China's Open Models Just Made 'Which AI Do We Bet On' a Board Question

Silicon Valley spent a week losing its mind over a Chinese AI model. Moonshot's Kimi K3 landed cheap, strong, and free to download. Then the lobbying and the open letters started flying. Behind all the noise sits a real question you now own: which models does your product bet on, and how much control do you keep? Let me catch you up.

The free weights that aren't really free

Kimi K3 shipped with open weights, meaning you can download the trained model, run it on your own machines, inspect it, and tune it. That gives you more control than piping every request to Claude or GPT. And it can beat some top US systems at a fraction of the cost, at least on paper.

But open weights are not a free service. As Fordham law professor Chinmayi Sharma put it, "A free set of weights is not a free AI service." You still pay for compute, engineering, security, and upkeep. The model is free. Running it well is not.

So the pitch is real but narrower than the headlines. You trade a monthly API bill for owning the stack. Whether that math works depends on your volume and your team, not on a benchmark screenshot.

Why the panic keeps repeating

We have seen this movie. TechCrunch's Sean O'Kane called the reaction a repeat of prior freakouts, where everyone expects one launch to blow everything away. His example: people cheered that Kimi built a copy of macOS in 30 minutes, except it was a graphical mockup, not an OS.

A week later, nobody thought the sky was falling. Kirsten Korosec flagged the part worth watching: a ban on Chinese models would force enterprises onto US labs. "Are we accelerating and ensuring that Americans win the AI race, or are we ensuring that certain frontier labs do better than others?"

Read the loud takes with that filter. When someone tells you a model is too dangerous to use, check who profits if you agree.

Anthropic drew a line, and it matters for your stack

Rumors said Anthropic wanted Chinese open models banned. Dario Amodei shot that down flat: "Anthropic has never advocated for a ban on open-weights models." He called capable open models without dangerous capabilities "a public good."

His worries sit elsewhere: authoritarian states building secret frontier models, plus bio and cyber misuse that is hard to guardrail once weights are out. His fixes are chip export limits, a crackdown on distillation, and mandatory safety testing for all capable models, open or closed. He wants that testing global, even China on board.

For you, the signal is that even a closed-model CEO expects open weights to stick around. The fight is over how they get tested, not whether you can use them.

The copying charge underneath it all

A lot of the heat is about distillation, where one model learns from another model's outputs. Anthropic has alleged DeepSeek, Moonshot, and MiniMax pulled more than 16 million Claude chats through 24,000 fake accounts. Trump's science chief says he has evidence Moonshot ran distillation attacks.

Here is the catch that changes your risk read: it is not clear this is even illegal. Nathan Lambert notes there is no legal precedent that model outputs are IP, and the US Copyright Office says AI-generated material is not protected by copyright. Arena's CEO argues Kimi K3 beats top US models in ways pure distillation cannot explain, and predicts American labs will start distilling Chinese models back.

Translation: the law is behind the tech, and the copying runs both directions. Do not build your roadmap on the assumption that Chinese models get regulated away.

The deep cut

Sharma named the outcome the US labs are steering toward: a portfolio strategy, keeping their best model proprietary while releasing capable open models to hold developer loyalty. Copy that on your side. Do not marry one provider. Build so you can route work across a proprietary flagship for hard tasks and cheaper open weights for the rest, and swap when price or capability moves.

The practical move for Monday: put an abstraction layer between your product and any single model, and run a real cost test of an open model like Kimi against your current API spend at your actual volume. If you are locked into one closed vendor with no exit, that is the risk to name in your next review, before a price change or an export rule names it for you.

Three questions for your team

• If our main model provider doubled prices or cut access tomorrow, how many days until we could switch, and what would break?
• Where in our product could a cheaper open-weight model do the job just as well, and have we actually run that cost and quality test?
• If a customer or regulator asks whether we run a Chinese model, do we have a clear, honest answer about what we use, where it runs, and what data touches it?