CoreWeave: GPU wiring and power delivery can swing AI latency by orders of magnitude

On October 2, 2026, SiliconANGLE reported that CoreWeave said the way GPUs are wired and powered can cause AI inference latency to differ by orders of magnitude. The claim is attributed to CoreWeave itself rather than to an independent study, and it concerns physical infrastructure choices — cabling and power delivery — rather than chip models or software alone.

The report places that statement in the context of a shift in the neocloud market, a term it uses for providers such as CoreWeave that grew out of efforts to fill gaps in GPU supply. That market, according to the report, is moving past its origins as a stopgap for scarce graphics processing units. AI-native startups, it says, now choose their infrastructure on the basis of latency, burst capacity and openness, not just on whether chips are available.

CoreWeave is expanding beyond GPU compute into networking, storage and software, the report says, a move it frames as a response to growing inference demand. The report does not give specific latency figures, benchmark results or product and version names tied to the wiring-and-power claim, and it does not describe how any comparison was measured or which configurations were tested. What is public at this point, then, is CoreWeave's characterization of cabling and power delivery as factors that can shift inference latency by orders of magnitude, set against the report's account of a neocloud market where latency, burst capacity and openness increasingly shape customer decisions.