# The model will never be your advantage > For equal performance, the cost of querying an AI model falls tenfold a year. Value lies elsewhere: in the context only you hold, once written down. Article · By Élie Sader, Network Effect · Published on November 4, 2026 You are torn between two artificial intelligence (AI) providers, you read that a new model has just overtaken another, and you wonder whether it would be wiser to wait for the best one before committing to anything. Turn the question around. The model you use costs far less than it did a year ago, for the same quality, and models look more and more alike. Something that gets cheaper every year and that anyone can buy has a name in economics: a commodity. And a commodity has never been anyone's advantage. The real question lies elsewhere: if the building block is worth less and less, where does the value go, and how much of it can a small or mid-sized company hold? > The advantage is never in the building block. It is in what you build with it. ## Why do people say the model is becoming a commodity? This is not an opinion. It is three facts. **Models are converging.** A model is the program that understands a request and answers it. On well-defined tasks, summarising, drafting, sorting, translating, the best ones are on a par, and every lead is caught up within months. **The price is collapsing.** At the end of 2024, a study by an American venture fund found that, for equal performance, the cost of querying a model had been falling tenfold every year since 2021. The author admitted the method was imperfect and the pace might slow; the direction is not in doubt. Over time, the price of an answer tends towards the price of the energy it uses. **Access is widening.** Open models, which anyone can download, now run on a desktop machine, or even on a server the size of a book. What used to be an intellectual scarcity, writing, analysing, designing, is becoming an everyday resource, like electricity. One image sums it up: you make a living building refrigerators; you make a fortune with what you put in them. AI models are the refrigerators. ## If the building block is worth less, where does the value go? It does not vanish; it moves. The history of infrastructure always gives the same answer. The internet's protocols are open, and nobody grew rich by owning them; the wealth was made on top, in what people built with them. Never in the infrastructure, always in the layer that uses it to solve someone's problem. One caveat, to avoid caricature. Commoditisation holds for well-defined, standardised tasks. For tasks that depend on context, preparing a proposal for a given customer, choosing between two suppliers, drafting an answer that reflects twenty years of shared history, interchangeability breaks down. Two companies using the same model do not get the same thing out of it. The one that wins is the one that understands its customer's problem better than anyone, and knows how to tell the machine. So value moves towards three things the model does not have: the company's own data and context; an understanding of the problem to be solved; and the way you tell the machine exactly what you expect of it. ## Won't the big providers take everything? They are trying. AI labs launch one product after another: tools for presentations, spreadsheets, agents, browsers. You could read that as strength. It is rather a sign that they are looking elsewhere for the rent the model no longer gives them: selling answers whose margin shrinks every year will never justify their valuations. So they bet on distribution, the customers they already have, and on the cost of leaving them. The mechanism is well known: an average product that is already installed beats an excellent one you would have to learn, and each additional integration makes leaving more expensive. But a great product comes from years of obsession with a trade, not from a few months of commercial rush, and a generic tool ends up disappointing those whose trade is not generic. For a small company, then, the conclusion is not "you are going to be eaten". It is: do not play on their ground. The question is not "how do I do as well as they do"; that is impossible. It is: "what value can I create that no generic offer will ever reach?" ## What does a small company own that nobody can buy? Its context. You are not behind on AI: you hold a context nobody else can acquire. The local market and its balance of power. The history of each customer relationship, which is not in the contract. What customers do not say, and you understand anyway. Feedback from the field, the rules of thumb of the trade, the reasons behind past decisions. That is your intangible capital: your knowledge of your customers, your craft, your processes, your weak signals, your dealings with the real world. The model is the commodity that reads it; the capital is the asset. We describe it layer by layer in [Intangible capital](https://networkeffect.fr/en/intangible-capital/). As models become commonplace, the nature of the advantage changes. It is no longer "they can do what we cannot", it is "they hold the context and the understanding of the problem that we cannot buy". And the skill that matters changes with it: knowing how to tell the machine exactly what you want, for which problem, in which context. Specifying well replaces executing well. The strongest objection follows from this. Scarce context is not a fortress; it is a lead with an expiry date. Context that stays in people's heads evaporates when they leave, and a better-equipped competitor can eventually catch up. It only becomes an advantage once it is written down, structured, readable by every generation of models, and alive, meaning kept up to date. That is the difference between an advantage and a habit. A last point: a small company is not a software publisher. What it puts in the refrigerator is not an AI product to sell; it is its own trade, served better than anyone else's, with rented models. ## What does Bitcoin say about what stays scarce? Bitcoin was designed around a question that AI makes urgent: what stays scarce when everything can be copied? The computer scientist Nick Szabo gave the answer a name: unforgeable costliness, a cost that cannot be faked and that the recipient can check without trusting the producer. He also tells what happens when that cost becomes forgeable. Wampum, beads made from shells, became legal tender in New England between 1637 and 1661; once Europeans learned to harvest and manufacture it in bulk, its value was inflated a hundredfold, and it stopped being money. Text, code and generic analysis have just gone through the same thing: they can now be produced at almost no cost. What keeps its value is what cannot be forged: twenty years of fieldwork, customers, decisions and the reasons behind them. Expensive to build, quick for a model to read. Bitcoin is the first purely digital object to carry such a cost: digitised capital, scarce and verifiable, held by keys. That is the meaning of our thesis: [AI makes everything abundant; Bitcoin stays scarce](https://networkeffect.fr/en/bitcoin-ai/). The second lesson is about rent. No great disruption has ever abolished it; each one moved it to the next limiting factor, from land to factories, from factories to finance. By making production almost free, AI does not abolish scarcity; it moves it. For the labs, the next bottleneck is distribution. For a small company, it is its context. ## What should I do now, before the market settles? Three decisions. **Never bet on the building block.** No model of your own, no waiting for "the best model", no project that only makes sense with one particular provider. Today's best model will be overtaken within months, and cheaper next year. **Treat the model as a commodity you replace.** The best value for money of the moment, local when possible, from a provider under contract otherwise, and always with the ability to switch. That is what we set up with [Infrastructure you own](https://networkeffect.fr/en/services/infrastructure/). **Put the time and money into your own ground.** Write down the context, structure the data, write the instructions, and keep them alive: that is the work of [CORPUS](https://networkeffect.fr/en/services/corpus/). Timing matters. Models are not yet fully interchangeable, and this is when the gap opens. The company that has structured its ground while others were comparing models will be on the right side once models are interchangeable; the others will remain dependent on whatever the standardised market is willing to offer them. ## What nobody will ever rent You will never have the best model for more than a few months, and it does not matter: it is for rent, to everyone, cheaper every year. What you can have, and nobody will ever rent, is your own ground, provided you have written it down. To write down and structure that ground: [CORPUS](https://networkeffect.fr/en/services/corpus/). To run the commodity in-house, and be able to replace it: [Infrastructure you own](https://networkeffect.fr/en/services/infrastructure/). And for the thread that ties the two together: [Intangible capital](https://networkeffect.fr/en/intangible-capital/). ## Sources 1. Guido Appenzeller (Andreessen Horowitz), "Welcome to LLMflation: LLM inference cost is going down fast", 12 November 2024: [a16z.com](https://a16z.com/llmflation-llm-inference-cost/). 2. Nick Szabo, "Shelling Out: The Origins of Money", 2002: [nakamotoinstitute.org](https://nakamotoinstitute.org/library/shelling-out/). ## Further reading - [Tools you own](https://networkeffect.fr/en/services/infrastructure/) - [Renting your AI: you have the use of it, not the ownership](https://networkeffect.fr/en/renting-your-ai/) - [All resources](https://networkeffect.fr/en/resources/)