Hewlett Packard Enterprise has booked its first paying customer for AMD’s answer to Nvidia’s dominant AI server racks. On September 30, 2026, HPE announced that Vultr placed a $1.2 billion order for AMD Helios AI Rack by HPE systems, the first disclosed commercial deployment of a rack-scale design the two companies spent nearly a year building together. The deal puts 72 AMD Instinct MI455X GPUs per rack directly against Nvidia’s NVLink-based racks, and it tests whether standard Ethernet can do a job Nvidia has owned for years. For AMD, long treated as the distant number two in AI data center chips, a named billion-dollar rack order shifts the conversation from roadmap slides to deployed hardware.
What HPE and Vultr Just Announced
HPE unveiled the order at its Networking investor day in Sunnyvale, California, and framed Helios as the main driver behind a raised financial outlook for the coming fiscal year. Vultr, which HPE describes as the world’s largest privately held cloud infrastructure company, will deploy AMD Helios AI Rack by HPE systems across its U.S. data centers, according to HPE’s announcement covered by StorageReview. Neither company disclosed how many racks the $1.2 billion covers or a firm delivery date for the hardware itself, a gap that makes the deal hard to price on a per-rack basis. DIGITIMES separately confirmed the order alongside comments from HPE’s networking and Juniper leadership at the same investor event.
J.J. Kardwell, Vultr’s CEO, framed the order around a supply problem that has defined AI infrastructure buying for the past two years. “Demand for high-performance AI infrastructure continues to outpace available capacity, and our work with AMD and HPE enables us to bring Helios rack-scale compute and scale-up networking online for customers faster,” Kardwell said, according to StorageReview’s reporting. “Together, we’re delivering the memory capacity, interconnect performance, and energy efficiency customers need to run the largest and most demanding AI training and inference workloads.”
Antonio Neri, HPE’s president and CEO, cast the order as proof of the company’s post-Juniper pitch to the market. “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,” Neri said, per the same StorageReview account.
Inside the AMD Helios AI Rack: The Numbers
GPUs, CPUs, and Memory
Each Helios rack packs 72 AMD Instinct MI455X accelerators, paired with sixth-generation AMD EPYC “Venice” CPUs, AMD Pensando Vulcano AI network interface cards, and AMD’s ROCm software stack, according to specifications reported by StorageReview and Fierce Network. The MI455X is AMD’s first chip built on its CDNA 5 architecture, and AMD specifies 432GB of HBM4 memory for each GPU. Multiply that across a full rack and the math gets dramatic fast.
Helios rack capacity (AMD figures, per GPU x 72):
432GB HBM4/GPU x 72 GPUs = 31TB total rack memory
Rack memory bandwidth = 1.7PB/s
Rack compute (FP4 precision) = 2.9 exaFLOPS
Scale-up bandwidth inside rack = 260TB/s
Scale-out bandwidth leaving it = 43TB/sThose are AMD’s own published figures rather than independently measured results, a distinction that matters once you get to how Helios stacks up against Nvidia later in this piece.
| Component | Specification |
| GPUs per rack | 72x AMD Instinct MI455X |
| GPU architecture | AMD CDNA 5 (first-generation) |
| Memory per GPU | 432GB HBM4 |
| Total rack memory | 31TB HBM4 |
| Rack memory bandwidth | 1.7PB/s |
| Rack compute (FP4) | 2.9 exaFLOPS |
| CPU | 6th-gen AMD EPYC “Venice” |
| Network interface cards | AMD Pensando Vulcano AI NICs |
| Scale-up fabric | UALink over Ethernet (UALoE) |
| Scale-up bandwidth | 260TB/s inside the rack |
| Scale-out bandwidth | 43TB/s leaving the rack |
| Switching | 6x HPE Juniper Networking QFX5252 trays |
| Switch silicon | Broadcom Tomahawk 6 |
| Software stack | AMD ROCm |
| Cooling | HPE direct liquid cooling |
Networking and Cooling
The 72 GPUs pool memory and behave like a single large accelerator over UALink over Ethernet, an open scale-up fabric that runs on standard Ethernet rather than a proprietary interconnect, according to StorageReview’s breakdown of the rack. HPE’s own contribution sits mostly in that networking layer. Six HPE Juniper Networking QFX5252 scale-up Ethernet switch trays occupy every Helios rack and tie all 72 GPUs together, running on switch silicon that AMD and HPE developed jointly with Broadcom’s Tomahawk 6 chip. Cooling runs through HPE’s direct liquid cooling system, which the company markets as derived from its exascale supercomputer work, with HPE Services handling installation and cooling integration for customers.
