On November 3, 2026, Hyperion Research announced that the on-premises HPC and AI-for-science server market reached $16.5 billion in the first half of 2026. These are servers that organizations install on their own premises for large-scale computing and simulation. The growth rates the firm gave diverge: revenue rose 19.9%, while shipments rose 4.4%. The figures cover computing infrastructure for scientific research and engineering, and they differ from figures for the AI server market as a whole, which includes chat services. What lies behind revenue growing faster than unit volume? Tracing the rise of costly configurations and the integration of AI into research workflows reveals a shift in demand that location alone does not show.
Breaking Down the Gap Between 19.9% Revenue Growth and 4.4% Shipment Growth
The $16.5 billion market size and the two growth rates come from Hyperion's announcement. The comparison period, however, needs care. The announcement uses wording that implies a year-over-year comparison, yet it also describes the revenue baseline as the second half of 2025, so the two descriptions do not match. There is no basis for deciding which is an error, so the 19.9% cannot be confirmed as either a year-over-year or a period-over-period figure.
Calculating from Hyperion's published 19.9% revenue growth and 4.4% shipment growth, aggregate revenue per unit rose roughly 14.8%. This calculation assumes the announcement's two growth rates cover the same period. The formula is (1.199 ÷ 1.044 − 1) × 100 ≈ 14.8%. Because it measures the change in average revenue per unit, it is not a simple subtraction of 4.4 from 19.9. The shipment figure here refers to the sales unit used in the announcement.
This 14.8% should not be read to mean that the same servers became that much more expensive. The aggregate average rises whether individual machine prices change or whether high-priced models account for a larger share of revenue. Even for the same single unit sold, differences in GPU configuration or system scale change its contribution to revenue. Nor can growth in unit volume be read as growth in GPU count or computing capacity.
Hyperion itself cites two background factors: an increase in expensive GPU configurations and rising memory costs. The company tracks the market through multiple surveys of both buyers and suppliers, but the public announcement does not reveal what share of the revenue increase came from pricing versus configuration mix. The calculated rise in average revenue and its breakdown must be read separately.
The next question after unit counts, then, is what is inside the machines. If research institutions build AI into conventional computing, the work they expect from the servers they buy changes as well.
Simulation and AI Linked in the Same Research Workflow
According to the public summary of Hyperion's generative AI and HPC user survey, 53% of respondents said they had entered the stage of integrating generative AI capabilities into existing work. Hyperion says this is the first time a majority has reached that stage. The 53%, however, is a figure for respondents to that survey, not an adoption rate for all companies or all HPC users. The public summary also does not disclose the sample size or questionnaire.
What does integrating AI into research actually involve? One concrete example is "Blue Lion," a project at Germany's Leibniz Supercomputing Centre (LRZ). In the plan LRZ presented on June 11, 2025, it described an approach that combines classical computing and simulation with AI. This is a concept for a system scheduled to go into operation in 2027, and it was not presented as a project that makes up the first-half sales figures discussed here.
LRZ describes two directions of use. One is supplementing conventional simulation with AI "surrogate models," which learn from simulation results and the like and approximate their behavior. The other is a flow in which data generated by simulation is used to train further AI models.
In other words, conventional computing and AI do not merely compete for equipment as separate uses. A relationship can also arise in which simulation produces material for training AI, and that AI is then used to supplement simulation. In this case, those choosing equipment must consider where in the research workflow to combine simulation computing with AI training and use. Server unit sales alone cannot capture this combination of tasks.
That said, the Blue Lion plan does not describe replacing all physical simulation with AI. The materials cited here also do not allow the reductions in computing time or cost to be quantified. And because the survey population differs from that of the sales market, the 53% response rate cannot be used to calculate how much AI integration pushed up the 19.9% increase. What can be said is that there is a movement to build AI into research workflows, along with facility plans premised on it.
Dell's Lead Backed by Mid-Range Demand and Waves of Large Procurements
Hyperion announced that Dell overtook HPE, which had held the top position for more than 25 years. Its explanation cites Dell's large deals, as well as a surge in customers for mid-range machines that bring AI into existing science and engineering work. The reading is that not only a small number of large systems aiming for peak performance but also mid-scale demand for adding AI to existing workloads is supporting vendor revenue.
As for HPE, Hyperion explains that in 2025 it booked large revenue from sales of multiple exascale-class systems. Procurements of this scale do not necessarily recur every year. When large deals cluster in a given period, equipment refresh and procurement timing affect the revenue ranking separately from the broadening of everyday demand.
The change at the top reflects both a widening customer base and the cycle in which large-deal revenue is booked. This ranking alone therefore cannot be used to judge the relative product performance of Dell and HPE, or to assume that customers switched en masse. In reading the market for research servers, one must check both which scale of demand grew and when a small number of large deals came in.
Cloud Use Growing Alongside On-Premises Expansion
In a separate public summary on cloud use for AI-for-science and HPC, Hyperion likewise says demand continues to expand. The reasons it cites for use are the ability to scale resources up and down quickly as needed, access to new hardware not available on-site, and the ability to run computations on a larger scale than on-site facilities allow.
These reasons do not fit a one-directional explanation such as "because in-house equipment becomes unnecessary." Even when research runs on owned equipment, additional resources may be needed depending on the scale of processing or the hardware wanted. The question of how to combine on-site equipment with external resources is compatible with growing on-premises demand.
The company forecasts that cloud spending on AI-for-science and HPC will grow at a compound annual rate of 19.3% and reach $30 billion in 2030. This is a forecast, not an actual result for the first half of 2026 announced this time. Moreover, spending on cloud use and sales of on-premises servers cover different scopes. The $30 billion and $16.5 billion figures cannot be set side by side to compare market sizes, nor can their ratio be treated as market share.
This sales data does not measure the share of the same customers who moved from the cloud back to on-site equipment. There is no basis for reading on-premises growth as a return from the cloud. At the same time, the integration of AI into research makes it harder to judge actual use from equipment purchase amounts alone. Which research stages use computing resources, how heavily on-site equipment is run, and when work is sent externally become the material for evaluating cost and capability.
If a breakdown showing the contributions of pricing and configuration were provided, along with actual processing capacity and utilization, it would become possible to check which research tasks the revenue increase is supporting. Research institutions choosing equipment can also make decisions about combining on-site infrastructure with the cloud more concrete by layering the resources each research stage needs and actual usage on top of the number of units purchased.