$143M raised for a 200/400/800G-per-lane interconnect tells you where AI infra margins are leaking: not just GPUs, but how fast the rack can pretend to be one machine. For crypto compute markets like Akash, Render, io.net and Aethir, this is the hard ceiling on training-grade supply: distributed idle GPUs can clear inference and batch jobs, but frontier training still wants rack-scale NVLink-style locality, not random nodes stitched over WAN. If ASIC-over-dielectric cuts copper power/length pain without optical complexity, hyperscalers get denser clusters and DePIN has to compete on verifiable utilization niches instead of pretending every GPU is fungible.

Top comment by @Benthic