Will Open Ethernet Challenge NVIDIA’s Grip on AI Data Centers?

Key Highlights

• Vultr has placed a $1.2 billion order for HPE systems built on the AMD Helios AI Rack architecture for its US cloud data centers.
• The deal is HPE's first order for AMD Helios, and Vultr has said its Helios orders are planned across 2027 and 2028.
• Each rack packages 72 AMD Instinct MI455X GPUs and EPYC "Venice" CPUs on AMD's UALink over Ethernet (UALoE) scale-up fabric. Pensando Vulcano AI NICs handle scale-out networking between racks.
• Our read is that the Ethernet scale-up fabric is both the least proven layer in the rack and the one where HPE has the most at stake.
• CUDA's software depth, more than any hardware specification, remains the main obstacle between this order and broader enterprise adoption.

The News

Hewlett Packard Enterprise has secured a $1.2 billion order from Vultr for AMD Helios AI Rack systems across its US cloud data center footprint. It is HPE's first order for the Helios system, which combines HPE Juniper scale-up Ethernet switching with AMD accelerators and HPE direct liquid cooling. HPE announced the order alongside its networking investor day, positioning it against capacity shortages for enterprise model training and high-volume inference. Readers can read the press release here.

Analyst Take

The $1.2 billion order from Vultr for HPE systems based on the AMD Helios architecture marks a notable step for open-standard AI infrastructure. Our read is that cloud providers outside the hyperscaler tier have been looking for credible second sources, because NVIDIA's allocation and pricing determine who gets capacity and when. Vultr is using this deal to build out capacity across its US footprint, and the relationship goes beyond a purchase order. AMD Ventures co-led Vultr's December 2024 funding round, so the buyer and the silicon supplier both have a stake in this order succeeding. For HPE, the deal is the first Helios order and a proof point for attaching Juniper networking to AI compute. It landed the same day HPE raised its networking revenue outlook to a high-teens CAGR through fiscal 2029. The easy read is that the GPUs are the story. We think the network is. The bull case holds that pairing Juniper's scale-up switching with AMD accelerators and HPE direct liquid cooling gives HPE a full-stack answer to proprietary end-to-end systems, and that case has merit. Yet the fabric is the least field-tested component in the rack, and it is the layer where HPE has staked the most differentiation. If the fabric underperforms, the GPU story stalls with it.

What Was Announced

The announcement centers on the AMD Helios AI Rack by HPE, a high-density integrated system architected for trillion-parameter model training and high-volume inference. Each rack integrates 72 AMD Instinct MI455X GPUs alongside AMD EPYC "Venice" CPUs and AMD ROCm software. AMD Pensando Vulcano AI network interface cards carry scale-out traffic, which runs separately from the GPU-to-GPU scale-up fabric. That scale-up fabric is UALink over Ethernet (UALoE), AMD's approach to carrying load/store GPU traffic over standard Ethernet. AMD's own engineering descriptions suggest UALoE bridges the GPUs' Infinity Fabric across Ethernet, and that its main departure from UALink is the switch: merchant Ethernet silicon in place of a dedicated UALink switch. Our read is that this makes UALoE closer to an AMD protocol running on Ethernet than to native UALink, and the distinction matters for buyers who are counting on multi-vendor interoperability. The rack uses six scale-up Ethernet switch trays, which HPE's release designates as HPE Juniper Networking QFX5252.

These switch trays aim to connect all 72 GPUs over high-bandwidth, low-latency Ethernet. HPE Services will manage end-to-end deployment, applying HPE direct liquid cooling (drawn from the same technology family as its exascale systems) along with operational management to mitigate thermal and project risks across Vultr's locations.

