The Model Is a Commodity. The Soil Is Not.

“Compute is all that matters.”

Or some variation of this is what I often hear. The first to artificial general intelligence wins. The owners of the strongest models will be the winners. So who cares about anything else in the AI conversation? I keep coming back to what that argument leaves out: the billions of people whose languages and ways of life are not well represented in the datasets shaping the future of AI.

Capability Scales, Culture Compounds

Computing power wins when it comes to capability. I’m not going to argue that. If you try to build a local model to beat the strongest frontier models on a general task, you will lose.

But capability and culture have different metrics and incentives. The capability of a model, or what some call “intelligence”, scales with computing power. Culture does not automatically scale with computing power; it only compounds with data, context, values, memory, and proper governance of that data.

And this is what is important to understand: more computation on dominant data does not preserve an eclectic mix of cultures, but is incentivized to amplify the dominant culture.

Seeds, Greenhouses, and Soil

If each culture is a seed, then AI is the greenhouse that allows the seeds to grow and make an ecosystem that is strengthened by its different parts, but only if the seeds exist within this greenhouse. And a greenhouse is something you can buy. With enough money and compute, anyone can build one, and next year’s will be better and cheaper than the one you built this year. That is what makes the model a commodity. What no greenhouse can manufacture is soil. Soil is made slowly, in one place, out of everything that lived there before it, and a seed dropped onto a concrete floor under perfect light still grows nothing.

Most of the internet has skewed toward English, understandably, with the rise of business on the internet. We must also remember that many languages and cultures have slowly been forgotten or have not had the same amount of exposure, be it a Welsh folklore tale, a Japanese village matsuri, or a Polynesian voyage song. Those are seeds with very little soil left underneath them.

Frontier models are polished mirrors of the data they are trained on, and that data represents online, English, and Western texts. I’m not saying this is a bad thing inherently, but the consequence is that the weakest digital records will become flattened as time progresses. It’s not that the AI tells you it doesn’t know; with thin data, it fills the gaps and delivers the answer with the same confidence as fact, which makes it harder to understand what is true.

The Arguments I Keep Hearing

I’ve heard arguments like “Just let AI get really powerful and scrape the internet for the data or textbooks.” But the most vulnerable cultures are often not digitized completely, and their knowledge is often oral, local, or contextual.

“A stronger model will reconstruct a language.” We have seen models attempt to finish or decipher the broken parts of Roman tablets, but they do not recover missing ground truth if it wasn’t there in the first place — models interpolate from what they have seen. Anyone who has played around with photoshopping with AI for fun will understand.

“Just use retrieval or fine-tuning later when it gets better” is another argument I hear, but it proves the point. The model still needs trusted cultural sources to retrieve this information from, and that’s an important part of our discussion.

Sovereignty Is About the Source, Not the Model

So what should the future be without a winner-takes-all mentality? There shouldn’t be one model that magically knows every culture or holds 51% of the computing power. The future should be an AI that is capable but can read from trusted, living cultural sources. A country’s defensible asset is not its greenhouse; it is its soil, its consented, context-rich record of how people actually speak, joke, remember, and live.

This isn’t just about making AI more culturally aware, or cultural awareness as something the model performs. Cultural sovereignty is about who or what group owns the source material, and how they can fine-tune and trust it. Who decides what is correct and can update it, and who benefits when it is used. Who has the power to let it thrive and grow after all the voices that made the model have passed.

Tend the Soil

For builders, governments, universities, and cultural institutions, the implication is simple: do not compete only on the model. Build the soil, the source layer that the models will eventually need, and build systems that collect, validate, govern, and refresh this cultural knowledge. I think it’s safe to say it starts where culture naturally appears: language, conversation, correction, memory, teaching, or intergenerational transfer.

The model is a commodity. There will be a better one next year, and a cheaper one the year after, and eventually a good enough one running on the phone in your pocket. The soil is not. It is made once, slowly, by the people living on it, and once it washes away no amount of compute will put it back. So it’s our responsibility as humans to tend the ground each of our cultures grows in, so that the seeds stay ours and remain available to AI without being flattened by it.