The AI infrastructure rewrite cycle is the stable state, not a passing phase
Every 12 to 18 months, AI builders face a complete infrastructure overhaul. The switching costs they assumed were negligible are not, and the competitive moats they thought they were building evaporate before the next cycle completes.
Logan Kilpatrick puts the rewrite cycle plainly: every 12 to 18 months, everything has to be rewritten. That is not a complaint about tooling quality. It is a description of what paradigm-level shifts do to whatever was carefully built in the previous window. The cause, in Kilpatrick’s framing, is structural: “The model is sort of chaining with the harness, the harness is powering the agentic product experiences.” When the model and its surrounding infrastructure become a single integrated system rather than separable components, any meaningful model change propagates into the entire stack.
Martin Casado narrows the window further than Kilpatrick does. A given model, in Casado’s view, stays relevant for somewhere between three and nine months before something newer supersedes it. That estimate matters because it sets the cadence against which any investment in deep model-specific integration has to pay off. If the integration work takes a meaningful fraction of the relevance window, builders are perpetually behind.
The operational consequence of that cadence shows up concretely at Applied Intuition. Peter, a leader there, reports that the company has gone through four complete technology stack overhauls over roughly two years. That rate of change is not a sign of poor planning. It reflects a market in which the underlying capabilities are moving fast enough that whatever was state-of-the-art when a stack was designed is genuinely obsolete before the next cycle completes.
The model is sort of chaining with the harness, the harness is powering the agentic product experiences. Logan Kilpatrick
What makes this dynamic harder to manage is that the switching costs teams assumed were negligible are not. Yasser Elsaid pushes back directly on the “cost of switching between models is zero” argument. Fine-tuning how a product should behave around a specific model can absorb three to four months of work. On a nine-month relevance window, that is a third of the useful life of the integration, consumed in the act of building it. Teams are forced into expensive rewrites on a cadence that barely allows the previous integration to stabilize before the next one begins.
The business-model pressure this creates is not abstract. Graham Weaver frames it as a moat problem: an AI application startup may only be six months ahead of where the underlying models are heading, which allows for real near-term revenue but no durable competitive position. The Forbes AI 50 list offers a blunter version of the same point. Marc Andreessen notes that 40 percent of the companies on that list last year dropped off this year. Turnover at that rate does not suggest a maturing market. It suggests one where the structural advantage of having built early evaporates quickly as the foundation shifts beneath it.
What this adds up to is a compounding problem for anyone building on top of AI infrastructure rather than inside it. The rewrite cadence Kilpatrick describes is not a phase the industry will pass through on its way to stability. It is the stable state, at least for now. Teams that treat their current stack as a durable foundation are mispricing the risk. The question facing builders is not whether to accept the rewrite cycle but how to architect around it: which parts of the harness are genuinely model-agnostic and can survive a transition, and which are so tightly coupled to a specific model’s behavior that they will have to be rebuilt from scratch when the next paradigm shift arrives. The answer to that question, more than any individual model choice, determines whether the work compounds or simply resets.