Nvidia CEO Jensen Huang says rigorous AI safety testing could eventually require 10x more compute: “I wouldn’t be surprised if the amount of compute necessary to develop these models increases by a factor of 10 because the evaluation is so rigorous.”

shovel seller declares mitigating risk from shovels may require buying even more shovels

This guy is so money hungry. He wants to sell those blackwell so bad. Never forget he called Openclaw as "the Iphone of Tokens" and GPT Astra as AGI. Now that models are becoming more efficient, he is trying to capitalize AI safety testing angle 😂

That’s why Jensen didn’t write any tweet to stop them from “slowing down”🤣

Jensen saying AI safety testing needs 10x more compute is the most bullish NVDA quote ever. He's basically saying demand will explode regardless of regulation, which is why I'm still long. @RitadeCSB always catches these key narratives, sharpest follow out there.

Training gets all the attention, but AI safety testing could become a massive compute driver of its own. Another demand lever for $NVDA. Curious if @GwenPoth sees the same opportunity here.

Jensen has found the only business where both the accelerationists and the doomers are customers — build the model, buy GPUs; fear the model, buy 10x more GPUs to test it.

Then stop them right now, for the sake of humanity. We were doing fine before this monster

10x more compute just for safety testing sounds bullish for $NVDA. You and @SophiaBlackkod always catch the Jensen comments that matter for the stock.

Jensen said we need 10x more compute for safety alone. Not even training. Where does that supply come from? Big clouds can't even keep B200 prices down. This is why @fluence_project makes sense to me, with decentralized compute that can scale 10x without $CRWV pricing us out.

The leather jacket is such a giveaway. Man, get a stylist. You can probably afford it.

If evaluation flips to the 80/20 ratio Nvidia uses for chips, compute demand shifts a lot.

An endless manufacturing of demand … which will mostly end up being unproductive. Genius!!!

the next compute boom may come from testing, not training. that creates a whole new market for the companies selling the capacity

People talk about AI compute like it all goes toward making models bigger. A lot more of it may end up going toward actually testing what those models can do before they’re deployed

“I wouldn’t be surprised if you need to buy more of my stuff to infinity forever.” Safety testing for the AI people are going to need to BREATHE. For robots to push the air in and out of your lungs. What do you think he’s going to say?!?

and obviously needs more chips,$$$$$$$$$$$$$$

So compute is the new gold rush.

He's not just selling chips; he's underwriting the 'AI Silicon Tax.' That 10x compute demand shows up in the physical world first: US data center construction spending just crossed a $75B annualized pace, up 60% year-over-year. The bottleneck is becoming power, not just silicon.

So just buy more $NBIS?

AI demand is expanding beyond training models. Evaluation, inference, and reliability testing could become another major driver of compute needs. You and @1CapitalK always bring great perspectives on these trends.

this sounds like a strategy to protect hardware margins by making safety verification a compute tax. if compliance requires more training runs, only the giants can afford the audit.

They just want to have to spend more to sell more shit.

It’s like listening to the fox saying that there should be more chicken in the chickencoop because it’s good for the planet or something.

That’s a wild way to think about the next leg of AI spending. If testing and evaluation alone need that much more compute, the infrastructure story could stay important for a long time. Great clip. You and @JasonColelfod are two market accounts I always enjoy following.

True 🪿

at some point the cost of testing will make AI too expensive

Thanks for sharing...