Two faces. That’s how Jensen Huang framed artificial intelligence during a media briefing in London on October 14-15, 2026 — enormously powerful, and in his words, “potentially dangerous.” Not two markets, not two superpowers. Two faces on the same technology. It’s an unusually plain description from someone whose company supplies the hardware underneath most of it.
The headline from that briefing: Huang said the U.S. “absolutely must” work with Chinese AI companies and keep a channel of dialogue open so the technology develops safely. His argument, roughly, is that talking across borders beats building walls around the science. He also told the Associated Press that society needs new social norms for the age of AI.
I review AI tools for a living. I poke at agent frameworks, model APIs, and the sprawling pile of wrappers that claim to do your job for you. So my first instinct when a chip executive starts talking about international cooperation is to check who benefits. Let’s do that honestly, then get to why this actually matters for the stuff on your laptop.
The messenger has a stake, and that’s fine
Nvidia sells the shovels. A world where Chinese labs buy compute and swap safety research with American labs is a better world for Nvidia than one where the map is carved into sealed zones. Huang isn’t a neutral party here, and pretending otherwise would be lazy.
But a self-interested argument isn’t automatically a wrong one. Plenty of positions are simultaneously profitable and correct. The question I care about is whether the substance survives contact with reality — and on the safety side, it mostly does. Capability research travels. Papers get published, weights get released, techniques get reproduced. Containment has a poor track record in software.
What walls actually do to the tool stack
Here’s where this stops being geopolitics and starts being a Tuesday afternoon problem for anyone building with these tools.
If you’ve evaluated open-weight models in the last couple of years, you’ve already run into the awkward truth that some of the most capable, most permissively licensed options in the space came out of Chinese labs. Teams I talk to deal with this constantly. They find a model that performs well on their task at a fraction of the cost of a frontier API, and then spend three weeks in meetings about whether procurement will approve it. That’s not a safety conversation. That’s a jurisdiction conversation wearing a safety costume.
Genuine safety work looks different. It looks like shared evaluation methods, agreement on what counts as a dangerous capability, red-team findings that circulate instead of sitting in a drawer, and some common language for incident reporting. None of that happens if the two largest concentrations of AI talent only communicate through press releases.
Huang also pointed to Anthropic’s Mythos result as evidence that this dialogue needs to exist. Reporting on that is thin so far, and I’m not going to pretend I know more about it than I do. The structure of the argument is clear enough though: a capability jump at one lab becomes everyone’s problem quickly, and the response works better when it’s coordinated.
New social norms is the vaguer half
The “new social norms” line is the part I’d push back on, not because it’s wrong but because it’s the easiest thing in the world to say and the hardest to cash.
Norms aren’t declared from a stage. They get built from thousands of small decisions — whether a tool discloses that output was generated, whether an agent asks before taking an action that costs money or sends an email, whether a product makes it obvious what data it retained. Those are design choices sitting in the hands of the people shipping software, not the people giving interviews about it. When I test an agent toolkit, the norm question shows up as: does this thing have a confirmation step before it does something irreversible? Half of them don’t. That’s the norm layer, and it’s being set right now by whoever ships fastest.
What I’d watch for
A few things would tell me this is more than a talking point:
- Safety evaluations and red-team methods published in ways researchers on both sides can actually reproduce
- Any shared vocabulary for classifying dangerous capabilities, even a rough one
- Agent tooling that treats consent and reversibility as defaults rather than enterprise upsells
- Model cards that are honest about limitations instead of marketing documents with a table in them
I’m cautiously on board with Huang’s core point. Shared safety standards are more useful than isolation, and the people closest to the technology usually know that first. I’d just prefer the follow-through arrive as published methods and better-designed tools rather than more briefings about how important dialogue is. The tools are the part I can test.
🕒 Published: