On 30 September 2026, Hewlett Packard Enterprise said Vultr had placed a $1.2 billion order for AMD Helios AI Rack systems built by HPE. Vultr will deploy them at its cloud data center locations in the United States. HPE called it its first order for the new system.
That makes it the first public, dollar-denominated commitment to Helios through HPE. It is also the first big commercial test of an idea AMD, HPE and Broadcom have been selling for most of a year: that standards-based Ethernet can do the GPU-to-GPU job inside a rack that Nvidia does with its own NVLink fabric.
Cowry Tech News is covering this as a cloud and GPU infrastructure story. Our readers buy compute, sell it, or finance the people who do. For them the useful question is what this order says about where rack-scale AI capacity is heading, and what it does not say yet.
What HPE announced
The facts in HPE’s release are short.
- Buyer: Vultr, which HPE describes as “the world’s largest privately-held cloud infrastructure company.”
- Order value: $1.2 billion.
- Product: AMD Helios AI Rack by HPE, part of HPE’s AI Data Center Solutions portfolio.
- Location: Vultr cloud data center locations in the United States.
- Workloads named: model training and inferencing for service-provider and enterprise customers.
- Status: HPE’s first order for the system.
HPE disclosed the order the morning of its HPE Networking Investor Day in Sunnyvale. The same language appears in Exhibit 99.1 of a Form 8-K that HPE furnished to the SEC under Item 7.01, Regulation FD. That is the routine path for investor-day material.
HPE CEO Antonio Neri framed it as an architecture win. “AMD Helios by HPE demonstrates HPE’s unique ability to engineer the complete AI architecture with rack-scale compute, scale-up networking, and direct liquid cooling, all delivered by our global services,” he said in the release.
Vultr CEO J.J. Kardwell framed it as a supply problem. “Demand for high-performance AI infrastructure continues to outpace available capacity,” he said, adding that the work with AMD and HPE lets Vultr “bring Helios rack-scale compute and scale-up networking online for customers faster.”
What is inside one Helios rack
HPE published the bill of materials for a single rack in detail.
- 72 AMD Instinct MI455X GPUs.
- AMD EPYC “Venice” CPUs.
- AMD Pensando Vulcano AI NICs.
- AMD ROCm software.
- Six HPE Juniper Networking QFX5252 scale-up Ethernet switch trays per rack, connecting all 72 GPUs.
- Direct liquid cooling and HPE deployment services.
AMD designed Helios, in the release’s words, as “a high-density system for trillion-parameter model training and high-volume inference.” The rack supports UALink over Ethernet, which the industry shortens to UALoE.
Fierce Network reported that the QFX5252 switch was developed with Broadcom on its Tomahawk 6 chip. SDxCentral noted that HPE first teamed with AMD and Broadcom on Helios networking in December 2025.
A quick definition helps here. Scale-up networking ties GPUs inside one rack together tightly enough that they can share memory and act like one very large accelerator. Scale-out networking connects racks to each other across a data center. Nvidia’s NVL72 racks do scale-up with NVLink, a proprietary interconnect. Helios does it with Ethernet switches that support an open standard. Vultr’s $1.2 billion is the first large check written on that approach through HPE.
What the disclosure leaves out
HPE gave a complete spec for one rack and a total order value. It did not give the number that connects them. Neither HPE’s release nor its 8-K exhibit states:
- how many racks Vultr ordered,
- how many GPUs that adds up to,
- when the racks ship or go live,
- what a rack costs,
- how much power the deployment draws,
- which Vultr sites get them.
Fierce Network reported that neither HPE nor Vultr disclosed a rack count or a delivery schedule. It also reported that HPE did not say how much of the $1.2 billion covers networking.
Without a rack count, $1.2 billion cannot be turned into compute. Without a per-rack price, it cannot be turned back into a unit cost. CTN is not going to back into either figure. Any story that prints a GPU total for this deal is estimating, and should say so.
So the accurate frame is a purchase commitment. Vultr has agreed to buy. It has not announced Helios capacity that customers can rent today.
When it might actually ship
The timing comes from outside the release.
Fierce Network reported that AMD CEO Lisa Su said on AMD’s August earnings call that initial Helios shipments would begin later in the third quarter, step up in the fourth quarter and keep ramping through 2027. HPE told investors at the same event that its fiscal 2027 networking revenue would be more back-end loaded than usual, partly because of when Helios systems ship.
Read together, that points to a 2027 ramp for meaningful volume. It does not give a Vultr go-live date. Buyers planning around Helios capacity on Vultr should ask Vultr for region-level availability in writing, not rely on the order headline.
Why Vultr picked AMD again
Vultr has history with AMD. Per Fierce Network, Vultr was among the first cloud providers to deploy AMD’s Instinct MI325X and MI355X GPUs. AMD Ventures co-led Vultr’s December 2024 funding round, which valued the company at $3.5 billion. A year later, Vultr said it would spend more than $1 billion on a 24,000-GPU MI355X cluster in Springfield, Ohio, and would move to Helios next.
AMD is a Vultr shareholder, and Vultr is now a lead Helios buyer. The order is still real money. Readers who treat it as a demand signal for the wider market should factor in that HPE’s first Helios customer is a company AMD has a financial stake in.
Vultr is also not betting on one GPU vendor. Fierce Network reported that in June, Vultr chose HPE-built Nvidia GB300 NVL72 racks connected with Nvidia Spectrum-X networking. With Helios, HPE now supplies Vultr with rack-scale systems on both Nvidia and AMD GPUs.