Why HPE Is Betting on Ethernet Instead of Nvidia’s NVLink
Rami Rahim, HPE Networking’s executive vice president, president, and general manager, put the stakes plainly at the investor day. “We are taking standards-based Ethernet to a new layer of the data center stack for the first time,” Rahim said, according to Fierce Network’s coverage of the event.
That statement targets the layer of the rack where GPUs talk to each other during training and inference, known as scale-up networking, as opposed to scale-out networking that links racks to each other across a data center. Nvidia handles the scale-up job with NVLink and NVSwitch, a proprietary fabric baked into every GB200 and GB300 rack it sells. AMD and HPE are instead running UALink over Ethernet, built on commodity switching gear rather than a single vendor’s custom silicon. The pitch to buyers is straightforward: open networking should mean more suppliers, more price competition, and less lock-in than a closed fabric controlled by one company. Whether UALoE can actually match NVLink’s latency and throughput at scale is a question Helios customers like Vultr are now positioned to answer in production, not just in lab tests.
AMD Helios vs Nvidia GB200/GB300 NVL72: Head-to-Head
Stripped down to GPU count, the Helios rack and Nvidia’s flagship NVL72 racks look nearly identical. Both pack 72 accelerators into a single rack-scale unit designed to act as one giant GPU for the largest training and inference jobs. The difference is everything surrounding those GPUs: the interconnect standard, the networking vendor, and the software stack buyers have to commit to.
| Attribute | AMD Helios (HPE) | Nvidia GB200 NVL72 | Nvidia GB300 NVL72 |
| GPUs per rack | 72x MI455X | 72x Blackwell | 72x Blackwell Ultra |
| CPU pairing | AMD EPYC “Venice” | Grace (36x, paired 2:1) | Grace (rack-scale config) |
| Scale-up interconnect | UALink over Ethernet (open) | NVLink/NVSwitch (proprietary) | NVLink/NVSwitch (proprietary) |
| Software ecosystem | ROCm | CUDA | CUDA |
| Cooling | HPE direct liquid cooling | Direct liquid cooling | Direct liquid cooling |
| Named 2026 customer | Vultr ($1.2B order) | Vultr (ordered June 2026) | Multiple hyperscalers |
| Independent benchmarks vs rival | Not yet published | Widely published | Widely published |
GPU-count parity does not mean performance parity. AMD’s 31TB and 2.9 exaFLOPS figures come from AMD itself, and no independently reproduced benchmark pitting the MI455X against Nvidia’s Blackwell Ultra or upcoming Vera Rubin generation has surfaced yet. Until Helios racks are actually running customer workloads at Vultr, and the results get tested by outside labs, the honest comparison is architectural rather than a verified performance ranking.
Vultr’s AMD Bet Predates Helios by Two Years
The Vultr order did not come out of nowhere. AMD Ventures co-led Vultr’s December 2024 funding round, which valued the cloud provider at $3.5 billion, and Vultr was among the first cloud providers to deploy AMD’s earlier Instinct MI325X and MI355X GPUs. About a year after that funding round, Vultr said it would spend more than $1 billion on a 24,000-GPU AMD Instinct MI355X cluster in Springfield, Ohio, with plans to move to Helios once it became available. In July 2026, Vultr opened pre-orders for reserved Helios capacity targeting deployments that run from 2027 through 2028, well before HPE’s order became public.
That history matters for reading the $1.2 billion figure correctly. This is less a surprise entry into AMD’s camp than the formalization of a relationship Vultr had already signaled it wanted, with HPE stepping in as the rack builder once the Helios design was ready to ship.
Vultr Isn’t Picking Sides: HPE Now Builds for Both Camps
Vultr’s AMD enthusiasm has not stopped it from buying Nvidia hardware too. In June 2026, Vultr chose HPE-built Nvidia GB300 NVL72 racks connected with Nvidia’s own Spectrum-X networking, meaning HPE now supplies the same customer with rack-scale systems on both GPU ecosystems. Microsoft separately said in July 2026 that it would ramp Helios at scale on Azure, giving AMD’s design a second hyperscaler validation point beyond a single neocloud.
