We treat intelligence like a utility. Flip a switch, ask a question, get an answer from somewhere in the cloud. Convenient. Powerful. And quietly dangerous.

The pattern is familiar. First we centralized money, then data, now thought itself. Each time the promise is the same: let someone else handle the hard part, and you’ll be free to focus on what matters. Each time the cost is the same: you lose control over the foundation.

Intelligence isn’t just another service. It’s the process by which we understand the world, make decisions, and shape our future. When that process runs on hardware you don’t control, under rules you can’t audit, trained on data you didn’t consent to, you aren’t using intelligence. You’re renting it.

The centralization of AI follows the same logic that created too-big-to-fail banks and surveillance advertising platforms. A few organizations control the models, the compute, the data centers. They decide what knowledge is included, what questions are safe to answer, how your inputs are stored and used. Their incentives aren’t yours. Their definition of “helpful” includes keeping you engaged, compliant, and profitable for their shareholders.

We’ve seen this movie before. Bitcoin emerged as a response to centralized money that could be printed, censored, and inflated at will. The core insight wasn’t technological. It was philosophical: trust should be minimized. Verification should replace authority. Ownership should be possible again.

The same principle applies to intelligence. A model running in a distant data center can’t be fully trusted because you can’t verify its behavior in real time. You don’t know if it’s been updated to reflect new corporate policies. You can’t be sure your private queries aren’t being logged, analyzed, or used to improve the model for others. Even with the best intentions, a centralized system creates a single point of failure for truth itself.

Local models change this equation. When the weights live on your device or in your private cluster, you decide the rules. You can run them air-gapped if needed. You can inspect the code, verify the inference, customize the system prompt, and know exactly what data it has access to. The intelligence becomes yours in the same way your private keys make Bitcoin yours.

This isn’t a rejection of scale. Larger models have clear advantages in breadth of knowledge and capability. But those advantages don’t require surrendering sovereignty for every interaction. The future isn’t one giant model serving everyone. It’s a hierarchy of models: massive foundation models used for training and research, distilled and specialized versions running locally for daily use, with cryptographic verification linking them where needed.

The technical barriers are falling faster than most realize. Quantization, efficient architectures, specialized hardware for inference — all point toward capable intelligence fitting in consumer devices within years, not decades. The question isn’t whether it’s possible. It’s whether we’ll demand it.

Of course, local intelligence comes with tradeoffs. Current consumer hardware still struggles with the largest models. Latency can matter for certain tasks. Keeping models updated without central servers requires new distribution mechanisms, perhaps built on the same decentralized networks that power Bitcoin.

These aren’t insurmountable problems. They’re engineering challenges that builders are already attacking. The open source community has shown time and again that distributed effort can outpace centralized giants when the incentives align with user sovereignty. Look at how Linux, Firefox, and Bitcoin itself evolved through passionate individuals and small teams rejecting the default path of convenience and control.

The real obstacle isn’t technical. It’s philosophical. We’ve been trained to accept rented services for everything. Our email, our photos, our conversations, our search history. Adding our thinking to that list feels natural until you consider the endpoint: a world where the most important decisions in your life are influenced by systems whose full behavior you can never truly know.

This is why the push for open weights matters so much. Not because open source magically solves alignment or safety, but because it makes verification possible. When the model is public, researchers can study it, users can audit it, and alternatives can fork it. Closed models concentrate power in ways that should make any thinking person uncomfortable, regardless of how benevolent the current owners claim to be.

Builders understand this instinctively. The best engineers I’ve known always want to see the source, run their own instances, and modify the system to fit their needs. They reject black boxes on principle. That same instinct needs to be applied to AI before the window closes.

The coming decade will be defined by whether intelligence follows the path of money toward decentralization or continues the trend toward ever greater centralization. My bet is on sovereignty winning eventually, but not without deliberate effort. The default path is always more convenient centralization.

Imagine a world where your personal AI truly belongs to you. It knows your context because you’ve fed it your own data, not scraped from the web. It reasons according to your values because you’ve fine-tuned or prompted it that way. It fails gracefully when it should because you control the guardrails. This isn’t science fiction. The pieces exist today. What remains is the cultural shift toward demanding ownership of our intelligence tools the same way we learned to demand ownership of our money.

The alternative is subtle dependence. Slowly, imperceptibly, letting the cloud think for us until we forget how to think without it. We’ve seen what happens when power concentrates. The pattern doesn’t change just because the domain is intelligence instead of finance.

Intelligence must be owned, not rented. The sooner we recognize this, the better our future will be.