HPE 称数据中心设备涨价未遭客户抵制

COMPUTE

The GenAI wave is indeed lifting almost all of the boats. While Dell has become the largest original equipment manufacturer in the world – a title that the original Hewlett Packard used to have after it bought Compaq in 2001, which had previously eaten Tandem and Digital Equipment Corp to embiggen itself – Hewlett Packard Enterprise is still in the top five and is still the server maker of choice for millions of companies worldwide.

This matters more in the long run than capturing a few AI system deals worth billions of dollars. And besides, HPE could not get the GPU allocations from either Nvidia or AMD to grow its business any faster than it is doing. The big will get bigger faster, as Dell shows.

And as the company’s financial results for its third quarter of fiscal 2026 show, HPE is finally seeing its fortunes improve and is set to help drive AI on various compute and networking engines into enterprises, sovereigns, service providers, and maybe some of the smaller neoclouds.

In the quarter ended in July, HPE posted revenues of $12.21 billion, up 33.7 percent year on year and up 14.4 percent sequentially. Thanks to cost controls and a richer mix of machines as well as passing through price increases on CPUs, GPUs, DRAM, flash, and networking, operating income was up by a factor of 5.6X to $1.39 billion, and net income rose by 5.5X in sympathy to $1.51 billion, representing a very healthy (for a box shifter, mind you) 12.4 percent of revenues. This is one of the best quarters in the company’s history in the datacenter, and you have to go back to the Unix systems boom of the late 1980s and early 1990s to find profitability in systems that is in the same league as what HPE has just turned in.

HPE ended the quarter with $6.22 billion in cash and investments, up 11.6 percent compared to the year ago period.

Here is the monster table describing HPE’s new groups and divisions, which were announced a few quarters ago:

The Cloud & AI group is by far the company’s largest business, lumping together servers, storage, financial services, and tech support for all that gear. The Networking group, which has the acquired Juniper Networks as its core but which also includes various HPE Ethernet switches, including the Slingshot HPC-centric version of Ethernet that its Cray supercomputing unit created two generations of machines back, is smaller but is growing faster.

In Q3, Cloud & AI sales rose by 25.4 percent to just a tad over $9 billion, and was importantly up 17.3 percent sequentially, showing that HPE is converting its backlog of AI systems and more traditional systems to revenue even as its backlog is growing. Cloud & AI operating income was up by more than 3X to $1.54 billion, which is wildly more profitable than the HPE systems business has been in a long, long time. Within this group, server sales were up 35.3 percent to $6.77 billion, while storage rose by 10.2 percent to $1.29 billion. Financial services was steady freddy at $883 million. HPE does not provide operating income at the division level.

The Networking group had sales of $2.89 billion in Q3, up 74.9 percent as HPE is getting an increasing attach rate for networking to its servers and storage. (This was the whole idea of buying Juniper.) However, campus and branch switching is still the main driver of this business, with $1.44 billion in sales, up 31 percent year on year and comprising just under half of revenues. Routing, which is a big business with service providers who don’t want to always buy Cisco Systems gear, drove $788 million in revenue, up by a staggering 3.7X year on year. Despite all that you hear, hyperscalers, cloud builders, and neoclouds buy a lot of routing equipment to lash their datacenters together.

One such company is none other than Oracle, which has its own cloud and which has a deal with HPE to supply it with routers and switches after Q3 F2026 closed. The bulk of this deal is for Juniper QFX switches based on the 102.4 Tb/sec Tomahawk 6 switch ASIC from Broadcom, which launched in June last year and which is shipping now. The deal also includes using Juniper’s AIOps software to manage the networks. HPE says this deal is part of a potential multi-gigawatt datacenter rollout and might span multiple years. As part of the deal, HPE gave Oracle warrants for up to 4.2 million HPE shares, and ahead of announcement day these would have been worth north of $200 million. The actual vesting for these warrants was not disclosed.

Like other deals we have covered, the idea is to use vendor shares as a way to cover some of the cost of the hardware acquired, and the deal itself makes the shares more valuable as soon as it is disclosed, which helps the vendor and the customer alike.

