A hardware wallet doesn’t ask you to trust anyone. The keys are generated on the device, they never touch the internet, and you can verify that fact yourself instead of taking a company’s word for it. That’s the whole appeal. Possession, not permission.

Open-weights AI models are starting to offer the same deal. When the weights are published and downloadable, you’re no longer renting intelligence from behind an API you can’t see into — you can run the model on infrastructure you choose, inspect what it’s doing, and keep a copy that works whether or not the company that trained it is still around next year.

The parallel, leg by leg

• Custodial exchange vs. closed API. Both ask you to trust a company’s internal controls you can’t inspect, with your funds sitting on their books or your prompts running on their servers.
• Hardware wallet vs. open-weights model. Both let you hold the actual thing — private keys, or model weights — instead of a claim on it.
• Seed phrase vs. checkpoint file. Both are portable. Move wallets, move providers, the asset comes with you.
• Losing exchange access vs. losing API access. A custodial exchange freeze locks you out of your funds. An API deprecation or price hike locks you out of the model you built a product on. Open weights mean that decision isn’t someone else’s to make.

None of this means closed models disappear or that every workload needs to run open. Plenty of custodial exchanges still make sense for active trading; plenty of closed APIs are the right tool for a quick prototype. The point isn’t purity, it’s optionality — knowing you can hold the thing yourself if you need to, the same way knowing you can self-custody doesn’t mean you never use an exchange again.

Why this is the open-weights movement’s real argument

The strongest case for open-weights models was never just “free” or “faster to fine-tune.” It’s sovereignty — the same case self-custody made for a decade before most people were ready to hear it. We laid out that self-custody logic in Why Self-Custody Matters, and it’s the backbone of the whole SSHL practice: frontier-quality open models are one of the four legs, alongside zero retention, a verifiable ledger, and distributed compute.

This is also, not coincidentally, evergreen. The specific model names change every quarter. The logic of “do you hold it or does someone else” doesn’t.

There’s also a practical builder upside beyond sovereignty: an open model you hold a copy of keeps working exactly as-is even after the next version ships, which matters a lot if your product’s behavior depends on a specific model’s quirks. A closed API gets upgraded out from under you whether you asked for it or not.

Curious what running open-weights models actually looks like for your product? Come talk it through at /build.