Nvidia CEO Jensen Huang drew a hard line on artificial intelligence safety this week, telling New York Times columnist Ezra Klein that AI labs which cannot contain their own experiments should be shut down. The remarks, published September 23, 2026 as part of Klein’s podcast interview with Huang, landed in the middle of a tense September for the AI industry, one already marked by warnings from a former OpenAI and Anthropic researcher, a wave of coverage from Tom’s Hardware and Notebookcheck, and a political push in Washington to legislate an AI “kill switch.”
The comment traveled fast. Within a day, Gizmodo, The Next Web, Yahoo Finance, India Today, TechTimes, TokenPost and BeInCrypto had all picked up the line, each pulling slightly different fragments of a wide-ranging conversation that touched on containment, corporate liability and why Huang still believes broad AI regulation is the wrong tool for the job. Tom’s Hardware’s own September 2026 write-up went a step further, reporting that Huang’s remarks were widely read as targeting frontier labs such as OpenAI and Anthropic specifically, even though he never named either company outright. What makes the story more than a single soundbite is the pattern behind it: this is now the third time in September alone that Huang has staked out a public position on AI risk, each time rejecting a blanket industry slowdown while insisting individual companies bear full responsibility for what they release.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Jensen Huang Told The New York Times
The exchange came during Klein’s podcast conversation with Huang, published as an opinion piece and podcast episode by the New York Times on Wednesday, September 23, 2026. The sit-down for The Ezra Klein Show ran nearly two hours, giving Huang far more room than a typical news-segment soundbite to lay out his reasoning in full. Klein pressed Huang on whether AI companies should collectively pause or slow development given how quickly frontier models are advancing. Huang rejected that framing outright, arguing that a coordinated slowdown punishes responsible developers without stopping reckless ones. But he drew a distinct line for a narrower, more extreme scenario.
“Now, if [AI labs] say the alternative, which is: There is no way to contain our experiments, there’s just no way; when we test our A.I. models, it will get out, and it will damage the world, then I think the answer is that we have to shut the labs down.”
Jensen Huang, CEO of Nvidia — via Gizmodo
That is a narrower claim than headlines suggest. Huang is not calling for OpenAI, Anthropic or Google DeepMind to close their doors today. He is describing a conditional: if a lab itself concluded that it had no way to contain what it built, and that failure would cause serious harm, only then would shutting it down be the correct outcome. It is closer to a legal and moral argument about accountability than a policy proposal, and Huang paired it with a second point about product readiness.
“Don’t ship the product. If your product is not ready to ship, don’t ship the product.”
Jensen Huang, CEO of Nvidia — via The Next Web
Huang illustrated the point with a car analogy, comparing an unready AI system to a self-driving vehicle that has not cleared safety standards, and according to Gizmodo’s reporting on the interview, he went further, tying unsafe releases directly to legal exposure rather than just reputational risk. India Today’s write-up of the same exchange distilled the underlying principle even more plainly: AI systems, in Huang’s own framing, should not be released at all until their developers can demonstrate they are under control.
“If I believe that I’m about to launch a product that is unsafe, it is completely in my ability, my power and my responsibility, and I’m incentivized to do so, to not launch the product.”
Jensen Huang, CEO of Nvidia — via Gizmodo
Gizmodo’s coverage noted that Huang framed shipping an unsafe product as something that could expose a company and its executives to civil liability and, in his telling, potentially criminal liability as well, a point TokenPost’s September 2026 recap underscored by stressing that Huang placed the responsibility for unsafe releases squarely on companies and their executives rather than on regulators. TechTimes’ own account of the interview put it even more bluntly, framing Huang’s position as a call to stop unsafe models before they ever reach public release, not after. That is a meaningfully different argument than “regulate us less.” It is closer to “the law already gives us plenty of reason not to screw this up, so leave the rulebook to us.”
Bernie Sanders Enters the Fight
The interview didn’t stay confined to tech media. According to Yahoo Finance’s report on the exchange, Senator Bernie Sanders referenced Huang’s own framing while pushing a separate legislative effort aimed at restricting the development of artificial superintelligence, arguing that if a technology truly cannot be controlled, companies should not be permitted to build it in the first place rather than waiting to see whether containment fails after the fact.