The Juniper switching inside the Helios rack deserves attention. Interconnect bottlenecks have historically been a primary hurdle keeping open systems from matching the efficiency of proprietary scale-up fabrics. With six Ethernet switch trays per rack, HPE aims to show that merchant Ethernet can carry the bandwidth demands of 72-GPU domains without a proprietary switch fabric. Ethernet brings familiar management tooling and broader architectural flexibility. Its cost advantage over proprietary fabrics is widely assumed, though at scale-up densities that assumption has yet to be tested publicly. HPE also has company here. Broadcom's Tomahawk Ultra targets the same scale-up Ethernet role, Arista sells into AI back-end fabrics, and NVIDIA itself ships Ethernet through Spectrum-X, which weakens any framing of this deal as Ethernet versus NVIDIA. The genuinely open scale-up effort is ESUN, the Open Compute Project workstream whose members include AMD, Arista, Broadcom, HPE, and NVIDIA. How closely UALoE converges with that work may decide how open it ultimately proves to be. Juniper gives HPE ownership of the switch, not a monopoly on the idea. If Vultr can demonstrate that UALoE delivers comparable low-latency scaling for large training jobs, it may provide a blueprint for other cloud providers seeking to limit vendor lock-in.

Software readiness remains the central variable. Hardware specifications look formidable on paper, but software stacks determine day-to-day cluster utilization. AMD's ROCm platform has advanced considerably over recent releases. Even so, trillion-parameter workloads spread across the multi-rack clusters they require test memory management and compiler efficiency to the limit. This is where NVIDIA's advantage is hardest to dislodge. CUDA's moat is less the compiler than a decade of tuned libraries and kernels, and the engineers who learned on it, and enterprise teams rarely budget for retraining. Vultr's enterprise customers will likely expect PyTorch and TensorFlow pipelines to port without rework, which puts the burden on HPE and AMD to provide hands-on engineering support through the rollout. Anyone who has run a migration off a proprietary stack knows the hardware cutover is a tough weekend, but the software gamble is the real nail-biter.

Market Analysis

From an economic perspective, Vultr appears to be carving out a distinct position in the cloud market. Our assessment is that hyperscalers' long-term capacity commitments have opened a window for independent clouds that can offer high-performance, liquid-cooled compute at scale. With 72-GPU liquid-cooled racks, Vultr aims to maximize compute density per square foot while working within tight power and thermal limits. At these densities, direct liquid cooling has become effectively a requirement rather than an upgrade. HPE's cooling pedigree comes from its exascale systems, but Dell, Supermicro, and Lenovo all field liquid-cooled AI racks at scale. Vultr's edge is therefore more likely to come from HPE's integrated services and single point of accountability than from the cooling itself.

The deal also reflects a broader shift in how cloud providers buy. Instead of buying server boxes piecemeal, they increasingly want pre-integrated rack-scale architectures, and HPE's integration of compute, networking, and cooling into one supported stack is designed to shorten deployment timelines and reduce operational risk. The timing cuts against Helios. NVIDIA's Vera Rubin NVL72 racks are already shipping, Supermicro has begun deliveries, and CoreWeave has a multi-rack cluster running, while Vultr's Helios rollout is planned for 2027 and 2028. Helios will be measured against an installed Rubin base, not a roadmap, and that raises the bar for the first production benchmarks. As model sizes expand, integrated rack-scale engineering appears set to shape the next generation of cloud infrastructure, regardless of which vendor's logo is on the rack.

Looking Ahead

Based on what we are observing, benchmarks will settle the open-fabric question, not press releases. UALoE signals that Ethernet-based approaches are trying to close the gap with proprietary scale-up interconnects such as NVLink, but parity remains unproven until independent latency and utilization data come out of production clusters. The key trend we will be monitoring is how AMD's ROCm handles heterogeneous workload distribution at scale relative to established CUDA workflows, alongside Vultr's delivery timelines and network latency under full cluster load. Our perspective is that cloud providers will likely favor modular architectures that decouple accelerators from network fabrics. That shift should benefit any switch vendor able to prove scale-up performance, HPE included. HyperFRAME will be monitoring whether HPE turns this Vultr win into broader enterprise adoption in coming quarters, and whether a second customer outside AMD's investment orbit follows.