“The future is already here - it’s just not evenly distributed.”
— William Gibson

We are not watching the rise of Artificial Intelligence.
We are awatching the formation of a new aristrocracy.

Not of land.
Not of capital.

But of compute.


For almost two decades, we lived in the comforting lie that the problem of scarcity had been solved by software.

Code was free, distribution was free, and some kid in a dingy hostel could out-code some incumbent with nothing but time, internet, and stubbornness. The internet was flattening everything, and it felt as if gravity no longer existed.

But that era has ended.

Let’s stop the TED talk for a second, shall we? You are not witnessing the dawn of universal intelligence. You are witnessing the land grab.

The people shouting about AGI on podcasts are not the futurists that they claim to be; they are the trespassers on someone else’s land, mistaking access for ownership.


Compute is the new land, and the land has already been occupied.

AI has not simply added another layer to the technology stack; it has, in effect, introduced physics into the game. The frontier is not run on cleverness; it is run on fabs, supply chains, air conditioning, and power consumption curves that are indistinguishable from heavy industry itself. The look and feel of the whole operation are silicon, the logic is oil, and the constraint is no longer your idea; the constraint is whether you can procure 20,000 GPUs before someone else does, and whether you can keep them powered.

TSMC makes the chips that are the foundation of almost every serious AI operation on the planet, and if someone were to disrupt that, that is not a supply chain problem; that is the world stalling on its axis. NVIDIA does not sell hardware; they sell possibility, and the H100 is not so much a product as it is the border crossing; if you cannot cross that, you are not at the frontier; you are downstream, watching the summaries of the decisions that were made elsewhere.

Scarcity is once again physical, and once again not equally negotiable.

A single modern data center can draw as much power as a mid-sized city. In various regions of the United States, the utilities are secretly restructuring their grids to accommodate AI clusters while domestic projects languish. In Ireland, data centers have surpassed entire industries in power consumption. The conversation is now about whether homes or hyperscalers get priority on the grid. In Arizona, they siphon water to cool machines that produce words while drought warnings blanket the state.

This is not hyperbole. This is allocation.

The average founder will not train a frontier model. The average nation will not either. Training is an industrial process now. It is capital-intensive and infrastructure-constrained. It is geopolitically complex. Most will simply lease inference. They will pay by token, by call, by tier. Intelligence will be a utility, indistinguishable from electricity except for the fact that it fundamentally changes how we think instead of lighting our rooms.

Ownership is reduced to access.

And from this, a hierarchy starts to form.

At the top are those who own compute: hyperscalers, chip makers, infrastructure nations. Amazon and their AWS platform, Google and their TPUs and vertically integrated data centers, Microsoft and their Azure platform and their close relationship with OpenAI. They do not compete on features. They compete on megawatts, on access to fabs, and on geopolitical strength. They decide who grows and who waits.

Under them, the renters of intelligence. Startups, developers, companies whose product is built on top of those APIs they cannot control. They call it leverage. It works like a dependency. The price change ripples through their business overnight. The model update changes the behavior they never agreed to. The innovation they bring into the world is real. It is just not limitless. Not by imagination, but by rate limits.

And at the bottom, growing quietly, are the automation victims. They are not present at the negotiating table when the GPU contracts are signed. They are not on the blueprints of the infrastructure. They just feel it. The workflows are faster. The headcount is lower. The expectations outpace the retraining schedules. The explanation is always the same. Always delivered with a smile, but a smile that seems practiced. PROGRESS.

No one talks about displacement. The word has been refactored.

This is not conspiracy. This is industrial logic speaking through a new language.

Every product of AI you touch today passes through the same narrow corridors. The silicon was designed here. The silicon was manufactured elsewhere. It was assembled. It was shipped. It was installed in clusters accessible only to a handful of entities. Each one of them taking a margin. By the time it reaches you, it has paid rent several times over.

The system doesn’t conceal this fact. It normalizes it.

We speak of openness, but open weights do not construct data centers. You may download a model, but you may not download the circumstances that made it relevant. Without compute at scale, openness is a performance: engaging, well-produced, structurally irrelevant.

Meanwhile, the map reconfigures itself.

Export controls are strengthened, and research infrastructures grind to a halt. A package of advanced computing chips is denied, and a nation’s aspirations for AI development quietly readjust downwards. Data centers are negotiated like treaties. Energy agreements start to resemble strategic partnerships. The dialogue changes from “what may we create?” to “what may we execute?”

Compute does not apportion itself equally. It accumulates. It solidifies around capital, geography, and power.

And in this accumulation, the form becomes familiar.

You will not own the model that makes you efficient. You will subscribe to it.