SaaS 没死：真正的护城河、主权智能与下一代算力平台

A large codebase is a weaker moat than before, but it was never the only source of defensibility for durable software companies.

Marketplaces, network effects, distribution, and deep operational relationships remain difficult for an AI-generated clone to reproduce.

Point solutions that sit directly in the model's path face more substitution risk than platforms embedded across workflows.

Companies can own the best model for a narrow task by training against proprietary data and specialized evaluations.

Moving reinforcement learning from simulations into safe online interaction is a major unsolved problem for real-world agents.

Compute planning has become a core executive discipline because demand, supply, and model economics can change by an order of magnitude.

Application companies are building internal research teams because intelligence is becoming part of the product rather than an external utility.

Owning weights, evaluations, and the data flywheel gives a company steerability and strategic control over its core product.

Falling inference prices can improve application margins while expanding total demand for model providers through Jevons-style effects.

Biological computing has demonstrated constrained token prediction, but it remains far from a general alternative to silicon.