Most AI inference bottlenecks aren't caused by lack of compute, they're caused by inefficient data movement. Here's how dataflow architecture changes that.