AI & ComputeLOT 261122 · 4 min read

Dell leads $16.48B HPC and AI-for-science market in H1 2026

Dell took the lead in first-half 2026 HPC and AI-for-science revenue, but higher system prices, scarce GPUs and cloud growth shape a $135 billion outlook.

By Sophie Lindqvist883 words

Casing The HPC And AI-Sci Market With Hyperion Research
Casing The HPC And AI-Sci Market With Hyperion Research — Pexels / pixabay

What changed in first-half 2026?

Dell led first-half 2026 on-premises HPC and AI-for-science system revenue with $3.93 billion, up 67.7% year over year, according to Hyperion Research data presented for SC26.

That result moved Dell ahead of Hewlett Packard Enterprise, whose revenue reached $3.44 billion, up 1.1%. HPE had long led on-premises installations, a position shaped by the SGI and Cray acquisition lineage.

Hyperion’s figures came from an SC26 breakfast presentation by Earl Joseph, Bob Sorensen and Mark Nossokoff. The report covers on-premises systems used for simulation, modeling and AI for science.

Worldwide on-premises revenue for January through June reached $16.48 billion, up 19.9% from the first half of 2025. Unit shipments rose only 4.4%. Hyperion attributes the gap to more nodes per system, slightly higher CPU counts per system, and higher prices for CPUs, GPUs, DRAM and flash.

The analysis offered a plausible but unproven explanation: HPC centers may hold CPUs and GPUs one or two generations back to reduce capital outlay. NVIDIA Grace Hopper and Intel Xeon-Hopper systems continue to appear in current and recent Top500 lists, although Grace-Blackwell has been available for two years. Scarce, expensive current GPUs support the delay explanation, but the source offers no unit data to prove it.

What does the 2030 forecast require?

Hyperion forecasts that the on-premises market will triple during the decade ending in 2030. Larger supercomputers account for most sales, while divisional, departmental and workgroup systems retain roughly their five-year-old share and are expected to keep that proportion near the decade’s close.

Earl Joseph said: “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.”

Cloud capacity spending for HPC and AI for science should reach nearly five times its 2022 level, around $30 billion, by 2030. North American spending should exceed both EMEA and Asia/Pacific. The forecast uses 2022, while the on-premises chart uses 2021; the source could not explain that mismatch.

The slides also place combined on-premises acquisitions and cloud rentals at just above $40 billion in 2022 and $135 billion in 2030, a 3.4-fold increase. The source does not reconcile those totals with the separate $54 billion on-premises and $30 billion cloud figures. They remain parallel forecasts, not a single reconciled market total.

Can Dell sustain its lead?

Dell’s first place may not mean it leads delivered computing power. The analysis says Dell received preferential NVIDIA treatment for AI systems during the prior two years and links that advantage to Michael Dell’s relationship with Jensen Huang. It offers no allocation figures, however, and cannot separate NVIDIA supply effects from AMD supply effects.

Revenue does not disclose aggregate FP64 throughput. One possible mix explains both results: HPE may be selling more CPU-only systems for traditional modeling and simulation, while Dell may have more GPU allocations and accelerated machines outside generative-AI workloads. The source provides no mix data to test either claim.

Vendors below Dell and HPE generated $9.12 billion in first-half sales, up 13.8%, across dozens of OEMs and ODMs. That growth trails the market’s 19.9% rate, so Dell’s size is lifting the vendor average.

It remains too early to judge second-half retention. No complete 2026 vendor totals were available. The next test comes at the ISC27 Hyperion Research breakfast in Hamburg in June 2027.

How large is the accelerator fleet?

A back-of-envelope estimate projects about 1.2 million GPU accelerators in HPC and AI-for-science systems by end-2026. The estimate extends the 2017-2025 installation trend and assumes 2026 matches the pace seen from 2024 to 2025.

The estimated installed accelerator fleet comprises:

  • GPUs deployed worldwide: 18 million to 20 million.
  • Custom AI XPUs: 13 million to 17 million.
  • Google TPUs: 9 million to 11 million.
  • Amazon accelerators: about 3 million Trainium chips and 1 million to 2 million Inferentia chips.
  • Microsoft Maia and Meta MTIA accelerators: hundreds of thousands each.

These are rough estimates, not audited fleet counts. They exclude China’s custom XPU population. Together, they imply a worldwide fleet of 31 million to 37 million GPUs and XPUs, placing HPC and AI-for-science at only 3% to 4%.

The source puts HPC and AI-for-science spending in the tens of billions per year, including power and data-center construction. Training and inference spending for AI proper remains two to three orders of magnitude larger.

Will cloud spending displace owned systems?

The apparent cloud discount shrinks under a time-normalized comparison. The source says five years of cloud spending should be set against one year of on-premises spending. On that basis, it judges owned systems slightly larger. Both forecasts include component price increases passed through to customers.

The commercial question remains whether each simulation or model workload costs less on rented cloud capacity or an owned system. Hyperion expects HPC centers to keep using both. Accelerator availability and component inflation should therefore shape near-term pricing before cloud substitution becomes decisive, with hybrid procurement remaining the likely outcome.

The Next Platform (Source)

Filed under

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

News editor covering money, funding and market shifts around Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,238 articles

Read also

« Previous article