EXECUTIVE SUMMARY
This Intersect360 Research report presents an overview of the technologies and trends shaping the HPC-AI market in 2026, with a focus on Networking and Interconnects. Other technology segments are covered in further reports in the “State of the HPC-AI Market” series.
This report is based on surveys of members of the HPC-AI Leadership Organization (HALO). HALO is a global, end-user organization, facilitated by Intersect360 Research, that helps steer the course of the HPC-AI industry by identifying key issues, providing input into planned Intersect360 Research studies, and participating in surveys. For more information on HALO, visit www.hpcaileadership.org.
The Intersect360 Research “State of the HPC-AI Market” report series includes the following technology modules:
- Storage and Data Management
- Systems and Clusters
- Facilities, Power, and Cooling
- Interconnects and Networking
- Quantum Computing
- Processing Elements (CPUs, GPUs, Accelerators)
- Cloud Computing (Including On-Prem as-a-Service Models)
Each report in the “State of the HPC-AI Market” series contains three sections: 1) end-user survey data on installations, major trends, and “satisfaction gap” analysis to reveal features where buyers are currently not satisfied with available technologies relative to their importance; 2) Intersect360 Research analysis of these data and trends; 3) submitted content from invited top vendors, as determined by end-user surveys, responding to a fixed template of questions on target market, technology differentiation, and future vision.
EXECUTIVE SUMMARY
The networking capabilities that HPC-AI practitioners value most are the ones they are already most satisfied with. The market’s unmet needs lie elsewhere — and that inversion is the central finding of this Intersect360 Research survey on networking and interconnects. The survey draws on HALO, a vendor-excluded community of senior practitioners who design, procure, and operate these systems, rather than the vendors who sell to them. Asked to rate networking features on both importance and current satisfaction, members placed the classic performance measures — bandwidth, latency, and message rate — highest in importance, yet these showed the smallest gaps, because satisfaction with them is already high. The widest gaps fell instead on the work of making the fabric operate cleanly across mixed AI-and-HPC stacks.
Those gaps cluster, but not at a single point. They fall on interoperability across accelerators, compatibility with AI communication models, a unified fabric spanning compute, storage, and AI traffic, and container and orchestration support — and, notably for a non-technical attribute, on OEM and vendor support and service. That last finding is the most telling: as systems grow more heterogeneous, the relationship with the vendor, not the speed of the link, is where practitioners feel least served.
The installed base explains why performance is no longer the differentiator. InfiniBand remains the dominant primary interconnect (76.5% of respondents), with Ethernet most often in a secondary role (68.1%). Bandwidth is widely deployed but not universal: 41.2% of sites run 100 Gbps, 35.3% run 200 Gbps, and 33.3% run 400 Gbps, while 27.5% still report segments below 100 Gbps. With performance well-rated across the board, unmet need has moved up the stack – and AI even demands the same things of it. Most practitioners say they do not: only 37.3% agree the two workload classes share the same networking requirements.
The pattern is consistent and strategically clear. As practitioners add AI to simulation-first environments rather than rebuilding around them, the strain shows not in how fast the network runs, but in how well it integrates. This is a largely satisfied market with specific, targetable unmet needs — concentrated in integration and support — rather than one defined by broad dissatisfaction.
These findings reflect a senior, research-weighted community: 64.2% academic and not-for-profit research, alongside national research facilities and selected commercial sectors, with no hyperscale or cloud-service- provider respondents. They are a precise read on the priorities of HPC centers and research computing, not a projection of the hyperscale AI buildout.