Rahim argued that dual-vendor flexibility is exactly what smaller, independent cloud providers want from a supplier like HPE. “Many neoclouds value the simplicity of buying a complete system and buying from a single technology provider,” he said, according to Fierce Network’s reporting on the investor day. HPE’s pitch, in other words, is not “choose AMD over Nvidia.” It is “buy whichever GPU you need, and let HPE build the rack around it.”
Market Reaction: How Investors Read the Deal
HPE shares rose roughly 3.5% in premarket trading on the news, while AMD shares slipped slightly, around 0.7%, according to market coverage of the announcement reported by GuruFocus. The asymmetry makes sense on closer reading. The order is primarily an HPE revenue and strategy story, since it validates the company’s post-Juniper rack-building pitch, while AMD’s upside from the deal was arguably already baked into its stock given Vultr’s existing, publicly disclosed AMD relationship going back to 2024.
It is a reminder that a single hardware order can move two stocks in opposite directions depending on what each company’s investors were already pricing in. In a video interview posted to Yahoo Finance, an HPE finance executive called the Vultr announcement one of the standout moments of the investor day, underscoring how central the deal was to the company’s messaging that week.
HPE’s Networking Revenue Bet Pays Off
HPE raised its fiscal 2027 networking revenue growth outlook to a range from the high teens to the low 20s in percentage terms, up from the 14% to 17% range it had forecast on September 2, 2026. Rahim identified Helios as the single new factor behind the revision, according to Fierce Network. He also sized the Helios networking opportunity, meaning HPE’s own switching and software slice rather than the value of the full rack, at more than $1 billion over the next two years, and he said that figure does not assume HPE gains additional market share beyond what it already holds in the Helios ecosystem.
Rahim was careful not to overclaim. “It’s still very early to talk about market share at this point in time,” he said. HPE is, for now, the only networking vendor inside Helios racks, a position that gives it outsized exposure to AMD’s success or failure in winning additional rack-scale orders.
The $14 Billion Juniper Acquisition Finally Shows Its Hand
HPE closed its $14 billion acquisition of Juniper Networks in July 2025 after a drawn-out antitrust review. The Helios order is the clearest commercial proof yet of what that deal was supposed to buy: the ability to sell a complete AI rack, including switching, rather than competing only on servers or standalone network gear. HPE pitched exactly this full-stack argument to investors at its Discover conference in June 2026, branding itself a one-stop AI infrastructure shop. Notably, the Vultr-Juniper relationship predates the HPE acquisition by nearly three years, which is part of why HPE frames the Helios order as a continuation of an existing partnership rather than a cold sale.
The Competitive Stakes for AMD in AI Data Centers
AMD has spent years positioning its Instinct GPU line as the credible alternative to Nvidia, but credibility built on roadmap announcements is different from credibility built on named, dollar-denominated deployments. A $1.2 billion order with a committed timeline gives analysts something harder to dismiss than a benchmark slide at a trade show. AMD CEO Lisa Su told investors on the company’s August 2026 earnings call that initial Helios shipments would begin later in the third quarter of 2026, step up in the fourth quarter, and keep ramping through 2027, a cadence that gives the market concrete dates to check AMD’s execution against.
The harder battle may be software, not silicon. Nvidia’s CUDA ecosystem has a years-long head start over AMD’s ROCm stack in developer tooling, library support, and the accumulated muscle memory of machine learning teams. Winning Vultr and reportedly Microsoft’s Azure as Helios customers helps AMD’s case, but converting GPU-count parity into workload-level parity still depends on ROCm closing that software gap in production, not just in AMD’s own benchmarks.
Historical Context: The Rack-Scale AI Arms Race
For several years, Nvidia’s rack-scale designs, including the GB200 NVL72 and its GB300 NVL72 successor, set the template for how the industry builds AI training and inference infrastructure. NVLink and NVSwitch became the signature of that approach: a tightly integrated, proprietary fabric that delivers strong performance but also locks buyers into Nvidia’s full stack, from GPU to switch to software. AMD, HPE, and Broadcom lined up the rack-scale design behind Helios last December, explicitly built around UALink over Ethernet as an open alternative to NVLink. Helios is the first product from that effort to reach a named, dollar-valued commercial order, which is why the Vultr deal carries more weight than a typical hardware partnership announcement, as Converge Digest and other infrastructure trade press noted in their own coverage of the announcement.
