· Compute Atlas
The merchant fabric tier is one company
Of the 21 interconnect edges in our dataset where the fabric came from someone other than the machine's integrator, 14 belong to NVIDIA. Two announcements this week land on exactly that segment.
Four days ago we wrote that the fabric layer consolidated twice and ended on an open question: whether commodity Ethernet would turn the interconnect into something you buy from anyone, or whether the two companies supplying the compute would end up supplying the network too.
Two items landed this week that touch that question, so we went back to the dataset to see what “buy it from anyone” has actually meant so far. The answer is narrower than we expected.
Splitting the edges
Our system_components table records, for each system, the interconnect it was built with and the company that integrated the machine. Comparing those two fields sorts every fabric edge into one of two piles: the integrator supplied its own fabric, or it bought one.
We hold 56 systems with a recorded interconnect edge. Of those:
- 35 carry a fabric from the company that built the machine. Cray’s SeaStar, Gemini and Aries, HPE’s Slingshot, Fujitsu’s Tofu, IBM’s BlueGene torus lines, the NUDT and Sunway fabrics. These are captive: the fabric is part of the machine’s identity and was never sold separately.
- 21 carry a fabric bought from someone else. This is the merchant tier: the entire market for a fabric you can put in a machine you did not design.
Then the number that made us write this up. Of those 21 merchant edges, resolved to current owners:
| Owner | Merchant edges |
|---|---|
| NVIDIA | 14 |
| Cornelis | 5 |
| Quadrics | 1 |
| Arista | 1 |
The merchant fabric tier in this dataset is InfiniBand, plus Omni-Path as a legacy tail, plus two single rows. There is no third supplier. “Buy your fabric from anyone” has in practice meant “buy it from NVIDIA,” and it has meant that for every non-captive system we hold since Summit and Sierra.
What landed this week
Both of this week’s items push on that same segment, from opposite directions.
NVIDIA said it will extend NVLink Fusion to d-Matrix’s inference silicon, connecting a third party’s accelerators over the rack fabric that until now bound only NVIDIA’s own. NVLink enters our parts taxonomy as a captive interconnect, the thing that makes an NVL72 one machine rather than 72. Selling it to someone else’s XPU moves it toward the merchant column.
Read against the table above, that is not a new entrant. It is the company that already holds two-thirds of the merchant tier extending from the network layer into the coherent rack fabric, which is the one part of the stack that was still captive by construction.
The second item runs the other way. Panmnesia and Meta published an architecture proposal for holding a cache-coherent CXL domain across a whole facility rather than a rack, the same cross-rack segment, but built on an open multi-vendor standard rather than one vendor’s link. It is a review article with claimed silicon and no deployment, so it changes nothing in the graph today. It is worth recording because it is the only vendor-neutral proposal we have seen aimed squarely at the segment NVLink Fusion is proprietary in.
How these numbers could be wrong
The split depends entirely on entity resolution, which is a choice we made. The table above rolls Mellanox up into NVIDIA and Omni-Path up into Cornelis. Under build-time names the same 21 edges read: NVIDIA 10, Intel 5, Mellanox 4, Quadrics 1, Arista 1, and the concentration story weakens considerably, because four of NVIDIA’s fourteen were bought from an independent Mellanox before April 2020. Both framings are in the data; we think the parent-resolved one answers “who supplies this today” and the build-time one answers “who supplied this then.” Neither is the true one. This is the same distinction the earlier post is about, and it is doing real work here.
These are edges in our dataset, not the market. We hold 56 systems. That is a curated set weighted toward large and historically notable machines, not a census, so treat 14-of-21 as a description of what we have recorded rather than a share estimate. The interconnect share analysis is the page with the full breakdown and its own caveats.
An edge is not a machine and not a FLOP. Meta’s GenAI clusters contribute two of the 21 rows, one RoCE and one InfiniBand, because the operator built one of each. Counting edges gives that operator double weight; counting systems or installed FLOPS would give different totals.
The captive/merchant test is cruder than the market. It asks only whether the fabric’s owner and the integrator are the same company. It cannot see a fabric that was technically purchasable but never bought by anyone else, which is precisely the shape of Eviden’s BXI: carried in our parts table, present in none of the systems we hold.
Why we are recording it now
Neither announcement changes a row in the graph. No system is built on NVLink Fusion, and the CXL work is a paper.
What they change is which question is live. The earlier post framed the open question as Ethernet versus InfiniBand: commodity economics arriving at a specialist layer. Both items this week suggest the contested segment has moved up: not the network between racks, but the coherent domain that decides how large a single machine can be. One proposal makes that domain a licensed extension of an existing vendor’s link. The other makes it an open standard.
If the licensed path is the one that ships, the merchant tier stays one company and gains a layer. That is a claim about the future and we are not making it. The claim we are making is that the table above is the baseline you would measure such a shift against, and every row in it carries its own source.