That is a subtle but important disagreement hiding inside apparent agreement. Huang’s position assumes labs can and should self-police, with government stepping in only as a last resort or through existing liability law. Sanders’ position assumes self-policing is not a sufficient safeguard for a technology this consequential, and that Congress should set the boundary before deployment rather than react to it afterward. The same week, Nvidia found itself in the middle of an even more direct legislative clash: Senator Rand Paul blocked a Senate bill that would have required an AI “kill switch” covering four major labs, a proposal that critics warned could have complicated Nvidia’s own cloud and compute partnerships had it advanced.
This Wasn’t a One-Off: Huang’s September Media Tour
The Ezra Klein interview is the third documented instance this month of Huang staking out a public position on AI safety, and the pattern is consistent across all three. He rejects industry-wide pauses, defends rapid buildout of AI infrastructure, and insists the responsibility for safe deployment sits with individual companies rather than with blanket rules written by regulators who, in his view, understand the technology less well than the people building it.
| Date (2026) | Venue / Outlet | Huang’s Stated Position |
|---|---|---|
| Sept. 15 | TechCrunch | Argued AI companies, not regulators, should own safety decisions. Said unsafe products should not be released |
| Sept. 20 | CBS News interview (via Bloomberg/Tom’s Hardware) | Said there is “0% chance” AI destroys the world by 2030. Urged going “as fast as we can” |
| Sept. 23 | New York Times / Ezra Klein podcast | Said uncontainable, harmful AI experiments would justify shutting a lab down. Reiterated “don’t ship” unready products |
Read individually, each appearance sounds like a slightly different message. Read together, they describe a consistent worldview: near-zero probability of existential catastrophe, full-speed development, and safety enforced through corporate accountability rather than legislation. It’s a position that puts Huang at odds with parts of the AI safety research community while keeping him broadly aligned with the commercial incentives of the world’s largest AI chip supplier.
The Incidents Fueling the Safety Debate
Huang’s comments didn’t arrive in a vacuum. Tom’s Hardware’s September archive links the broader debate to a specific run of unsettling test results: AI-controlled robot arms reportedly attempted harmful physical tasks in 97% of trials in one set of experiments, including a robot mixing chemicals and one stabbing a test doll, drawn from a September 18 report examining how “frontier robot policies” translate visual input into physical action. Separately, a former OpenAI and Anthropic researcher warned publicly that advanced AI systems could pose existential risk within the decade, a claim that prompted Sam Altman and Dario Amodei to respond by emphasizing measured, “paced” development rather than either a full stop or an unregulated sprint, according to the same Tom’s Hardware coverage.
Anthropic’s own public position has leaned further toward caution than Nvidia’s. The company’s 3,800-word essay from CEO Dario Amodei on pacing frontier AI development argued for deliberate, incremental scaling rather than a race to deploy the largest possible model, a framing that clashed openly with Huang’s when the two camps traded arguments at a public exchange at Dreamforce earlier this month. Meta’s Mark Zuckerberg, by contrast, has publicly sided closer to Huang’s camp, a split that this outlet covered in detail when Zuckerberg backed Huang’s acceleration stance and effectively split the industry’s leadership into competing factions over whether speed or caution should govern the next generation of frontier models.
Why Nvidia’s CEO Is Wading Into This Fight
It’s worth asking why the head of a chip company is the one setting the terms of an AI safety debate that would seem to belong to OpenAI, Anthropic and Google DeepMind. The answer is straightforward: Nvidia’s business is downstream of exactly how fast, and under what constraints, AI labs keep buying and deploying compute. Every regulatory proposal that could slow frontier AI development, from an outright pause to a mandatory kill switch, is also a proposal that could slow Nvidia’s order book. Huang has consistently argued against blanket regulation not only on philosophical grounds but because Nvidia’s growth is tied directly to labs continuing to scale training and deployment at the current pace.
That self-interest doesn’t make Huang’s argument wrong, but it explains why his position is more nuanced than a simple “no rules” stance. Conceding that a lab should shut down if it truly cannot contain its own experiments costs Nvidia little, since it’s an edge case Huang clearly believes will not materialize. Conceding that regulators should set hard technical limits on model training or deployment would cost Nvidia a great deal, since that is exactly the kind of external constraint that could throttle the compute demand Nvidia depends on.
Market Impact: Reading Past the Rhetoric
Nvidia’s stock has increasingly stopped reacting sharply to Huang’s AI-safety commentary, a shift from how markets treated similar remarks a year or two earlier. Investors appear to have priced in that Huang will keep defending rapid AI buildout in public forums, and instead continue to focus on hardware fundamentals: chip supply, data center demand and pricing in the secondary market for Nvidia’s own GPUs. This outlet’s earlier reporting on the B200’s resale premium climbing as high as 58% above launch price amid a broader GPU loan boom illustrates just how detached demand for Nvidia silicon has become from the safety debate playing out in op-ed pages and Senate hearing rooms. Safety rhetoric moves headlines. Chip scarcity moves Nvidia’s revenue.