Predictions: Where the AMD-Nvidia Rack Fight Goes Next
- More neoclouds will likely follow Vultr’s dual-sourcing model, buying both Nvidia and AMD/HPE racks to hedge against GPU supply constraints, rather than standardizing on a single vendor.
- AMD’s fourth-quarter 2026 and early-2027 earnings calls should show whether Helios shipments are ramping on the schedule Lisa Su outlined in August. Any slippage would undercut the deal’s credibility quickly.
- Independent, reproducible benchmarks comparing the MI455X to Nvidia’s Blackwell Ultra and Vera Rubin generations should start appearing once Vultr’s racks go live, and the industry will scrutinize them closely given AMD’s figures so far are self-reported.
- HPE will likely disclose additional Helios orders in coming quarters if its raised fiscal 2027 networking guidance is going to hold up, and a further guidance increase would signal the bet is paying off faster than planned.
- Adoption of UALink over Ethernet beyond AMD’s own ecosystem, such as other accelerator vendors building on the same open standard, would be the clearest sign that open scale-up networking is taking real share from Nvidia’s proprietary approach rather than just giving AMD a single showcase deployment.
What This Means for Enterprise AI Buyers
For enterprises shopping for AI infrastructure, the practical takeaway is more competition at the rack level, which historically pushes pricing and terms in buyers’ favor. But the disclosure gaps in this announcement are real: no public rack count, no fixed delivery date, and no independent benchmark data comparing Helios to Nvidia’s current or upcoming racks. Treat the HPE-Vultr order as a capacity and roadmap signal worth watching, not yet as proof that AMD’s rack-scale hardware matches Nvidia’s on performance. That evidence will come from Vultr’s own production results once the racks are actually running customer workloads, likely sometime after the fourth quarter of 2026.
Frequently Asked Questions
What is the AMD Helios AI Rack by HPE?
It is a rack-scale AI server system built jointly by AMD, HPE, and Broadcom, combining 72 AMD Instinct MI455X GPUs, AMD EPYC “Venice” CPUs, AMD networking hardware, and HPE’s switching and liquid cooling into one integrated unit designed for large-scale AI training and inference.
How much is the HPE-Vultr deal worth?
HPE announced a $1.2 billion order from Vultr on September 30, 2026, marking HPE’s first disclosed commercial order for the Helios platform.
How many GPUs are in a single Helios rack?
Each rack contains 72 AMD Instinct MI455X GPUs, the same GPU count Nvidia uses in its GB200 and GB300 NVL72 rack systems.
How does a Helios rack compare to Nvidia’s GB200 or GB300 NVL72?
Both platforms pack 72 GPUs per rack, but Helios uses AMD’s CDNA 5 architecture with an open UALink over Ethernet interconnect, while Nvidia’s racks use proprietary NVLink and NVSwitch fabric. No independently verified head-to-head benchmark between the two has been published yet.
When will Vultr’s Helios racks start running?
AMD CEO Lisa Su said on the company’s August 2026 earnings call that initial Helios shipments would begin later in the third quarter of 2026, step up in the fourth quarter, and keep ramping through 2027. Vultr’s own pre-orders, opened in July 2026, target deployments running from 2027 through 2028.
Is Vultr only using AMD hardware now?
No. Vultr also ordered HPE-built Nvidia GB300 NVL72 racks with Nvidia Spectrum-X networking in June 2026, making it a customer for both GPU ecosystems through HPE.
Does this deal affect HPE’s Juniper Networks acquisition story?
Yes. The order is the clearest commercial evidence so far that HPE’s $14 billion 2025 acquisition of Juniper is paying off, since Helios routes HPE Juniper Networking switch hardware into every rack it ships.
Has anyone independently benchmarked the MI455X against Nvidia’s latest chips?
Not yet publicly. The memory, bandwidth, and compute figures cited for Helios come from AMD’s own specifications rather than third-party testing, and that comparison is expected to firm up once racks are deployed at customers like Vultr.
