete: HPE now says one cloud company has placed a $1.2 billion order for its version of the design.
How to measure whether the order succeeds
The first observable milestone would be shipment and installation dates, preferably with an identified site or service availability statement. A second would be the opening of Vultr cloud instances based on MI455X GPUs, including locations, pricing, memory configuration and software support. A third would be independent workload testing or customer reports. Neither HPE’s press release nor the initial wire accounts supply those outcomes. Until then, the commercial value is an announced order and a potential deployment pipeline.
One should also watch how HPE reports revenue from AI systems and networking over future quarters. The $1.2 billion is the announced total order, not automatically one quarter’s revenue. Recognition depends on delivery terms and accounting treatment that the release does not spell out. Vultr, likewise, will have to convert equipment into rentable compute and then win customers who use it sufficiently to justify the capital cost.
There are questions the release leaves open: How many racks are in the order? Which US sites receive them first? When will customers be able to rent the capacity? What share of the total price is networking or services? How much of the intended use is training versus inference? HPE, AMD and Vultr may answer these in later product, investor or customer updates. Filling the gaps today with estimates would make the story sound more complete than the evidence allows.
The confirmed news is still substantial. HPE has a named first customer for its AMD Helios rack, and the announced order is $1.2 billion. Its commercial significance rests on whether HPE can integrate AMD compute with Juniper networking and liquid cooling at cloud scale, and whether Vultr can turn those racks into usable AI services. That is the standard against which this deal should be judged.
Frequently asked questions
What did Vultr order from HPE?
According to HPE’s September 30 announcement, Vultr ordered $1.2 billion of AMD Helios AI Rack by HPE systems for its US cloud data centres. The disclosed product design includes AMD accelerators, HPE Juniper networking and HPE integration services. HPE did not publish the rack count or installation timetable.
Is this AMD Helios equipment already live?
Neither the HPE release nor the cited Reuters, Bloomberg and Dow Jones reports establish that this specific order has been delivered or made available to Vultr customers. An announced order and an operating cloud service are different stages.
Does the AMD order replace Vultr’s Nvidia systems?
There is no statement that it does. HPE separately announced Vultr’s selection of HPE-and-Nvidia AI infrastructure in June 2026. The September AMD order is best read as an additional announced architecture unless Vultr says otherwise.
Why is Ethernet networking central to the story?
Training and serving large AI models require rapid communication among GPUs. HPE says its Juniper switch trays provide standards-based scale-up connections within the Helios rack. That is part of its product design claim; production performance at Vultr remains to be demonstrated.
Source note: The commercial announcement and product specifications come from HPE’s September 30 release, compared with original reports from Reuters, Bloomberg and Dow Jones Newswires. AMD’s product and July partnership materials and HPE’s June Vultr–Nvidia announcement provide context. Company descriptions of performance, openness or future customer demand are attributed rather than treated as independently verified results.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
