Summary

The infrastructure part went below the model and then above it. It began with accelerator bandwidth, then the software substrate in between: autodiff and the frameworks that industrialized it, and the compiler and kernel layer where bytes moved, not FLOPs performed, set the price and one vendor's software sediment forms the industry's deepest moat. From there it covered cluster orchestration, data infrastructure, silicon, power, and failure at scale. Then it returned to the frontier itself, where data, learning signal, measurement, and verification decide whether more compute still turns into accepted capability.

The real limit is rarely the one on the model card. It may be HBM, a network tier, a power interconnect, an export rule, a checkpoint system, a failure distribution, a depleted data source, a saturated benchmark, or a claim that is cheaper to generate than to verify. The frontier is a moving set of physical, operational, measurement, and acceptance constraints.

The open question is no longer only whether capability acceleration is durable. It is also whether proof, replication, oversight, and operating contracts can scale quickly enough to keep accepted knowledge from becoming the bottleneck. That is why the next part can turn to economics without leaving the technical argument; markets form around these constraints.