We are entering a new phase of the machine age. To understand where we are going, we have to look at where we’ve been.
In the early days of the web, power was pushed to the edges. Tools like Java and Ruby on Rails allowed anyone with a laptop to build and ship global software. The cost of participation was a few hundred dollars and an internet connection. This was the era of sovereignty.
Our current era-AI 1.0-is the exact opposite. It is a return to the Mainframe. Instead of owning our tools, we are renting intelligence from a handful of massive data centers. We have traded the freedom of the early web for the convenience of an API.
This is Techno-Feudalism. The lords own the land-the GPUs, the weights, the data pipelines-and we are digital sharecroppers, paying rent with every token.
The Infrastructure Squeeze
The feudal lords didn’t seize power through superior technology. They seized it through scarcity.
We are living through a GPU squeeze. Demand for AI compute has outstripped supply, and the shortage has concentrated power in the hands of those who already owned the hardware: AWS, Google, Microsoft, OpenAI. If you want to train a model, fine-tune at scale, or run serious inference, you go to them. There is no alternative. The shortage is the moat.
This is the mechanism of feudal control. It’s not that the cloud providers are smarter or more innovative-it’s that they had the capital to hoard GPUs before the gold rush, and now they control access to the means of production. You cannot compete with someone who owns the only mine in town.
The squeeze creates dependency. Dependency creates lock-in. Lock-in creates rent extraction. Every startup building on OpenAI’s API, every enterprise running inference on AWS, is a vassal paying tribute to a lord whose only real advantage is that they got there first with enough money.
The Feudal Economics
Once the lords control the infrastructure, the rent extraction follows a predictable pattern: separate software from hardware, then charge for the software as if it were scarce.
Consider a typical cloud AI product. The hardware costs are real-GPUs consume power, take up space, require cooling. But the software layer sitting on top? Zero marginal cost. Every “feature” locked behind an enterprise tier that doesn’t require additional hardware to run is pure rent extraction. They are taking something that costs nothing to duplicate and creating artificial scarcity.
This is the pattern: give people enough free capability to create dependency, then extract rent once they’re locked in. It’s the same model as feudal agriculture-you work the land, but you never own it, and the lord can change the terms whenever he likes.
A change in terms of service. A “safety” update. A price hike. You have no recourse because you never owned anything to begin with.
The Great Filter: From Squeeze to Glut
The GPU squeeze cannot last forever. Scarcity is the source of feudal power, and the lords are destroying it themselves.
Greed is the engine. During the squeeze, every major player can use all the capacity they can get-demand is real, the money is flowing, and no one wants to be caught short. So everyone builds. Billions pour into GPU clusters. Nvidia sells shovels at gold-rush prices. Meanwhile, researchers publish efficiency breakthroughs weekly-quantization, distillation, sparse attention-each one reducing the compute required to achieve the same results.
This is the trap. Each player is acting rationally in the moment, but collectively they are building toward a cliff. The squeeze incentivizes simultaneous over-building and over-optimization. One day the music stops, and suddenly there is more compute than anyone knows what to do with.
The same greed that created the feudal concentration is building the infrastructure that will destroy it.
We don’t know exactly when the bubble bursts. Markets can stay irrational longer than you can stay solvent betting against them. Every short seller learns this the hard way: you can be right about the fundamentals and still go broke waiting for the correction. What looks like months away often takes years. So don’t bet on timing. But the destination is certain. When the burst comes, the squeeze inverts. Overcapacity floods the market. Efficiency gains multiply the effective supply. The lords find themselves sitting on depreciating assets, competing for customers instead of rationing access.
Just as the 2000 dot-com crash left thousands of miles of “dark fiber” that made the modern internet possible, the AI crash will leave behind a GPU glut. And suddenly, the moat disappears.
This is the birth of AI 2.0.
AI 2.0: The Architecture of Sovereignty
AI 2.0 is not about better models. It’s about who owns them-and who has options.
It is the shift from renting intelligence to owning the means of deduction. But sovereignty doesn’t require running everything yourself. It requires leverage. The glut creates leverage by turning the feudal lords into commodity providers competing on price.
You can start building this way now, but don’t expect enthusiasm from American investors. They are optimized for the squeeze. They want to fund platforms with network effects and lock-in-the exact characteristics of feudal systems. The AI 2.0 strategy is about independence, which makes it a poor fit for the VC model that profits from dependency.
To survive the transition, focus on three pillars:
1. Own the Weights (The Mind)
If you rely on a closed API, you can be evicted at any time. Sovereignty requires running open weights models. If you can’t download it and run it on your own hardware, on your own terms, you don’t own it.
This doesn’t mean open weights are “as good” as the frontier closed models today. It means the gap is closing, and for most practical applications, open weights are already sufficient. The 10% of capability you sacrifice buys you 100% of control.
