HPE Pops 5% on $1.2B Vultr Helios Deal While Dell and SMCI Sit Still — The Rack-Scale Tell for 2027 AI Capex

Hewlett Packard Enterprise closed its first commercial order for the AMD Helios rack-scale AI platform on Sept. 30, 2026, a $1.2 billion commitment from privately held cloud operator Vultr that lifted HPE about 5% to $64.55. Dell Technologies added only 1% to $543.65 and Super Micro Computer slipped 0.4% to $40.85 on the same session, even though both also sell AMD-powered AI systems. The dispersionmaps ordering instinct into Big-Three integrators with unequal upside: the Helios contract is a turnkey rack that includes the GPUs, the scale-up fabric, the operating software, and the deployment services, and HPE owns pieces of that stack none of its rivals do.

Each Helios rack is built around 72 AMD Instinct MI455X GPUs backed by 31 TB of HBM4 memory and 1.7 PB/s of aggregate memory bandwidth, with 260 TB/s of scale-up bandwidth inside a single scale-up domain. Per AMD's Helios launch in July 2026, the platform delivers up to 2.9 exaflops of dense FP4 compute per rack and up to 15% more AI compute than Nvidia's competing Vera Rubin NVL72 platform, alongside 50% more HBM capacity. The HPE value-add isn't the silicon — Advanced Micro](amd) designs that. It is the rack-scale enclosure, the liquid-cooling loop, the QFX5252 switch tray built on Broadcom's Tomahawk 6, and the global services team that ships it.

Those are exactly the line items the market is rewarding. When HPE raised its networking segment outlook alongside the Vultr deal, it guided FY27 networking growth to the high-teens to low-20% range and disclosed that data-center networking should compound at a low-50s to high-50s CAGR over the FY27–FY29 window. Routing (which now includes the QFX5252 and the MX series) was guided to a low-20s to high-20s CAGR over the same period. The networking segment, not the server segment, is where the Helios economics land.

The mute reaction from Dell Technologies and Super Micro Computer tells investors where the rack-scale procurement dollar is actually settling. Dell closed its fiscal Q2 FY2027 on Aug. 1, 2026 with $46.97B in revenue, $4.13B in net income and a record AI server backlog of $60.9 billion in a single three-month period. The Vultr order is a rounding mistake against that backlog, and Dell's PowerEdge XE9712 platform already supports AMD MI455X GPUs across multiple SKUs. The market correctly concluded that one $1.2B order changes nothing for Dell.

Super Micro is the more interesting non-mover. SMCI launched its own AMD Helios rack-scale platform at Computex in June 2026, three months before this deal, and delivered Q4 FY2026 revenue of $11.12B with EPS of $1.89 — the strongest sequential ramp in its history. But Super Micro's Helios box runs on its own building-block architecture without an owned networking tier equivalent to HPE Juniper's QFX5252, and it doesn't sell a Tomahawk-class switch. SMCI captures the GPU and motherboard dollars; HPE captures the GPU, the switch, and the services dollars. The market already priced that into SMCI when the Computex launch hit, so the Vultr news didn't add a multiple.

Every Helios rack carries 72 AMD MI455X GPUs, each equipped with 12 HBM4 stacks for a per-GPU memory capacity of 432 GB. Multiply that across racks for Vultr, plus the Microsoft Azure deployment AMD committed to in July 2026, plus Meta's roughly 6-gigawatt infrastructure agreement with AMD reported in February 2026, plus the OpenAI and Oracle commitments AMD disclosed at its Advancing AI event, and the demand line for HBM4 in 2027 exceeds anything the industry has built. HBM4 supply stays the tightest constraint through 2027, and SK hynix is forecast to hold roughly 70% market share in HBM4 supply for next-generation AI accelerators in 2026 — including for AMD's MI455X and Nvidia's Rubin platform.

Two upstream enablers capture margin that HPE cannot. SK hynix is the sole-source HBM4 supplier for the MI455X at the volumes needed; HBM4 pricing has already moved into a $31–$36 per GB range in the spot data, roughly double HBM3e levels. Taiwan Semiconductor Manufacturing provides both the foundry wafer and the advanced2.5D/3D packaging that lets12 HBM4 stacks ride alongside each MI455X die — the CoWoS-S and SoIC-X lines are the bottleneck. And Broadcom — which doesn't design the GPU but does design the Tomahawk 6 silicon inside every QFX5252 switch tray that ships in a Helios rack — collects a per-port royalty on the102.4 Tbps of scale-up bandwidth in each tray.

Vultr's $1.2B order matters as much for who Vultr is as for what it bought. Vultr has been privately held and self-funded for most of its life; it took its first outside equity in December 2024 at a $3.5 billion valuation, with AMD's own corporate venture arm, AMD Ventures, on the cap table. It then raised $329 million in debt from Bank of America, Citi and Goldman Sachs in June 2025 explicitly to expand global AI infrastructure. The Sept. 30, 2026 order is the spending of that capital — a privately held operator, backed by debt and a strategic AMD investor, deploying the entire AMD stack at scale. The buyer of $1.2B in rack-scale AI is no longer the named hyperscaler — it is a privately funded alternative cloud provider with an AMD-aligned cap table.

That shifts the addressable buyer pool. Microsoft committed to deploy Helios across Azure in July 2026, Meta took a 6-gigawatt infrastructure agreement with AMD in February 2026, OpenAI and Oracle are listed as early Helios customers alongside Vultr, and Sovereign AI buyers (Saudi Arabia, UAE, India) have been chasing AMD Helios capacity through 2026 per AMD's launch disclosures. The buyer pool is broadening from four named hyperscalers to include well-capitalized Tier-2 clouds and sovereigns. For HPE, that expands the customer universe for which a turnkey rack-scale product — not just a server SKU — is the right sales motion.

Over the next two quarters, the catalysts are order-flow and segment prints: HPE reports Q4 FY2026 in early December 2026, and investors will parse the networking segment for evidence that the Vultr-class deals are arriving in the order book at the rates implied by the raised outlook (data-center networking at a low-50s to high-50s CAGR through FY29, routing in the low-20s to high-20s). AMD reports Q3 FY2026 in late October 2026 and is expected to begin recognizing full Helios revenue in the quarter; the Instinct data-center segment will be the line to track. Dell's Q3 FY2027 arrives in late November 2026, and investors will check whether the $60.9B AI server backlog grew or shrank.

Over the 1–3 year horizon, the structural shift is more important than any single print. Helios is the first rack-scale AI platform designed by AMD — not by Nvidia — that hyperscalers, sovereigns, and alternative clouds are buying in volume at the same time. The economics of the deal flow toward the network and rack layer, not the GPU layer. That favors integrators with owned networking IP and services scale — currently HPE alone — over commodity box-builders. The2027 build is not a GPU story; it is a rack-scale integration story, and only one of the four names here owns it end-to-end. The risk to the thesis is that Dell or SMCI signs a comparable rack-scale networking partnership — Broadcom Tomahawk-powered, scale-up Ethernet-based — that closes the structural gap before the FY29 CAGR compounds.