GPU Transitions, Aggressive Server Pricing Squeeze HPE Profits

Dell saw a sequential slump in server sales its most recent quarter as customers were awaiting access to systems using Nvidia’s “Blackwell” GPUs, and rival Hewlett Packard Enterprise had a similar issue when it turns in its first quarter of fiscal 2025, which ended in early February.

But HPE apparently had some other issues in its systems business, too, with aggressive pricing from rivals and reactive discounting that ate into its server operating profits in Q1 F2025. HPE chief executive officer, Antonio Neri, and chief financial officer, Marie Meyers, also said on a call going over the numbers that there were some inventory issues as well – presumably relating to earlier generation CPU and GPU compute engines it had in stock and couldn’t move – that hurt operating income. HPE had been expecting operating margins of 10 percent to 11 percent, but they came in at 8.1 percent.

Add it all up, and HPE now is going to be laying off 2,500 workers in the next twelve to eighteen months and taking a writeoff of about $350 million to do so, to get its server business staff more in line with the sales and profitability of its systems business. About $70 million of those savings will be realized in fiscal 2025, with the remaining $280 million in savings coming off the books in fiscal 2026 and fiscal 2027.

“Overall, the quarter was within our guidance but given the performance in servers, we took immediate actions to limit hiring, travel, and discretionary expenses,” Meyers explained on the call. “We also decided to take further actions in the form of workforce reductions and attrition management to better align our cost structure with the current business needs. This tough decision will help streamline our organization, improve productivity, and speed up decision making. We will support affected team members with resources and assistance during this transition.”

That seems pretty vague, but if we had to guess, we would guess that AI augmentation is being deployed, managing functions that are currently being done by people. (This happened a year and a half ago with IBM’s supply chain and certain back office functions.)

The news did not get better. Looking ahead to Q2 F2025 and the rest of the fiscal year, Myers said that HPE expected for revenues in its Server group will only grow in the low single digits year on year, and will decline sequentially in the “mid to high single digits” including a slight sequential decline in AI server revenues. (What HPE has been calling APU Systems until now, and which includes both AI and HPC systems accelerated by GPUs.) Server operating margins will be in the “mid single digits” due to expected tariffs as well as other issues like competitive discounting and AI accelerator availability.

The good news on the revenue front is that HPE booked $1.6 billion in new AI server orders, and Neri said that 70 percent of those orders were for Blackwell systems. But it is not at all clear that the profitability of Blackwell B100, B200, and GB200 systems is higher than for systems based on the prior generations of “Ampere” A100 and “Hopper” H100 and H200 accelerators.

The real issue, which neither Nero nor Meyers discussed, is that all of the ODM and OEM system manufacturers were expecting for Blackwell GPUs to have already long since ramped in volume so they could have ramped their own revenues. Companies have had a year to think about Blackwell devices, which were unveiled in March 2024, and in many cases they are willing to wait to get them rather than accept the H100 GPUs that launched way back in March 2022 or even the H200s with their fatter GPU memory that was announced in November 2023. The OEMs are used to being at the front of the compute engine line with Intel and AMD, but with Nvidia they come after the hyperscalers and cloud builders, who as a group buy as many compute engines as the OEMs who build machines for tens of thousands of large enterprise customers.

HPE has a big focus on enterprise AI, and 40 percent of the orders booked in Q1 F2025 were for such customers. HPE is also chasing governments and the sovereign AI institutions that they are setting up (or are already in existence thanks to national HPC centers), and the telco and other service providers who also need smaller scale AI systems than the cloud builders and hyperscalers are building with the help of the ODMs (which include parts of Lenovo and Inspur, Supermicro these days, as well as Foxconn, Invensys, Quanta, Wywinn, Jabil, and the usual subjects).

Neri added that the second half of the fiscal 2025 year would have a boost in AI server revenues, and that the Server division would be back to 10 percent operating margins by Q4 F2025.

With that as a backdrop, let’s go through HPE’s Q1 F2025 numbers.

In the quarter, HPE’s revenues rose by 16.3 percent to $7.85 billion, with operating income off by 17.5 percent to $433 million and a net income that would have been a whole lot lower had it not been for a $244 million gain on sales of a business. Thanks to that, HPE was able to bring $627 million to the bottom line, which divides out to 8 percent of revenues.

The company exited the February quarter with $13.42 billion in cash and equivalents. Which means it has time and maneuvering room to get its business of selling IT to enterprises on a more stable footing.

In the quarter, HPE’s Server division brought in $4.29 billion, up a very healthy 29 percent, and the sequential 8.8 percent revenue drop was consistent with HPE’s historical trends given that its fiscal year end comes in October. Operating income was off 9.1 percent to $348 million, but were off 36.1 percent sequentially, showing the pressure HPE was under.

Interestingly, sales of traditional servers – meaning those not used in HPC and AI settings and including GPU accelerators – were up by 15.2 percent to $3.39 billion in the quarter, and representing 79 percent of server sales in the period. By contrast, AI servers drove $900 million in revenues, which is 2.2X higher than in the year-ago period, but which is down 40.2 percent sequentially.

As we said above, HPE booked $1.6 billion in new AI system orders in the first quarter. And its cumulative orders for these machines rose to $8.3 billion. It’s current backlog of AI system sales stands at $3.1 billion, and its pipeline is many times larger than this number according to Neri.

What is obvious in the chart above is how most of the AI systems sales are for hardware with a tiny bit of software, and some services add ons. The services are growing, but building full AI clusters with burn in and software installation takes a real datacenter and a huge amount of power. Any vendor offering this full service has to invest in facilities to do this.

You will also notice that the quarterly AI system backlog tends to be a little north of $3 billion in the past seven quarters, and revenues are averaging a little north of $1 billion. Strictly speaking, this is not growth, but steady state with some wiggles. This may be a function of where HPE’s customers are at with AI at this point, minus the traditional HPC systems sales the company does.

One last thing that is important for HPE in the datacenter: The company thinks that it can close the $14 billion deal to acquire Juniper Networks, despite a lawsuit brought by the US Department of Justice blocking the deal. HPE doesn’t need Juniper – it has AI and HPC switching thanks to Cray’s “Rosetta” Slingshot interconnect – but we could argue that Juniper needs HPE to better compete against Cisco Systems and Arista Networks in the datacenter.

It will be interesting to see the HPE arguments and the Justice Department counter-arguments. Had a lawsuit not been launched, one could imagine this whole thing being solved by one trip by Neri to visit President Trump in the White House. The trial starts on July 9.