A quick attempt to join the dots. A discussion today with the ever-reliable Dimitris Mavrakis at ABI Research (ostensibly about Ciena; coverage to follow) covered interesting ground. The deployment model for data centres is changing; where large training clusters have been concentrated in large venues, they are being broken into smaller regional centers. This will happen over the next few years, he says; the logic is to take advantage of energy availability – essentially to split power supplies between regions, just to be able to draw enough to power these monster AI factories. It’s not about sovereignty or latency; just about power. 

But these new clusters will need to perform as one, he says – so the connections between them will be highly available and reliable, and super fast. They need “thousands of fibres between them”, he says. Which is why Ciena is doing amazing, of course; which is why Nokia and Cisco are also doing well (for fibre optic cable). It is why Amphenol and CommScope are being strategic, as written below: when AI is distributed, the bottleneck is in the interconnect layer, and value moves down into cabling, connectors, fibre management.

The same logic extends inside the data centre – and increasingly, inside the rack. If clusters are decomposed geographically, they are also decomposed architecturally, with GPUs, memory and accelerators no longer assumed to sit on a single board or chassis. And as Mavrakis said, the old adage about the use of “copper if you can, fibre if you must” will no longer apply – because AI needs fibre, even to connect rack-level gadgetry. Which is why Marvell is buying XConn and Celestial AI, as we wrote last week – as a pincer movement in rack comms (UALink switching and/or ‘photonic’ light/optics, taken from each) to control near-chip interconnects. 

If AI performance is now defined as much by how fast data moves between processors, in rack or in region, as it is by the processors themselves, then interconnect IP becomes a strategic asset, not a component – and the winners in AI infrastructure will increasingly be those that control the fabric, not just the compute. Why is why… it is all about fibre now. Them’s your bets.