Law II, Value Accrues at Bottlenecks

If Law I tells you who gets absorbed, Law II tells you where the money goes when they do. Durable value rarely sits in the model or the UI. It sits at whichever layer of the stack is structurally scarce, the layer that competitors cannot easily build, buy, or bypass. Find that layer. Own it. Everything above and below it pays you rent.

The model is not the bottleneck. The model is the commodity. This is the part most operators get backwards. They look at the AI stack, see that GPT-class models cost hundreds of millions to train, and conclude that L2, the model layer, must be where the money is. They are right that the layer is expensive. They are wrong that it is scarce. Three labs ship GPT-class models, two more are 18 months behind, and open-weight models track the frontier within a year. The model layer is rapidly commoditizing because it is rapidly being supplied. Commodities are not bottlenecks. Commodities are pass-throughs.

The bottleneck is whatever the model cannot do alone. NVIDIA owns L0 silicon, every model trains on their chips, every inference runs on their chips, and there is no second source at scale. That is a bottleneck, and the market prices it as one. Bloomberg owns L1b, forty years of structured financial data the models were not trained on and cannot license cheaply. That is a bottleneck. Vanta owns L3, the SOC 2 and ISO compliance gate that every B2B SaaS company has to pass to sell into enterprise. That is a bottleneck. Apollo and ZoomInfo own L1c, proprietary B2B contact graphs the public web does not contain. Bottleneck. Salesforce owns L4, the system of record that the workflow already runs through. Bottleneck.

The pattern is not 'own AI.' The pattern is 'own the layer everyone else needs to cross.' Once you own a bottleneck, you do not need to win on features. You win on the fact that the alternative is to rebuild your layer, which competitors cannot do quickly and the model layer will not bother to do at all. The model layer's incentive is to expand upward into surface and orchestration, not sideways into your trust gate or downward into your data well.

How do you find your bottleneck? Ask three questions. First, what does my product require that competitors cannot replicate within twelve months? If the answer is 'a better prompt,' you do not have a bottleneck. If the answer is 'eight years of customer behavioral data tied to outcomes,' you might. Second, what does my product require that the model layer is not incentivized to supply? Foundation model labs will not build vertical compliance, regulated-industry trust, or enterprise system-of-record integration. That is space the model layer leaves alone. Third, what would my customer have to rebuild if they left me? If the answer is 'a Stripe integration,' you do not have a bottleneck. If the answer is 'a decade of accumulated context that makes the product work,' you do.

Bottlenecks are layer-shaped, not feature-shaped. This is the second mistake operators make. They think the bottleneck is a feature, a clever workflow, a unique UI pattern, a better integration. Features get copied. Layers get owned. The bottleneck is structural: you own the data nobody else has, the trust nobody else can issue, the distribution nobody else controls, the memory nobody else accumulates. Features sit on top of bottlenecks. Bottlenecks sit underneath features. Confuse them and you build the wrong moat.

The corollary is uncomfortable for surface companies: if every layer below you is a commodity, you do not have a bottleneck. You have a brand. Brand is real value, but it is not structural value. It does not protect you from a bigger brand showing up with the same model, at the same price, with more distribution. Surface companies that survive Law I almost always do it by sprinting downward into a bottleneck layer before the model owner sprints upward into theirs.

Law II is the most actionable of the three. It tells you what to build, not what to fear. Identify the scarce layer in your category. Decide whether you can own it. If you can, own it relentlessly, every dollar of product investment should compound the bottleneck, not decorate the surface. If you cannot, do not build the company. The market is currently funding hundreds of products at surface layers where the bottleneck is already owned by an L4 platform or an L1b incumbent. Most of those funding rounds are buying time inside Law I, not escaping it.

Find the bottleneck. Own it. Everything else is rent, and rent gets repriced.