Huang says AI safety testing needs its own data centers and calls that new demand

CBS aired Jo Ling Kent's interview with Jensen Huang on Sunday, September 20, and the headline wrote itself. Nvidia's CEO called warnings that AI could end humanity within a decade "not grounded on facts" and put the odds of the world ending in 2030 at zero. Every outlet ran that line. I think they took the wrong one. Five days earlier Huang had already explained where all the new safety work is supposed to run, and the answer is a data center nobody has bought yet.

The line that should have led

Dario Amodei and Sam Altman spent the first half of September asking the industry to pace itself at the frontier, and I covered the evaluator proposal when it landed. The bear case that followed was simple. Labs that slow down buy fewer chips. Huang's answer never argues with the premise. It reprices it. On stage with Marc Benioff at Dreamforce on September 15, asked about pacing and safety, he went straight to infrastructure.

However, safety is an engineering problem. We're developing software after all. We're developing computing systems after all. It's a complicated computing system but it's ultimately a computing system. And so the first thing is to make sure that we create the test environments for testing these complex systems and those test environments have to be built in a good way, built in a safe way.

Jensen Huang, September 15, 2026, source

In a television interview the same week he put a size on those environments. Training and inference are the two workloads the entire capex cycle was built on. Huang added a third.

We now will have another third system. That third system is basically testing, environmental testing labs. And so these testing systems will be extremely large so that we can have products that are functional, capable, safe, and secure. And that requires massive data centers to test them. So we just created yet another new demand for AI.

Jensen Huang, September 15, 2026, CNBC interview

If he is right, the safety argument is not a brake on the buildout, it is a line item inside it. That is either the most useful reframing anyone offered this month or the most self-serving one, and it is worth sitting with the possibility that it is both. TechCrunch covered the Dreamforce session and led with the regulation line. The compute claim went unremarked.

Kent asked him the obvious question

To her credit she put the conflict on the table directly, asking why Americans should trust his read on safety when his net worth is tied to selling chips.

Our company's success is directly connected to the safe deployment of products and services. If the computer industry doesn't provide value to the world safely and productively, provide enormously positive impact as we have so far, if we don't continue to do that our value would be diminished.

Jensen Huang, September 20, 2026, CBS Sunday Morning

It is a decent answer to a different question. Nobody thinks Huang wants unsafe AI. The objection is narrower. A man who sells the picks has an interest in defining every new problem as one more pick, and on the testing point he did exactly that, on camera, with no follow up. CBS ran the doomsday angle instead. So did nearly everyone else.

He also put a number on next year

On September 17, at an AI summit convened by King Charles III in Scotland, Huang gave the figure: he expects Nvidia to sell twice as many chips next year as it does this year, and Fox Business carried the quote from the event. That is a unit forecast rather than a revenue one, and he offered it to reporters without a slide or a caveat. It lands two days after nobody who sells compute moved a number in response to Amodei's slowdown essay. The sellers are not holding their forecasts steady. They are raising them while the safety fight is still live.

Howard Marks will not put a number on anything

The same Sunday, at the Qatar Economic Forum's New York edition at the Plaza, Howard Marks was asked whether AI looks like a bubble. He declined to say.

We certainly have exuberance. The question is whether it's irrational. And since nobody, in my opinion, can spec out exactly what it is that AI is going to be doing and with what effect and with what profitability, I don't think anybody can say that it's irrational, but you have to consider the possibility.

Howard Marks, September 20, 2026, source

Then he went further, and this is the part that sits badly next to a third data center.

I just think that there's an unusual degree of uncertainty in coming up with any quantification of earnings potential and thus of what we call intrinsic value. I think it's going to be hard. And Warren Buffett always said, oh, we put that on the too hard pile.

Howard Marks, September 20, 2026, source

So Huang has a workload nobody has budgeted for, and Marks cannot value the workloads that already exist. Two people the market takes seriously, same day, and the statements do not fit together.

Bill Ford has sat through four of these

At the same forum, Bill Ford of General Atlantic got the question that matters most to anyone holding AI paper. Does the safety anxiety pause capital allocation?

Every tech cycle I've invested in, the PC cycle, the Internet cycle, the mobile cycle, the cloud cycle, and now AI, there's always moments of overcapitalization and overvaluation. But over a ten, twenty-year period, there's plenty of opportunity.

Bill Ford, September 20, 2026, Qatar Economic Forum

He added that the safety risks are real, that they should be handled with a mix of industry self regulation and government rules, and that none of it stops the impact of AI. That is the house view among people writing checks. It is also a thirty year investor declining to say anything at all about the current price.

Gary Cohn found the invoice

Also on Sunday, Gary Cohn spent most of a Face the Nation segment on diesel prices and the Fed, which raised rates on September 16 by a quarter point to a range of 3.75% to 4% on a unanimous vote. Then he named the buildout as a force in the debt market.

We have enormous demand from the data centers in the AI world. There's more debt being issued by corporates as well as the governments than we've seen in a long period of time.

Gary Cohn, September 20, 2026, CBS Face the Nation

Earlier in the same interview he said real wages have turned negative and households are covering the gap out of savings. Put the halves together and you get the uncomfortable version of Huang's story. A third class of data center means more issuance, while the Fed tightens into supply driven energy inflation, in an economy where the median household is already drawing down to stay level.

The apology, and the claim that needs a number

The most human moment in the CBS interview had nothing to do with extinction. Asked about the bipartisan backlash against data centers in small towns, Huang apologized, and said the industry should have gone to talk to those communities long before the trucks arrived. I believe he means it. What follows is thinner. He said water consumption is a myth now because modern sites run closed loop cooling, and compared a year of use to a swimming pool evaporating. He gave no figure for any specific campus, and the claim describes new builds rather than the installed base towns are actually fighting. If Nvidia wants that argument to land, the industry should publish site level water draw and grid contribution. Today the protesters have numbers and the industry has an analogy.

Does any of this slow the buildout?

On the evidence of one Sunday, no, and the reason is that the pacing argument never had a mechanism behind it. Amodei asked for evaluators. Huang priced them, and the price is more compute. Ford says the capital keeps moving. Marks will not call it irrational. Cohn says the paper is landing in the bond market. The only thing that actually cuts the number is a buyer who stops buying, and last week the buyers were raising money rather than cancelling orders. So watch the testing claim, because it is falsifiable. If Huang is right, some hyperscaler discloses a dedicated evaluation cluster within two quarters. If none does by spring, the third system was a talking point that happened to arrive exactly when the bear case did.