25 companies signed a letter backing open AI — and OpenAI, Anthropic, and Google are nowhere to be found

The backdrop: Washington evaluates restricting Chinese open models

The letter didn't come out of nowhere. Four days earlier, outlets like Axios and TechCrunch reported that the Trump administration is reviewing the possibility of restricting access to Chinese AI models — the Department of Commerce has considered adding Chinese labs to its Entity List, which would cut off access from the U.S. without a special license. The immediate catalyst was Kimi K3, the open-weight model with 2.8 trillion parameters that Moonshot AI launched on July 17 and which, according to the company itself, outperforms GPT-4.5 and Opus 4.8 on code benchmarks.

The problem, flagged by Tom's Hardware, is that an open-weights model already released is nearly impossible to "ban" in practice: once the weights are downloaded and replicated across thousands of private servers, no sanction erases them from the internet. That's why the approach being discussed isn't a frontal ban, but regulatory friction — licenses, sanctions, IP reviews — that functions as a de facto ban without saying it openly. On July 21, Treasury Secretary Scott Bessent announced the government will review Chinese open-source models for possible IP theft, with sanctions as a possible consequence.

Why this isn't just Washington politics

It's easy to read this as a distant fight between governments and labs. It's not, for two concrete reasons:

• The self-hosting option is already real and already cheap. Models like Kimi K3 or DeepSeek can be downloaded and run on your own infrastructure, giving any company — not just the giants — the option to keep its data within its own perimeter and drastically cut inference costs versus paying for an API to a closed model.
• The regulatory dispute can change that option overnight. If Washington decides to restrict access to or commercial use of open Chinese models, any stack that depends on them — even running on your own infrastructure, outside the U.S. — ends up in a legal gray area that no serious client wants to inherit without knowing it.

In other words: the question "what model do I use?" stopped being purely technical (which reasons better?) or economic (which is cheaper?). Now it's also geopolitical and regulatory, and that third variable can shift value from a decision your agency and your client didn't even make.

Data sovereignty: the angle that matters most in LATAM

Huang's letter uses the word "sovereignty" on purpose. For an agency or client in Mexico, Colombia, Chile, or Brazil, this isn't abstract: it means being able to decide on what server — and under what jurisdiction — your users' or patients' or clients' data lives. An open model you can run on local infrastructure (or on a cloud provider inside the region) gives a LATAM company a degree of control over its data that no closed API, hosted on a U.S. data center, can offer in the same way. That's probably the real reason why this U.S. policy debate should also be on the radar of any founder or CTO in the region.

What this means for agencies and developers

If your agency is defining or reviewing its AI stack for the coming months, this fight between open and closed models should influence at least four concrete decisions:

• Don't bet your entire stack on a single model or provider. The regulatory dispute shows that access to a model can change for reasons outside your control. Design architectures that let you swap models (open or closed) without rewriting your entire application.
• Evaluate self-hosting as a real option, not an experiment. If you handle sensitive client data in LATAM, running an open model on your own or regional infrastructure could be the difference between meeting local privacy expectations and depending on another country's foreign policy.
• Document model provenance in every proposal. A client asking "where does the model come from and who controls its weights?" deserves a clear answer — and being able to answer it well is already a business differentiator, not just a technicality.
• Follow policy closely, not just benchmarks. Until a few months ago, choosing a model was a decision of performance and price. Today it's also a regulatory risk decision, and that risk moves faster than model release cycles.

The underlying lesson: the fracture between open and closed AI is no longer a niche debate for researchers. It's a business variable that determines how exposed your agency — and your clients — are to regulatory decisions made in another country. Whoever understands this first will be able to offer something almost no one offers explicitly today: a model strategy with a Plan B.

If you're reviewing what models to use in your stack, how to evaluate self-hosting, or how to explain to a client why this matters, write to us at comercial@geekvibes.agency. It's exactly the kind of conversation we like to have before the client asks.