This Oracle deal is not to be confused with a $3.5 billion deal that HPE has inked separatelywith an unknown hyperscaler. “It is a very important use case,” Antonio Neri, HPE’s chief executive officer, explained on the call with Wall Street analysts going over the numbers. “It is a hyperscaler customer, but think about them as an enterprise customer who is going to use our AI inferencing for their own internal usage. So our strategy has not changed from selling large amount of infrastructure for them to serve, like it used to be in the past the cloud business. This is about a multi-billion-dollar AI inferencing for their own internal usage as an enterprise customer. It just happened to be they are labeled as a hyperscaler customer.”

Back to drilling down into the HPE numbers for the quarter, then.

HPE booked just under $1.6 billion in AI server revenues in Q3 F2026, up about 7 percent from the year ago period. HPE added $2.4 billion in new AI system orders in the third quarter, and we presume that HPC systems are being put in this bucket because there is not an HPC deal that is not also an AI deal at this point. HPE’s AI server backlog stood at $6.8 billion as the quarter came to a close. For AI networking, which it has begun to track separately since acquiring Juniper, orders were $700 million for AI-specific systems and the backlog stood at $800 million. HPE did not disclose the AI-related revenues for the Networking group. But we can tell you that AI networking orders are up by 3.5X year on year.

The traditional, non-AI server business was booming in Q3, with sales up 52 percent to $5.17 billion. As was the case with Dell, there is a natural upgrade cycle for vintage HPE ProLiant servers in the installed base, and the machines that companies are buying to replace this old gear are packed with CPUs, DRAM, flash, and hard drives. Such fat boxes have inherently higher costs (justified by the density of the machines and the power savings they deliver compared to racks of older machines) coupled to ever-increasing costs for just about every component in the system.

HPE’s history in the datacenter business is, of course, much longer than the GenAI boom and even the tenure of its current top brass. We have been tracking the company for four decades, and the dataset we have built tracks the core datacenter systems business of HP and then HPE since the Great Recession, which was definitely a dividing line in history. Take a gander at how the financials of this core systems business has changed in the past two decades:

The GenAI boom started helping lift revenues in fiscal 2024, and the addition of Juniper to the HPE portfolio has only accelerated that. And HPE has set record datacenter revenues for core systems – servers, storage, networking, tech support, and financing for the gear – for the past five quarters in a row. Yes, it is dwarfed by Dell, Supermicro, Inspur (IEIT), and Lenovo in the server business, but when you add in networking and financing, the gap is not as large.

Going forward, revenue growth depends on GPU allocations that HPE and its customers can get, and that depends on the power and facilities that those customers can line up.

In any event. HPE’s core system revenue in the third quarter came in at $10.49 billion, up 35.1 percent and up 15.6 percent sequentially. Our allocation of operating income shows the core systems business generated $1.86 billion at the middle line.

Looking ahead, HPE expects to do $13.9 billion to $14.8 billion in business in its fiscal fourth quarter, with the Networking group expected to grow by 11 percent to 13 percent and the Cloud & AI group growing by 60 percent to 72 percent. Clearly, it depends on how the power and facilities line up to GPU allocations.

As far as Neri is concerned, demand is there and people are rebalancing their budgets and increasing them in a way that we have not seen since the Dot Com boom. And there is not a lot of pushback on costs.

“At the beginning of this hyper cycle, on the cost, obviously they were a little bit in shock, and they're trying to navigate through that timing by focusing on understanding the trends and looking at the spot market and the like,” Neri explained. “Once they understood that and understood that particularly on AI, you need to go faster, they figured it out. Look, waiting is not an option. But they're getting smarter about where to land their budgets and how to optimize for these token economics. When you do it on premise, and we do that ourself, and we share the number, we can see up to 60 percent cost benefits on a token basis. And so, no, I don't see hesitation at this point in time, and that's why demand continue to be exceptionally strong.”

SPONSORED FEATURE: HPE and NVIDIA help organizations apply accelerated computing, software, enterprise infrastructure, networking, control plane, and services to the workloads and business outcomes that matter most.

October 1, 2026

September 30, 2026

AMD Buys World Labs For $8.2 Billion To Get AI Model Cred