2. Own the Compute (The Body)
Sovereignty requires control over your inference. In AI 1.0, that meant begging for access to centralized mega-clusters. In AI 2.0, you have options at every layer:
- On-device: NPUs in smartphones and laptops handling sophisticated tasks without ever talking to a server. Apple, Qualcomm, and others are shipping serious local silicon.
- On-premise: Consumer and prosumer GPUs running inference in your home or office. A few thousand dollars gets you meaningful capability.
- Commodity cloud: The ex-feudal lords, now competing on price. When compute is abundant, you rent on your terms, not theirs. You can switch providers. You can negotiate. You have leverage.
The goal isn’t to run everything locally. The goal is to never be dependent on a single provider who can change the terms. Sovereignty means options.
3. Own the Data (The Soul)
The only durable moat is data that cannot be scraped from the public internet.
This means friction-heavy, real-world data: private operational logs, physical sensor streams, high-trust human interactions, institutional knowledge that lives in people’s heads. This is the messy reality that a centralized model sitting in a data center cannot see.
If your competitive advantage can be reproduced by crawling the web and running a training job, you don’t have a competitive advantage.
The Transparency Paradox
There’s a counterintuitive truth that the feudal model obscures: transparency is a strategic advantage when you actually own your tools.
Companies building on open infrastructure can share their metrics, their architecture, their learnings publicly. Why? Because if you’re not pretending to be bigger than you are, you have nothing to hide. Your moat isn’t secrecy-it’s capability, execution, and trust.
The feudal players can’t do this. They’re locked into an escalating game of marketing narratives and artificial scarcity. They can’t show you the internals because the internals would reveal how much of what you’re paying for is margin, not cost. The squeeze lets them hide behind “capacity constraints” and “premium access.” The glut will expose them.
Open players compete on reality. Closed players compete on perception. In the long run, reality wins.
The Barbell Strategy
So what do you do?
First, accept that the traditional “Knowledge Worker” is being phased out. For decades, you could make a living as a symbol manipulator-writing code, drafting documents, analyzing spreadsheets. This worked because intelligence was scarce and expensive. You were selling access to your brain.
AI has broken this model. When intelligence can be rented by the token, the scarcity disappears. If you are a “prompt engineer” or a “wrapper startup” sitting between the user and the API, you are in the kill zone. You are adding no value that the infrastructure owner cannot capture themselves.
The question is no longer “how smart are you?” but “what do you own?”
In AI 2.0, the middle is death. The future belongs to the extremes:
The Hyper-Sovereign: Hackers and builders who own their stack-hardware, weights, and data pipelines. This is the “Ruby on Rails” moment for AI, where specialized, local tools finally let technology penetrate the actual workplace. These are people building narrow, deep solutions to specific problems, not general-purpose chatbots competing with OpenAI. They own, so they cannot be evicted.
The Hyper-Human: People doing work that requires physical presence, legal liability, or high-trust chaos. A nurse. A trial lawyer. A plumber. An executive who has to fire someone. These are roles that an algorithm cannot insure, cannot bond, cannot hold accountable. The human is the trust layer. They are irreplaceable, so they cannot be commoditized.
The middle-the “AI-assisted” generalist, the “wrapper” startup, the knowledge worker who rents their intelligence from someone else’s API-is the kill zone. You’re competing with commodity intelligence on one side and irreplaceable humanity on the other.
Pick an extreme.
The Natural State
Software wants to be free. Not in the political sense-in the thermodynamic sense. Code has zero marginal cost to reproduce. Any business model that depends on restricting access to something that costs nothing to copy is fighting entropy. It requires constant energy to maintain-legal threats, DRM, API keys, artificial limits.
Hardware is scarce. Compute takes energy. Real-world data requires physical presence to collect. Human trust requires time and skin in the game.
The feudal lords have been exploiting a temporary inversion: the GPU squeeze made compute artificially scarce, which let them charge monopoly rents on what should be a commodity. The glut will restore the natural order. Compute becomes a commodity. The lords become vendors. And the serfs become customers with choices.
Build on the scarce. Let the free be free. Any company charging monopoly prices in a commodity market is running a temporary arbitrage. Any company building on truly scarce resources-proprietary data, human relationships, irreplaceable expertise-is building something defensible.
The Outlook
AI 2.0 isn’t a utopian promise. It’s an existentialist shift toward survival.
The acceleration will likely be even faster than AI 1.0, because it will finally be unencumbered by the bottlenecks and “safety” theater of the feudal lords. Sovereign intelligence doesn’t ask for permission. It doesn’t wait for API access. It doesn’t pay rent.
You cannot rent your way to freedom. The only durable wealth is that which cannot be duplicated by a token generator: proprietary data, human trust, and the leverage to walk away.
Build your castle on land you own. Or prepare to pay rent forever.