Twice as many chips. That’s what Jensen Huang says he expects Nvidia to sell next year compared to this one, and it’s the number I keep circling back to as I read his comments on AI safety. Doubling the supply of the hardware that everything runs on is not a subtle move. It’s the infrastructure equivalent of widening every road in the country at once and trusting drivers to figure out the rest.
Which is roughly the position Huang staked out. Speaking on Tuesday, September 15, 2026, the Nvidia CEO dismissed the need for new AI security regulations, arguing that market forces will keep development honest. Earlier in the year he’d told the AP that society needs “new social norms” in the age of AI. And the line that gave this story its legs, aimed squarely at developers: if it’s not ready, just hold it back.
I build bots for a living. That last sentence is the only part of this that lands as practical advice, and it’s worth taking seriously even if you disagree with everything around it.
Shipping discipline is a real skill, and most of us are bad at it
“Hold it back” sounds obvious until you’ve had a half-working agent sitting in a branch on a Friday afternoon with a stakeholder asking when it goes live. The pressure to ship is not abstract. It comes from demos that went well, from a competitor’s launch post, from your own sunk-cost attachment to the thing you spent three weeks on.
What makes it harder with bots specifically is that “ready” is fuzzier than it is for a CRUD app. A form either validates or it doesn’t. A conversational agent works beautifully in eighty test cases and then confidently tells someone the wrong thing in the eighty-first. You can’t unit test your way to certainty about behavior in an open input space, so the temptation is to declare it good enough and let production be the test suite.
What holding it back actually looks like in practice
If you want Huang’s advice to mean something in your repo rather than in a keynote, it has to translate into mechanics. A few that have saved me:
- Define your refusal behavior before your happy path. Write down what your bot should do when it doesn’t know, when the request is out of scope, and when a tool call fails. If those three paths aren’t implemented, you’re not ready, no matter how good the demo felt.
- Keep an adversarial eval set that grows every time something breaks. Every weird production transcript becomes a permanent test case. This is the closest thing to a regression suite you’ll get.
- Gate the blast radius, not just the quality. Read-only first. Then writes behind a human confirmation. Then autonomous writes. Shipping a partially ready bot with no write access is a very different risk than shipping the same bot with your production database credentials.
- Log everything, and actually read the logs. Cheap to add, and the only way you’ll know whether “ready” was true.
- Ship to ten people before ten thousand. A staged rollout is a soft form of holding it back that still gets you signal.
None of that requires a regulator. It requires you to have decided, in advance and in writing, what your bar is, because you will not be a neutral judge of your own work at 6 p.m. on a Friday.
Where the argument gets thinner
Huang’s framing is that the market handles this. It’s a coherent position, and there’s a version of it that’s true: products that embarrass people get abandoned, and reputational damage is a real cost. I’ve watched teams quietly pull agents that made customers angry, and no law forced them to.
But market discipline works on a lag. The feedback arrives after the harm, and it arrives loudest for companies with a brand to protect. The developer shipping a scraping agent under a throwaway name doesn’t feel that pressure at all. Neither does the internal tool nobody outside the company will ever review. Market forces are a filter with fairly large holes, and “hold it back” is advice, not a mechanism.
There’s also a structural oddity in a chip vendor telling application developers to slow down. The company selling twice the compute next year is asking the people buying it to exercise restraint about what they do with it. I don’t think that’s cynical, exactly. Huang has been consistent that AI is now essential infrastructure rather than one breakthrough product, and infrastructure providers genuinely don’t control what gets built on top. But it does put the burden entirely downstream, on us.
The part that’s ours either way
Whatever happens with regulation, the decision about whether your agent goes live tomorrow belongs to you. That’s not a comfortable amount of responsibility, and I’d rather have it than not.
So take the advice and skip the ideology. Write down your bar. Build the refusal paths. Stage the rollout. And when the thing isn’t ready, be the person willing to say so out loud in the standup, which is genuinely harder than any of the engineering.
🕒 Published: