I have gone down this path myself for doing 8k editing. TB3 attached SSDs IO to my Mac at about 3GB/s. The ones on my server connected over 10GB fiber Ethernet actually only reach about 800MB/s and I suspect that that macOS networking stack is just not at all optimized for 10G.
10GBit networking is really only 1.25GBytes per second, so 800MBytes/s isn't saturating the link, but is 64% of the way there. TB3 has a theoretical throughput of 40GByte/s so 3GByte/s => 24GBit/s is 60%. Realistically both are lower to the theretical link performance than I would have guessed, so there may be some bottlenecks involved beyond just computational overhead, but it makes sense TB3 was going to win assuming the storage had the bandwidth.
