The next wave of AI products may be won on latency and cost rather than benchmarks