The fabric layer consolidated twice in a year

Here is a question that sounds trivial and is not: who supplies the interconnect in the world’s large supercomputers, and how has that changed?

You cannot answer it from spec strings, because a spec string names a product. It says “Dual-rail Mellanox EDR Infiniband” or “Aries interconnect”: the thing that was shipped, under the name it was shipped as. It does not say who owns that business now.

What happened

Between September 2019 and April 2020, two acquisitions closed:

• HPE completed its acquisition of Cray in September 2019, having already absorbed SGI in 2016. Cray’s SeaStar, Gemini and Aries fabrics, and the Slingshot line that succeeded them, moved under HPE.
• NVIDIA completed its acquisition of Mellanox in April 2020. InfiniBand (which by then was the default fabric for essentially every non-Cray large system) moved under NVIDIA.

Intel had discontinued Omni-Path a few months earlier, in 2019; it was taken over by Cornelis Networks in 2020.

In under a year, the independent fabric business essentially ceased to exist as a category.

Why it matters more than it sounds

The consequence is not that a market got smaller. It is that two variables that used to be independent stopped being independent.

Before 2020, you could buy accelerators from one company and fabric from another, and analysts modelling concentration risk in the compute supply chain could treat those as separate exposures. After 2020, the two companies supplying most accelerator FLOPS also supply most fabric FLOPS.

Our interconnect share analysis shows what that looks like when you resolve the entities. Read the same data without resolving them and Summit’s dual-rail EDR belongs to a company that no longer exists independently, Titan’s Gemini belongs to another, and the concentration simply does not appear.

The entity resolution is the work

This is why our schema records the supplier at time of build on every component edge, and keeps a separate vendor alias table mapping raw vendor strings to the parent that owns them today, with an era range and its own source.

Both facts are load-bearing and neither may erase the other:

• Aggregate by today’s parent and you see the consolidation.
• Display the historical name and you keep the record honest. Titan was built by Cray. It was not built by HPE, which did not own Cray in 2012.

So the site shows “Cray (now HPE)” and rolls the FLOPS up to HPE. HPE’s company page currently shows 18 systems: 11 shipped under HPE, six under Cray, one under SGI.

Get this wrong in either direction and a thirty-year dataset produces a wrong chart. Treat Cray in 2012 and HPE in 2022 as unrelated and you understate consolidation. Silently rewrite the 2012 row as HPE and you lose the fact that an independent company built it.

The one that got away

Omni-Path is the counterexample worth keeping, and it complicates the tidy story.

Intel discontinued it in 2019. It did not die. Cornelis Networks took it over and continues it independently, which makes it the only case in this dataset of a fabric outliving its vendor’s interest in it.

Our parts table records Cornelis as the current supplier of record for Omni-Path, while the edges from Stampede2, SuperMUC-NG, and TSUBAME3.0 still correctly name Intel as the supplier at build. Same distinction, applied to a transfer rather than an acquisition.

What to watch

The live question is whether Ethernet takes the AI fabric market. Ultra Ethernet, RoCE and vendor-specific enhancements are all attempts to bring commodity economics to a layer that has been specialist for twenty years.

If that works, the consolidation described above matters less, because the fabric becomes something you buy from anyone. If it does not, the two companies that already supply the compute will also supply the network for the next generation of AI infrastructure, and the concentration in this chart is a floor rather than a peak.