Law I, Intelligence Commoditizes Downward

Every wave of AI startups runs the same play. A new model lands, GPT-3, GPT-4, Claude 3, Gemini, the next one, and within ninety days a thousand companies wrap a prompt around it, polish a UI, ship a Stripe page, and call themselves a product. For a year the metrics look like a generational software business. ARR doubles every quarter. Boards funnel money in. Then the model layer underneath them ships the same feature for free, and the entire category compresses overnight.

This is not a market accident. It is a structural force. Law I of Supply Chain of Intelligence states it plainly: if your product depends only on generic model capability, the platform layer below you will eventually absorb it. The model owners are not standing still. They have every incentive to move up the stack, because that is where the margin is, and every capability to do so, because they already own the substrate everyone else is renting.

The clearest case study is Jasper. At its October 2022 peak Jasper raised $125M at a $1.5B valuation as the canonical 'GPT wrapper for marketers.' Defensibility lived at L7, the surface, and only at L7: prompt templates, brand voice presets, a polished editor. Then ChatGPT launched. Free. Conversational. The same underlying GPT-3.5. Within six weeks the entire premise of Jasper was a feature inside a product the user had no reason to pay for. By 2024 the company was reportedly trading at roughly $300M. An 80% mark-down. Not because the team got worse. Because the layer they owned got absorbed by the layer below it.

Why does this keep happening? Because the AI stack has a gravity. Value flows downward through it the way water flows downhill. Anything that can be done at L2, inside the model itself, eventually will be done at L2, because the model owner controls the marginal cost. If GPT-5 can write marketing copy in 'Jasper voice' with the same quality, OpenAI will ship that capability, and they will ship it bundled into ChatGPT Plus at $20/month. The wrapper does not lose on feature parity. The wrapper loses on price floor. You cannot charge $59/month for something that costs the layer beneath you nothing to include.

The pattern repeats across categories. Chegg sat at L7, generic educational content, and lost 99% of its market cap when ChatGPT made the same Q&A free. Stack Overflow's traffic compressed when models absorbed the answers their community had volunteered for fifteen years. Presentation-generation startups like Gamma now sit one Microsoft demo away from being a free PowerPoint feature. Every category whose entire moat lives at the surface is on the same clock.

The first law tells you who gets absorbed. It does not tell you who survives. That requires the next two laws. But Law I gives you the diagnostic: ask which layers your product owns that the model layer below it does not. If the honest answer is 'none, we own the prompt and the UI,' you are not a company. You are a feature waiting for its acquirer, and the acquirer will not bother to acquire you. They will simply ship the feature.

The escape, when it exists, is to own a layer the platform structurally cannot. Proprietary data the model was not trained on (L1). A trust gate the platform cannot legally cross (L3). A distribution surface the platform does not own (L4). An execution depth that requires years of workflow embedding (L5–L6). A memory of the user that compounds over time and would be painful to migrate (L8). One of these, owned with conviction, beats five rented at the surface.

If you are building today, the question is not 'is my product useful?' Useful products get absorbed every quarter. The question is: 'when the model layer below me ships my feature for free, what is the user still paying me for?' If you cannot name it in one sentence, you are inside Law I. You have time, but not as much as you think.

Observation (corollary of Law I). We are not entering an era where AI replaces creators. We are entering an era where taste becomes the moat, because once generation collapses into L2, the scarce input is no longer *who can produce* but *who can choose*. Taste is not a single layer; it is an L1c (behavioral history) + L5b (curation playbook) + L8b/c/d (compounding profile of what worked) package. That stack is where the value Law I displaces eventually lands.