Spending on on-premise scientific HPC and AI systems grew by 19.9% in the first half of 2026, reaching $16.48 billion. These figures are from the presentation by Earl Joseph, Bob Sorensen, and Mark Nossokoff for the SC26 breakfast, organized by Hyperion Research.
Over the same period, global unit shipments increased by only 4.4% compared to the first half of 2025. According to the source, the gap between the two indicators is explained by a higher number of nodes per system, a slightly higher number of processors, and rising prices for CPUs, GPUs, DRAM, and flash memory. The source’s editor is surprised that the increase in component costs does not weigh more heavily in this gap, and perhaps sees it as a sign that data centers remain one or two generations behind to limit their capital expenditure.
Dell Ahead of HPE in Revenue
For several years, Hewlett Packard Enterprise (HPE) has been the leading provider of on-premise scientific HPC and AI systems, following the acquisitions of SGI and Cray. In the first half of 2026, Dell takes the lead in revenue. Its sales in this area surged by 67.7% to reach $3.93 billion, while HPE saw its revenues grow by 1.1% to $3.44 billion.
The editor notes that these figures are expressed in dollars, not units, and that it is difficult to know if this gap is due to GPU allocation from Nvidia or AMD. He recalls that Dell has benefited from preferential treatment from Nvidia for AI systems for two years. Other manufacturers, OEMs and ODMs, total $9.12 billion in sales, up 13.8%. According to the source, it is too early to say whether Dell will retain its position in the second half; the answer will be provided at the Hyperion Research breakfast at ISC27 in Hamburg in June.
An Installed Base of GPUs and XPUs
Hyperion Research also presented the evolution of the installed base of GPUs in HPC and scientific AI between 2017 and 2025. The editor estimates that in 2026 the pace will remain comparable to that of 2024-2025, which would lead to approximately 1.2 million GPU accelerators installed worldwide for these uses by the end of the year.
This figure takes on its full meaning in relation to the entire installed base. According to source estimates, the installed base of GPUs would be between 18 and 20 million, and that of custom AI chips, or XPUs, between 13 and 17 million. Google holds the largest share, with 9 to 11 million TPUs, while Amazon would have about 3 million Trainium and 1 to 2 million Inferentia. Microsoft and Meta each have a few hundred thousand, with Maia and MTIA chips. The total installed base would therefore be between 31 and 37 million GPUs and XPUs, placing the 1.2 million for HPC and scientific AI at 3 to 4% of the total.
Forecasts Until 2030
Hyperion Research forecasts that the on-premise systems market will triple by 2030. Joseph estimates that the broader HPC and scientific AI market, including storage, software, services, and cloud, will exceed $100 billion in 2030, compared to $54 billion for on-premise system sales alone.
Spending on cloud capacity is expected to grow by about five times between 2022 and 2030, reaching nearly $30 billion. Total infrastructure spending, on-premise purchases and cloud rental, would increase from over $40 billion in 2022 to $135 billion in 2030. The source notes that many data centers combine both approaches and are expected to continue doing so.
