藉助 Hyperion Research 數據剖析 HPC 與 AI-Sci 市場
Casing The HPC And AI-Sci Market With Hyperion Research
Hyperion Research 在 SC26 早餐會的簡報顯示,2026 年上半年 HPC 與 AI-Sci 現場系統開支按年上升 19.9%,達 164.8 億美元;系統出貨量則只增加 4.4%。
When you go to the Super Computing conference in America or the International Super Computing conference in Europe, one of the hardest things to do is get up early on a Tuesday morning and hustle on over to the IDC and now Hyperion Research breakfast to get a good greasy breakfast and some strong coffee. But you do it so you can be alert for an update on the state of the HPC and now AI for Science markets, and because this breakfast usually follows a very social evening on Monday, if you catch my drift.
If you can make it to that breakfast, you will make it through the week.
NEXTPLATFORM AD
I was not able to attend ISC26 or SC26 this year for medical reasons, but I did get my hands on the presentation put together by Earl Joseph, Bob Sorensen, and Mark Nossokoff for the SC26 breakfast, which gives us a sense of how the first half of 2026 turned out in HPC and AI-Sci Land and also gave out forecasts for spending for on premises gear and for capacity on the clouds aimed at simulation, modeling, and AI for science workloads.
The revenues in the chart above for on premises HPC and AI-Sci systems are in millions of dollars, and as you can see, the spending from January through June was up 19.9 percent to $16.48 billion. (It would be helpful to put the dollar signs in the chart so people know what the numerical units are for each category.) While revenue was up smartly, worldwide unit shipments of on premises HPC and AI-Sci systems only increased by 4.4 percent compared to the first half of 2025. The gap between the two come from having more nodes per system, slightly higher numbers of CPUs per system, and price increases for CPUs, GPUs, DRAM, and flash.
Frankly, we are surprised the increased pricing for components is not more of the gap, and that might be an expression that HPC centers are being more conservative and hanging back one or two generations on CPUs and GPUs so they can get a lower capital outlay than it would take to buy a current system. We do see a lot of Grace-Hopper systems or Intel Xeon-Hopper systems being added in the current Top500 supercomputer rankings and for the prior couple of lists, even though Grace-Blackwell systems have been available for two years now. More current GPUs are too scarce and too expensive for most HPC centers.
While Hewlett Packard Enterprise has been the leading in HPC and AI-Sci on premises installs for many years, since acquiring SGI and Cray, in the first half of 2026 Dell has emerged as the revenue leader.
NEXTPLATFORM AD
Again, this table is in dollars, not unit shipments. Dell’s HPC and AI-Sci system sales were up 67.7 percent in the first half of 2026 to $3.93 billion, while HPE only grew its on premises technical computing revenues by 1.1 percent to $3.44 billion. It is hard to say if this is driven by allocations of GPUs by Nvidia or AMD – Dell has been getting preferential treatment in the past two years from Nvidia for AI systems in general. (It was a billionaire thing between Michael Dell and Jensen Huang, the latter of which no longer runs Nvidia.) It would be interesting to see how many aggregate FP64 flops HPE and Dell sold against those revenues. For all we know, HPE is still winning in the flops race against Dell in the HPC and AI-Sci space.
The revenues per vendor drop off pretty fast below Dell and HPE, as you can see, with dozens of system builders – both OEMs and ODMs – accounting for $9.12 billion in sales for the first half of 2026, up 13.8 percent compared to 19.9 percent for the market at large. Dell’s growth and its size are lifting the class average bigtime. It is too early to tell if Dell will retain its position for the second half of 2026, and we won’t know that until the ISC27 Hyperion Research breakfast next year in Hamburg next June.
It could turn out that HPE is selling a lot more CPU-only machines for traditional HPC ModSim workloads, and Dell is getting more GPU allocations and is selling more machines with accelerators – and not just for GenAI workloads.
We don’t have any numbers for 2026 as yet, which is still not finished, but Joseph flashed up this interesting chart showing how the installed base of GPUs in the HPC and AI-Sci market have changed between 2017 and 2025:
NEXTPLATFORM AD
To my eyes, we keep going at the same pace as between 2024 and 2025 this year, and that means 2026 should end with about 1.2 million GPU accelerators installed worldwide in HPC and AI-Sci systems.
