As AI clusters swell to thousands of chips, a quieter race is heating up over how those chips talk to each other. The connections between accelerators are becoming as important as the accelerators themselves, and the industry is pushing past standard networking to keep up.
According to Semiconductor Engineering, one effort centers on a "production-ready" optically connected rack — using light rather than copper — designed to deliver the bandwidth density and efficiency needed to scale AI compute clusters to 1,000 accelerators.
Meanwhile, the established Ethernet standard is being reworked for the AI era. The Next Platform reports that OpenAI, Microsoft and other partners are building "a better, more scalable Ethernet" — a sign that some of the biggest names in AI want the open networking standard, rather than proprietary alternatives, to carry their workloads.
The Next Platform also covers how networking specialist Arista is riding demand for AI "scale-out" networks, expanding into what it calls "scale-across" connections, and waiting to move into "scale-up" — the tightly linked clusters of chips that sit closest together.
A newer entrant is in the mix too. According to The Next Platform, a company called Eridu is "cutting to the AI networking chase" with a high-radix switch system — switching gear built to connect large numbers of devices efficiently.
Together, the reports point to a single theme: the plumbing that links AI chips is now a competitive battleground, drawing in cloud giants, networking incumbents and startups alike.
Why it matters: how fast and efficiently AI chips can be wired together increasingly determines how big and capable AI systems can get — and who controls that technology shapes the cost and pace of the entire AI build-out.