Segment
AI networking: wiring thousands of chips into one machine
Training a large model is a parallel computation in which every accelerator must repeatedly share its results with every other one. The network is not plumbing around that job — it is part of the machine, and when it is too slow the accelerators simply wait.
In one sentence
AI networking is the set of interconnects that join accelerators to each other — a short-reach, very high bandwidth fabric inside a rack, and a longer-reach switched fabric across a cluster.
Distributed training has a rhythm: every accelerator computes on its slice of the work, then all of them exchange and combine results before the next step can start. That collective exchange is a synchronisation point. Until the slowest participant finishes, nothing proceeds — which is why tail latency and consistency matter more here than average throughput.
Two very different networks answer two different halves of the problem. Inside a rack, a short-reach fabric joins a few dozen accelerators tightly enough that they can share memory; this is where bandwidth per link is highest and the technology is most proprietary. Across the cluster, a switched fabric connects thousands of nodes over metres to hundreds of metres, where optics rather than copper carry the signal.
The economics follow the distance. Copper is cheap, reliable and short. Optics reach further and cost more in money and power — enough that the number of optical links a design needs is a first-order cost question, not an accessory decision.
How this breaks down
Split by reach — what can be done in centimetres, in metres, and across a building.
- Switch siliconThe high-radix switch ASIC almost every AI cluster network is built around.Definition pageChokepoint
- SerDes and retimersThe analogue circuits that push tens of gigabits down a wire, and the chips that clean them up.Definition pageChokepoint
- Cables and connectorsPassive and active copper assemblies, backplanes and the connectors that terminate them.Definition page
- Scale-up fabricsThe short-reach, very high bandwidth links that create a single memory domain inside a rack.Definition page
- Cluster fabricsThe switched network across thousands of nodes, and why lossless behaviour matters more than raw speed.Definition pageChokepoint
- Optical interconnectTransceivers, DSPs and co-packaged optics — the components that carry cluster traffic over distance.Definition pageChokepoint
What this depends on
Technology dependencies are solved by engineering; supply dependencies are solved by building something, which takes years.
- Technology
High-speed SerDes
Every link in this branch rests on the circuits that push tens of gigabits down a wire or into a laser. It is among the harder analogue problems in the industry.
Semiconductor IP
What depends on this
Other pages in this map that name AI networking as something they cannot do without.
- Artificial intelligence · AI compute · AI memoryAI storageStorage bandwidth is delivered over the same fabric as everything else; a storage tier faster than its network is a storage tier that cannot be used.
- Artificial intelligence · Data layerData pipelinesBatches are streamed to every accelerator continuously through the run. If the front-end network cannot sustain that rate, the most expensive hardware in the building waits on data.
- Artificial intelligenceSoftware stackEverything above the kernel layer assumes gradients and activations can be exchanged between devices quickly. The libraries are written against the fabric's topology, so a change in one forces a change in the other.
Companies across AI networking
Every company named on a step below this page, ordered by how many of those steps it appears at. Compiled from the pages themselves rather than written separately, so the two cannot disagree. Not a ranking and not a recommendation.
- Huawei TechnologiesPrivate3 steps
- Cornelis NetworksPrivate2 steps
35 more companies appear at a single step each; they are named on the pages for those steps.
How these pages are written
Each page explains one technology in plain language, states what it depends on, and names companies by what they supply at that step. Company roles are described qualitatively and deliberately carry no market shares, revenue figures or rankings — those change faster than an explainer can, and a stale number is worse than none. Ticker links point at company pages on this site and are provided for reference only.
Nothing here is investment advice, a recommendation, or a forecast. A company named on a page about a technology is not thereby a good investment, and the chokepoints described are structural facts about supply chains rather than predictions about prices. Technology moves; where a page describes something as unresolved or in development, that was true when it was written.
Plutux no es un asesor de inversiones. Los datos de mercado y el análisis generado por IA son solo informativos y educativos, no asesoramiento de inversión. Aviso legal