HPE Just Sold $1.2 Billion of Hardware Because the Network Became the Point

Hewlett Packard Enterprise held its Networking Investor Day on September 30, 2026, and used it to make an argument the market accepted quickly. HPE told investors that networking, not servers, is where its AI growth now lives, and it raised its fiscal 2027 networking revenue growth outlook to the high-teens to low-20s range from a previous forecast of 14 to 17 percent.

The order it announced alongside that forecast is the part worth noting. Vultr placed a $1.2 billion order for HPE's AMD Helios AI racks, which bundle AMD chips with HPE networking switches and software for deployment in Vultr's US data centers. It is HPE's first AMD Helios order, and the company has not disclosed how the $1.2 billion splits across networking, compute, software, services, and other components.

The longer-term numbers show how far the outlook moved. HPE now expects networking revenue to grow at a compound annual rate in the high teens from fiscal 2026 through fiscal 2029, up from a prior forecast of 5 to 7 percent annual growth. Data center networking alone is projected at a low-to-high 50s percent CAGR through fiscal 2029, routing at a low-to-high 20s percent, and campus and branch at a high single-digit rate. It also raised its target for annual run-rate cost savings from the Juniper Networks acquisition to $800 million by the end of fiscal 2028, up from at least $600 million.

Why the bundle is the strategy

For years, buying AI infrastructure meant picking servers first and treating networking as whatever connected them. HPE's bet is that customers have flipped the order. Teams building large training and inference clusters care about latency, congestion, and interconnect behavior between accelerators, and those are properties of the fabric rather than the box. A pre-tested stack that includes switches and the software that runs the network removes a class of integration risk the buyer would otherwise own.

The Vultr order is a clean illustration. The Helios racks are AMD-based, and they pull HPE switches and software along with them. HPE's networking chief Rami Rahim said at the event that orders grew 3.5 times faster than revenue in the third quarter, and framed supply availability as the constraint on sales rather than demand. The company doubled its networking supply-purchase commitments in the latest quarter, which is what a supplier does when it believes the backlog is real and wants to convert it. HPE also raised its fiscal 2026 networks-for-AI cumulative order expectation to more than $3 billion, above its previous $2.5 billion to $3 billion range.

The reported financials behind the forecast are unusually strong. HPE posted record fiscal third-quarter 2026 networking revenue of $2.9 billion, up 74.9 percent year over year. CFO Marie Myers said on the earnings call that demand continues to outpace supply, and repeated that assessment in an interview the same week. The stock closed several percent higher on the investor day and reached a record, up roughly 170 percent for the year.

The part investors should scrutinize

A raised forecast and a large order are two different kinds of evidence, and HPE's own framing hints at the gap between them. The immediate test is converting orders into shipments, revenue, and cash flow. A $1.2 billion order with an undisclosed split across five categories is not the same as $1.2 billion of high-margin networking revenue, and the mix will determine whether the margin expansion HPE is projecting actually arrives.

HPE expects networking operating margins in the mid-to-high 20s in fiscal 2027, up from the low 20s expected for fiscal 2026. That improvement is supposed to come from Juniper integration synergies and operating efficiency. Juniper brings a strong enterprise portfolio, and it also brings overlapping products, separate go-to-market motions, and a channel that has to be rationalized. Integration programs of that size tend to surprise on cost before they surprise on savings.

The competitive position is genuinely contested. Cisco and Arista both compete for the same networking dollars, and Arista in particular has spent years winning AI cluster fabric deals on the strength of its software and its Ethernet roadmap. HPE's advantage is that it sells the rack and the fabric together, which gives it a chance to attach higher-margin networking to compute orders that a pure networking vendor would have to win separately. That advantage is real and it is also the reason the strategy is fragile. If customers decide they would rather source the fabric from a specialist and the compute from someone else, the bundle loses its pull, and the networking attach rate becomes the first thing to erode.

The AMD angle nobody is talking about

Read the Vultr order again and notice which accelerators are inside the racks: AMD Helios, rather than anything from NVIDIA.

That is a smaller detail than the dollar figure and a more interesting one. HPE's first Helios win landing at $1.2 billion with a private cloud provider suggests that the second-source market for AI accelerators has a buyer who is willing to commit at scale before buying it in small increments. Vultr competes with the hyperscalers on price and on latency to market rather than on volume, and a provider in that position has every incentive to build on whichever accelerator gives it the best cost per unit of work.

HPE's decision to lead with an AMD-based rack rather than an NVIDIA-based one also says something about where the company thinks it can win. NVIDIA's rack-scale systems are tightly bound to NVIDIA's own networking and software stack, which limits how much of the fabric a partner can sell around them. AMD's ecosystem is less vertically closed, which leaves room for HPE to attach its switches, its software, and eventually its services. Bundling is only a strategy if there is something left to bundle.

The risk in that bet is symmetric. If the AMD route stays a niche, HPE has built its networking growth story on a platform with a smaller addressable market. If it works, the company has positioned itself as the integrator for every customer that wants an alternative to vertical lock-in, which is a larger story than a single rack order.

What the shift signals about AI infrastructure

The wider story is about where the money goes as AI buildouts mature. GPU procurement dominated the headlines through 2025 and 2026, and it is now the most negotiated, most financially engineered, and most thoroughly benchmarked line item in the stack. Networking has received a fraction of that attention, which is precisely why it looks like the next contested layer. It is also stickier than compute. Once a fabric design is built into a data center, swapping it means re-architecting, which makes expansions and refresh cycles more likely to follow the original blueprint.

That stickiness cuts both ways. It is what makes HPE's forecast plausible, and it is what makes the first few reference deployments so consequential. Vultr is an early proof point rather than a settled trend. The real signal will arrive over the next few quarters, when the shipments land, the mix becomes visible, and the operating margin either moves into the mid-to-high 20s or explains why it did not.

Networking has been the quiet part of the AI stack for two years. HPE just bet a large part of its growth story that it is about to become the loud part, and it has one $1.2 billion order to show so far.