One company has just ordered a building block the rest of the industry cannot build without. Nebius, a European cloud provider, has told investors it plans to spend about $40 billion on cloud infrastructure over the coming years, according to TechCrunch (Nebius tests the neocloud playbook). The money would flow into data centers and the computer equipment inside them, the physical layer on which every large language model and every application now depends.
If the plan holds, Nebius becomes one of the biggest compute spenders in the world. That figure raises an unglamorous question: what is one bottleneck resource worth when every model and every application depends on it? It is what a bottleneck is worth: the price the last buyer pays.
The answer depends on which side of the argument you are on. Nebius’s spending concentrates capital on the one resource everything else needs. Hugging Face’s vow keeps the open-source side focused on the same resource’s distribution. The two positions differ on who should set the pace of capability. The two sides are really arguing about who owns the next decade of computing capacity.
The Two Sides of One Market
Nebius’s spend matches the concentrated side of the story. It pours capital into a single resource and bets that whoever controls the most of it sets the pace. Hugging Face’s vow belongs to the distributed side. Its chief executive has said he will keep fighting for open-source AI, even as the industry around it consolidates.
TechCrunch reports the vow; Hugging Face CEO vows to keep fighting. The two positions differ on who should set the pace of capability: the builders inside the walls, or the builders outside them.
Both sides are fighting for access to the same input. The neoclouds need computing power. The open-source projects need the hardware and the engineering talent to keep their models competitive. They disagree about whether capability rises from concentration or from distribution. The disagreement is about the future of the market, not about the models themselves.
Two readings of the same market explain the fight. The first says that capability is a prize handed to the payer. The second says that capability is a practice that survives once enough people are using it. Each reading is a bet about the future of computing. Both sides buy into the same market; they simply disagree about who is destined to lead it.
But the market they are fighting over is not the market for models. It is the market for the single bottleneck input, the one thing every other product must buy. The models are the products; the resource is the commodity. That is the argument in miniature, stated once, plainly.
The Single Bottleneck
The common assumption is that the buildout is about producing better models, or about getting more compute for the same money. Both readings miss the real structure. The actual structure is that one input controls everything the rest depends on. The assumption that more compute means better models ignores who controls the supply.
Every model architecture, every application layer, every startup trying to ship an AI product must buy its way into this one market. The neoclouds and the open-source projects are not at war with each other; each is fighting for access to the same scarce input. The fight only looks like a models race because that is what everyone can see.
The mechanism is older than the current boom. Sir Lawrence Freedman, in Strategy, describes how concentrated spending on a single resource sets the pace competitors must match. He records that “cutting costs and product improvements could be easily emulated and so left the relative competitive position unchanged.” Outperformance, in his account, required a difference that could be preserved.
Freedman’s book is about navigating the gap between intention and reality, where plans collide with will, friction, and opposing minds. Strategy, in his account, is the art of navigating that gap. Leaders must think in levels, separating tactics from strategy. Tactics ask whether a plan will survive a single month. Strategy asks whether the plan is worth choosing at all. They may wield narrative and deception as legitimate instruments. They must treat every doctrine as provisional rather than timeless.
The AI industry is navigating that gap too. To spend $40 billion on one input is to make a strategic choice, not merely a financial one. The bet is that buying the resource closes the gap between what builders intend and what reality delivers. It makes strategy a matter of balance sheets instead of doctrine. The rest of the industry then has no choice but to follow the spender’s pace, because its own plans depend on the input the spender controls.
The Cheaper Input That Consumes More
The planned $40 billion is a floor, not a ceiling. The spending first pays for construction and equipment. Once the data centers are running and the machines are cheaper to deploy, the industry buys more of them, and the total consumption of the input rises.
George Reisman’s Production Versus Consumption makes the same point in a different field. Its central claim is that consumptionism’s fatal error is treating desire as the engine of prosperity rather than production. Reisman argues that production, not stimulated desire, is what raises living standards. Productionism names the real engine of prosperity: capital accumulation. The error is to treat the appetite as the engine. The engine is the work done to satisfy the appetite. A growing stock of tools, machines, and buildings lifts how much a working day can produce.
The book’s example is concrete. “A given number of brick houses requires a larger supply of capital goods than the same number of wooden houses of the same size,” Reisman writes. The material and the machine choices determine how much capital the industry must stock, not how many customers want the end product. The example is a comparison of inputs, not a prediction about sales. The physical structure of production sets the capital stock; desire does not.
Map those parts onto the AI industry and you get the same arithmetic. The $40 billion is that stock. Once the compute is cheap, the bottleneck moves. It stops being a wall and becomes a toll. And once it is deployed, the cheaper compute it contains gets consumed in larger total quantities.
The open-source projects that could not previously afford the bottleneck are pulled into the same market as the neoclouds. Cheaper compute does not just lower the cost of building models; it widens the circle of people who can buy into it. The fight over the single input stops looking like a models race and starts looking like a competition for access to a scarce resource.
A Market That Keeps Paying
The open-source model survives only as long as the closed walls keep paying for shared standards. The open-source projects that can no longer afford the bottleneck will keep their weights, the learned settings of a model, available only inside the very companies that price the resource.
Chris Dixon’s Read Write Own describes the deeper rule. Power accrues to whoever owns the network layer, not the tool built on top of it. Open protocols like DNS and HTTP created durable user sovereignty because they were shared standards anyone could build on; corporate platforms extract value from them because they own the pipes. A protocol is a set of shared rules, a tool is whatever you build on those rules, and the owner of the rules keeps the value.
Blockchains, Dixon argues, are a third architecture: programmable, token-incentivized networks where ownership, governance, and reusability are encoded into the infrastructure itself. “Other developers can then take those components and reuse them, without either party having to understand the details of what the other is doing.” The bricks just snap into place. In a blockchain, the protocol itself distributes a token that rewards builders. Those tokens make the network worth running.
The open weights in Hugging Face are the same kind of shared standard. They are worth something because the neoclouds keep building on them. The walls that keep paying for the compute are the walls that keep the standards alive. Hugging Face’s vow to fight for open-source AI therefore looks less like a moral position than a business problem: the standards are only as strong as the capital that sustains them. The standard is not free. Someone pays for it, and the payment determines its worth.
If that concentration holds, the open-source world does not disappear; it gets absorbed. By the early 2030s the distributed ideal might be a service tier priced by carbon and capital inside the closed walls, rather than a separate channel for model building. The neoclouds would keep pouring money into the single bottleneck, and the open-source projects would survive as priced components within the very companies that control it. The likelier outcome is that the boundary between open-source and closed shifts, and what had been a separate channel becomes a paid tier inside the dominant one.
Close on the Single Resource
The open-source fight and the neocloud buildout both come down to who controls the concentrated supply that every model and every application depends on. Nebius pours capital into the bottleneck, betting that whoever controls the most of it sets the pace. Hugging Face fights to keep the standards open, betting that distribution and shared infrastructure can outlast any single builder. The market’s answer is that control follows the concentration of capital. The open-source ideal gets priced into a service tier, and the neoclouds get to set the pace. The race runs where the capital runs.
Sources
Dixon, Chris. Read Write Own.
Freedman, Sir Lawrence. Strategy.
“Hugging Face CEO vows to keep fighting for open-source AI after Jensen Huang’s exit.” techcrunch.com.
“Nebius tests the neocloud playbook with $40 billion in planned spending.” techcrunch.com.
Reisman, George. Production Versus Consumption.
Leave a Reply