Neocloud is the industry’s term for GPU-focused cloud providers that sit outside the big three hyperscalers. For one that sells to enterprises as an alternative to the hyperscalers, that dual-vendor position is part of the pitch. SDxCentral noted that Gartner placed Vultr in its “challengers” quadrant for cloud AI infrastructure, citing cost-optimized offerings and access to both AMD and Nvidia hardware, along with concerns about support response times on urgent, complex issues.
Why HPE wanted this order
For HPE, the deal is about Juniper as much as GPUs.
HPE closed its $14 billion Juniper Networks acquisition in July 2025. The release says the Vultr order builds on nearly three years of collaboration between Vultr and Juniper. SDxCentral reported earlier work in late 2024 that put AMD, Broadcom and Juniper equipment into Vultr’s Chicago cloud region.
Rami Rahim, who runs HPE Networking, made the pitch directly at the investor event. “It’d be very difficult for Juniper to do that on a standalone basis,” he said, according to SDxCentral, pointing to HPE’s liquid cooling and full-rack engineering.
He also put a number on the networking side. “We believe Helios represents more than a billion-dollar networking opportunity over the next two years, with networking tray orders having already exceeded $200 million,” Rahim said, per SDxCentral.
Fierce Network reported that HPE raised its fiscal 2027 networking revenue growth outlook to a range of high teens to low 20s percent, up from the 14 to 17 percent it forecast on 2 September. Rahim named Helios as the one new factor since that call.
Rahim also explained why full racks sell to neoclouds. “Many neoclouds value the simplicity of buying a complete system and buying from a single technology provider,” he said. Hyperscalers tend to buy compute, networking and storage separately and run their own integration. Smaller clouds often would rather buy the rack.
Ethernet versus NVLink, in plain terms
At bottom, this order is a choice of interconnect.
Nvidia’s rack-scale systems use NVLink for scale-up. It is fast, proven and Nvidia’s. Helios uses Ethernet switches inside the rack and supports UALink, an open interconnect standard backed by AMD and others. Rahim described it at the event as “taking standards-based Ethernet to a new layer of the data center stack for the first time,” per Fierce Network.
The appeal for buyers is choice. An open fabric could, over time, let a cloud mix vendors and avoid a single supplier for GPUs and the network that ties them together. The risk is maturity. NVLink racks are already in production at scale. Helios-by-HPE now has its first order. Performance, reliability and software readiness at full rack scale will be measured once Vultr’s racks are running customer jobs. The release offers no data on any of them yet.
For ROCm, AMD’s software stack, the stakes are similar. Large training jobs only move to new hardware when frameworks, kernels and libraries keep up. A billion-dollar buyer gives that ecosystem more reason to invest. Whether ROCm keeps pace with CUDA on large multi-rack jobs is still an open question for buyers to test.
What buyers should ask Vultr and HPE
If you rent GPU capacity, sell into neoclouds, or underwrite them, this order raises practical questions. None of them are answered in the release.
- Availability: Which US regions get Helios, and when can customers reserve or rent it? Get dates by region.
- Pricing: How will Helios instances be priced against MI355X and against Vultr’s Nvidia GB300 NVL72 capacity?
- Software: Which ROCm version, which frameworks, and which reference models are validated for multi-rack training on Helios?
- Scale-up behavior: How does UALoE over the QFX5252 trays perform on your communication pattern? Ask for benchmarks on your model size, not vendor peak numbers.
- Support: Given Gartner’s note on support response, what are the SLAs and escalation paths for rack-level faults?
- Liquid cooling: Who owns facility readiness and failures in the direct liquid cooling loop: Vultr, HPE Services, or the colocation operator?
- Portability: If a workload starts on Nvidia racks, what does it take to move it to Helios, and back?
How this fits with the rest of the GPU buildout
This is the third rack-scale order story CTN has tracked in a week. On 1 October we covered CoreWeave moving Nvidia Vera Rubin NVL72 into production. On 5 October, our Morning Brief carried AM Intelligence’s two new orders for 20,000 Nvidia Rubin GPUs for India and Malaysia, with delivery slated for the second quarter of 2027.
The pattern is consistent. Specialist clouds are committing to rack-scale systems a year or more ahead of delivery, and the headline is usually a dollar figure or a GPU count, rarely both, and almost never a go-live date. Vultr’s order is the AMD-and-Ethernet version of that pattern. What it tells you today is which architecture and which rack integrator Vultr is backing. It tells you little about capacity coming online this quarter.
What this piece is not
Vultr has not announced Helios instances for rent, so this piece does not describe capacity available today. CTN does not estimate rack counts, GPU totals or unit prices the companies did not disclose. Nothing here is investment advice on HPE, AMD or Vultr.
Disclosure: Lone Cowry has no commercial relationship with HPE, AMD, Broadcom or Vultr.
Sources
- HPE, “HPE Secures Its First AMD Helios Order in $1.2 Billion Deal With Vultr,” Business Wire, 30 September 2026: https://www.businesswire.com/news/home/20260930335555/en/
- HPE Form 8-K, Exhibit 99.1 (HPE Networking Investor Day), SEC EDGAR, 30 September 2026.
- Mitch Wagner, “HPE lands first AMD Helios order in $1.2B Vultr deal,” Fierce Network, 1 October 2026: https://www.fierce-network.com/broadband/hpe-lands-first-amd-helios-order-12b-vultr-deal
- Dan Meyer, “HPE Networking chief links $1.2B Vultr neocloud win to ‘revenue synergy’ opportunity,” SDxCentral, 1 October 2026: https://www.sdxcentral.com/news/hpe-networking-chief-links-12b-vultr-neocloud-win-to-revenue-synergy-opportunity/