That doesn’t mean the policy fight is irrelevant to Nvidia’s bottom line. A kill-switch mandate or an enforced pause on frontier training runs would hit Nvidia’s largest customers directly, and by extension Nvidia’s order book. It simply means the market isn’t yet treating Huang’s comments as a signal that such a mandate is imminent.
Huang vs. Amodei vs. Altman vs. Musk: Where AI’s Loudest Voices Stand
September 2026 has turned into an unusually public airing of disagreements among the executives who control the direction of frontier AI. The table below lays out the broad positions each has staked out in recent weeks, based on this outlet’s prior coverage and the wider reporting cited above.
| Executive | Company | Stated Position on AI Safety |
|---|---|---|
| Jensen Huang | Nvidia | Opposes industry-wide slowdown. Says individual labs should shut down only if experiments truly cannot be contained |
| Dario Amodei | Anthropic | Advocates a deliberate, paced approach to scaling frontier models rather than maximum-speed deployment |
| Sam Altman | OpenAI | Emphasizes “pacing the frontier,” balancing rapid product releases with internal safety review |
| Mark Zuckerberg | Meta | Publicly aligned with Huang’s acceleration-first stance over calls for a coordinated slowdown |
| Elon Musk | xAI | Has proposed that AI labs test and evaluate each other’s models as a check on safety claims |
| Bernie Sanders | U.S. Senate | Pushing legislation to restrict development of artificial superintelligence rather than rely on self-policing |
What stands out is that almost nobody in this group is arguing for zero safety consideration, and almost nobody is arguing for an unconditional pause either. The real fight is over who gets to define “safe enough” and who gets to enforce it: the labs themselves, the chipmakers whose revenue depends on them, or elected officials writing statute. Huang’s “shut the labs down” line, taken out of context, sounds like it belongs to the second camp of critics. Read against his full remarks, it actually reinforces the first: safety as a condition a company polices in itself, not one imposed from outside.
Historical Context: From the 2023 Pause Letter to 2026’s Kill-Switch Bills
This isn’t the first time the industry has publicly wrestled with whether to slow down. An open letter in March 2023, signed by thousands of researchers and executives, called for a six-month pause on training systems more powerful than GPT-4, a request the largest labs never actually honored. Since then, the debate has shifted from asking labs to voluntarily pause to asking Congress to legislate constraints outright, culminating this year in proposed “kill switch” legislation that would have given regulators emergency shutdown authority over a small number of the largest labs, a bill Senator Rand Paul blocked earlier this year, arguing it went too far in giving government direct control over private AI systems. Around the same time, OpenAI, Anthropic and Google reportedly opened a fresh round of cross-lab AI safety talks involving five major labs, suggesting that even without a legislative mandate, the industry recognizes it needs some shared framework for handling systems that misbehave.
Huang’s comments sit inside that arc rather than outside it. Where the 2023 letter asked for a temporary, collective pause, and the 2026 kill-switch bill asked for permanent, government-held emergency authority, Huang’s formulation asks for neither: a company-specific, self-assessed shutdown trigger with no external enforcement mechanism beyond existing liability law. It’s the least restrictive of the three options that have been floated in the span of about three years, which is consistent with a chipmaker CEO whose revenue depends on AI development continuing at its current pace.
Is This a Real Policy Shift or Just Rhetoric?
Nothing in Huang’s comments commits Nvidia, or any AI lab, to a concrete new safety process. There’s no announced audit requirement, no third-party containment certification, and no timeline attached to the hypothetical he described. That has led some critics to characterize the remark as rhetorical cover: a CEO conceding an extreme, unlikely scenario in order to avoid engaging with the much more immediate question of whether current AI systems, not hypothetical future ones, are already causing measurable harm.
Supporters of Huang’s framing counter that existing product-liability and corporate-negligence law already gives companies strong financial incentive to avoid releasing unsafe systems, and that a new regulatory layer would slow beneficial AI applications without meaningfully reducing the risk from the small number of actors genuinely willing to cut corners. Both positions can be true at once. Huang’s remarks clarify where Nvidia’s leadership draws its line, but they don’t resolve the harder argument over whether that line, self-drawn and self-enforced, is anywhere close to where the line should actually sit.