This may sound like a lot, but you always need a compared-to-what context to put things into perspective. Doing some back of the envelope math, There is an installed base of somewhere between 18 million and 20 million GPUs – and that means everything that has shipped in the past six or seven years that is still in the field because no one is unplugging a GPU that isn’t reused somewhere immediately. There is probably another 13 million to 17 million custom AI XPUs in the fleet as well, with Google having the lion’s share with its TPUs (somewhere between 9 million and 11 million). Amazon might have 3 million Trainiums and maybe 1 million to 2 million Inferentia chips. Microsoft and Meta Platforms have hundreds of thousands of Maia and MTIA accelerators, respectively. The Chinese custom XPUs are noise in this data, and China hates that this is true. So that 1.2 million GPUs for HPC and AI-Sci represents somewhere between 3 percent and 4 percent of the total worldwide fleet of 31 million to 37 million GPUs and XPUs that are up and running in the world. (The lower percent is against the higher GPU and XPU installed base, obviously.)
For further perspective, HPC and AI-Sci spending is on the order of several tens of billions per year, including power and datacenter construction, while AI proper with training and inference together is between two and three orders of magnitude larger.
The AI tail that used to be on the HPC dog a decade and a half ago has been cloned and is now dozens of sled teams of dogs that are larger in size than the original parent dog. Which has got to be very weird for the HPC crowd that has spent decades at the bleeding edge of systems architecture and high performance.
NEXTPLATFORM AD
Now, I have a DGX Spark baby supercomputer – and that is what it is – under my desk, which I will be using for some AI model experimentation. If I could send this baby supercomputer back in time to 1979 and land it on Seymour Cray’s desk, he would have wept with joy.
We truly do live in the future. . . .
Speaking of the future, let’s look at some Hyperion Research forecasts, starting with the forecast for on premises HPC and AI-Sci servers:
Basically, this market will triple in a decade ending in 2030, which is decent growth even if it is not exponential like the GenAI boom has been for the past four years. As you can see, the bulk of the sales of on premises HPC and AI-Sci gear is for larger-scale supercomputers, but divisional, departmental, and workgroup systems are still selling at roughly the same proportion today as they were five years ago and moreover are expected to sell at about the same rate four years from now as the decade comes to a close.
Joseph says that the broader HPC and AI-Sci market, including storage, middleware, applications, services, and cloud compute, networking, and storage capacity, will exceed $100 billion in 2030 against $54 billion in on premises system sales.
Of course, HPC centers rent clusters as well as buying them, and you have to count that, too:
As you might expect, there is more spending on the cloud for HPC and AI-Sci applications in North America than in the EMEA and Asia/Pacific regions even though the European and Asia economies as two groups are formidable economic regions in their own right.
Over the forecast period from 2022 through 2030, spending on cloud capacity for HPC and AI-Sci applications will grow by nearly 5X by 2030 to around $30 billion. (Why this forecast didn’t have 2021 as the baseline year like the on premises chart did is unclear.)
Just to make it simple and readily apparent, Hyperion Research put the on premises acquisitions and cloud rental spending for infrastructure for HPC and AI-Sci workloads into a single chart:
Back in 2022, the combined infrastructure spending was just north of $40 billion, and it will grow by 3.4X or so to $135 billion in 2030.
It may look like cloud spending is a lot lower, but you have to remember that the on premises spending is for five years. If you add up five years of cloud spending and compare it to the on premises spending for a particular year, the amount of money being spent for a given number of compute-hours over time is a lot closer than it looks. On premises spending is probably a little larger than cloud spending when looked at this way. And remember: Both on premises revenues and cloud rentals include price increases driven by higher component costs that are definitely being passed on to the end user customers.
The issue is whether the amount of HPC and AI-Sci work is cheaper to be done on premises than on the cloud, and there are a lot of ways to argue about that. What is clear is that many HPC centers are doing both and are expected to continue in this practice.
來源:The Next Platform · nextplatform.com