The AI safety debate is often framed as a choice between trusting the people closest to the technology and dismissing their warnings as self-interested. The record is more complicated: public warnings, documented safety interventions, rapid product development, and preparations for possible public listings are happening at the same time.
Nvidia chief executive Jensen Huang represents one clear pole in that argument. In a July 2026 interview with Axios, Huang called the idea that AI would end humanity “complete nonsense.” He also suggested that some companies could benefit from regulation written in their favor, although Axios noted that he did not name OpenAI or Anthropic when making that argument.
The capital-markets side of the story requires more precision. Anthropic announced on June 1 that it had confidentially submitted a draft S-1 for a proposed initial public offering, but it explicitly said the number of shares and the offering price had not been set.
OpenAI announced its own confidential S-1 submission on June 8. Its statement said the company had not decided on timing and might remain private for a while because some things would be easier to do outside the public markets.
Those filings establish that both companies took formal steps that could lead to IPOs. They do not establish that either company had priced a listing, and they do not prove that executives privately reject the safety concerns expressed inside the industry.
There is also concrete evidence that safety concerns have changed development decisions. In July, OpenAI disclosed that it paused access to a long-running model after limited internal use revealed failures its existing pre-deployment evaluations had not captured. The company said it created new evaluations, strengthened safeguards, and later restored limited access under monitoring.
That kind of disclosure is more useful than a vague reference to unnamed researchers or social-media discussions because it identifies what happened and what the company did in response. It does not settle the larger debate over existential risk, but it demonstrates that at least some safety concerns have translated into operational decisions.
Huang’s criticism still matters. His argument challenges policymakers to distinguish between demonstrated risks and scenarios that remain uncertain, and he has openly questioned whether some regulatory proposals could advantage incumbents. But his interview does not establish that safety warnings from OpenAI, Anthropic, or their employees are fabricated or merely strategic.
The same caution applies in the other direction. A model failure, an internal pause, or a warning from an employee can be important evidence that a risk deserves investigation. None of those things, by itself, establishes a precise probability that advanced AI will cause human extinction.
The tension is therefore less tidy than the familiar “doomers versus optimists” framing suggests. Companies can believe that advanced systems create serious risks while also believing that continuing development is commercially or strategically necessary. Those positions can be uncomfortable together without logically cancelling each other out.
The IPO filings are useful for the same reason: they put limits around what can responsibly be claimed. A confidential S-1 preserves a path toward a public offering. It is not proof of a fixed listing date, a final valuation, a share price, or a corporate judgment that catastrophic AI risk is impossible.
For regulators, investors, and the public, it helps to separate three different categories of evidence. There are forecasts about what advanced AI might eventually do, documented incidents and safety interventions involving systems that exist now, and corporate decisions about financing, releases, and governance. Collapsing all three into a single story about either hypocrisy or panic makes the debate harder to evaluate.
The behavior that can be audited matters most. Did a company pause deployment when something went wrong? Did it change evaluations or safeguards? What did it actually disclose in regulatory filings? What commitments can outside parties verify?
That leaves a real contradiction, but a narrower and more defensible one. Frontier AI companies are trying to manage risks they say may become extremely serious while remaining inside a competitive system that rewards speed, capability, and scale.
Huang thinks the extinction narrative has gone much too far. Others inside the field take the possibility seriously enough to change how systems are tested and deployed. The public does not need to choose between those positions on faith. It can start with the record: what was said, what was filed, what was paused, and what actually changed.