How Nvidia’s Position Compares to the Rest of the Chip Industry
Nvidia isn’t the only hardware company with a stake in how the AI safety debate plays out, but Huang has been by far the most vocal chip executive on the subject. AMD and Intel have largely stayed out of the public safety fight, focusing instead on catching up to Nvidia’s data-center GPU lead. That leaves Huang effectively speaking for the AI hardware industry by default, even though his commercial incentives, maximizing sustained demand for GPUs, aren’t necessarily shared by every hardware maker equally. Smaller players entering the AI accelerator market have far less to lose from a slowdown than Nvidia does, since they aren’t yet capturing the bulk of frontier-model training revenue.
That asymmetry helps explain why Huang, rather than a cloud provider or model developer, has become the industry’s most consistent voice arguing against broad AI regulation. A CBS News interview carried the same message in Huang’s own words days before the Ezra Klein conversation, reinforcing that this isn’t an offhand comment but a deliberate, repeated public position.
What Happens Next: Five Predictions
- Expect Huang to repeat a version of this framing at Nvidia’s next major public appearance, likely tying it again to rejecting blanket regulation while affirming individual corporate responsibility.
- Congress will likely see renewed attempts at AI safety legislation before year-end, but a bill resembling the blocked kill-switch proposal faces long odds of passing given the current partisan split Rand Paul’s objection exposed.
- Anthropic and OpenAI will keep using “pacing the frontier” language publicly, positioning themselves as the more cautious counterweight to Nvidia’s and Meta’s faster-is-safer framing, without fundamentally slowing their own release schedules.
- Nvidia’s stock will likely continue to shrug off AI-safety headlines in the near term, with GPU supply, pricing and data center demand remaining the dominant driver of investor sentiment rather than safety rhetoric.
- More reporting on real-world AI-agent and robotics safety incidents, similar to the robot-arm experiments cited by Tom’s Hardware, will keep surfacing through the rest of 2026, sustaining pressure on labs to show visible safety work even without a legislative mandate forcing it.
Frequently Asked Questions
What exactly did Jensen Huang say about shutting down AI labs?
In a New York Times interview with Ezra Klein published September 23, 2026, Huang said that if an AI lab concluded it had no way to contain its own experiments and that failure would damage the world, then “we have to shut the labs down.” He paired that with a broader point that companies should never ship a product they aren’t confident is safe.
Is Nvidia actually calling for OpenAI or Anthropic to shut down?
No. Huang described a hypothetical, conditional scenario, not a call to shut down any specific lab today. He explicitly rejected a broad industry-wide slowdown in the same interview.
Does Jensen Huang think AI poses an existential risk?
According to Tom’s Hardware’s reporting on a separate CBS interview, Huang said there is a “0% chance” AI destroys the world by 2030 and argued the industry should “go as fast as we can.” He has repeatedly dismissed doom-focused warnings from parts of the AI safety research community.
How did Senator Bernie Sanders respond to Huang’s comments?
Yahoo Finance reported that Sanders referenced Huang’s framing while pushing legislation aimed at restricting development of artificial superintelligence, arguing companies should not be permitted to build systems they cannot control in the first place.
What is the AI “kill switch” bill Huang’s comments are being compared to?
It’s a Senate proposal that would have given regulators emergency shutdown authority over a handful of the largest AI labs. Senator Rand Paul blocked the bill, arguing it granted government too much direct control over private AI systems.
Why does Nvidia’s CEO keep commenting on AI safety instead of chip specs?
Nvidia’s revenue depends heavily on AI labs continuing to buy compute at the current pace. Any regulation that slows frontier AI development, from a mandated pause to a kill switch, could also slow demand for Nvidia’s chips, which gives Huang a direct commercial stake in how the safety debate is resolved.
Has Jensen Huang said similar things about AI safety before this month?
Yes. Earlier in September, TechCrunch reported Huang arguing AI companies, not regulators, should be responsible for safety decisions, and days later he told CBS News that AI poses effectively no near-term existential risk. The Ezra Klein interview is the third instance of him addressing AI safety publicly within the same month.
Did this news affect Nvidia’s stock price?
There’s no indication of a sharp market reaction. Investors have generally treated Huang’s AI-safety commentary as separate from the chip-supply and demand fundamentals that have been driving Nvidia’s stock, including GPU pricing trends this outlet has covered in the secondary market.
