Jensen Laughs Off OpenAI’s “Nvidia Killer” After Jalapeno “Cooks” Blackwell in Lab Tests (Rigged?)

Scam Altman: Your time is almost up, old man. Enjoy it.

Jensen Huang: Adorable. Same card as usual?  

Before the Super Bowl (read: Nvidia’s most anticipated earnings report ever), our AI daddy had the audacity to jump on an interview with none other than Jim Cramer.

Of course, Cramer couldn’t resist the urge to ask the Nvidia commander and chief whether he was offended that OpenAI (one of his largest customers and the recipient of billions in Nvidia investment) was simultaneously building a custom chip designed to buy fewer Nvidia chips. 

Jensen responded with the verbal equivalent of patting Sam Altman on the head.

“Lots of projects get started. Lots of projects get canceled,” Huang said. “I don’t have to take anything personally.”

Oh, he’s taking it personally.

The increasingly hot semiconductor in question is Jalapeno, OpenAI’s first custom AI chip developed alongside Broadcom.

Unlike Nvidia’s world famous GPUs, Scam Altman’s favorite new vegetable was designed specifically for inference… the process of running already-trained models so they can answer prompts and operate agents.

OpenAI says Jalapeno will reach its data centers by year-end, allowing ChatGPT to hallucinate faster and more efficiently than ever before.

And according to OpenAI, the little pepper can cook. 

Jalapeno reportedly (key word doing a lot of lifting) completed more AI work per watt and returned responses faster than Nvidia’s GB300 systems during testing. 

Supposedly, SemiAnalysis visited OpenAI’s labs and found Jalapeno beat Nvidia’s Blackwell platform on performance per watt in nearly every tested scenario.

There’s just one tiny detail standing between Sam and total semiconductor domination.

OpenAI’s chip uses newer HBM4 memory, while the Blackwell systems it defeated use HBM3E. That makes Nvidia’s incoming Vera Rubin platform (which also uses HBM4) the more appropriate comparison.

Another important distinction is that Rubin systems are already beginning to ship. OpenAI, meanwhile, is still working with engineering samples and hoping to deploy Jalapeno later this year.

Translation: Sam brought next year’s equipment to fight last year’s Nvidia… and still hasn’t made it out of the laboratory.

Still, Jalapeno doesn’t have to murder Nvidia in broad daylight to become a serious problem.

Inference is expected to become the fastest-growing portion of the AI chip market as models graduate from expensive science projects into products people use every day. If OpenAI can move a meaningful chunk of that work onto its own silicon, it could lower electricity costs and buy fewer chips from Daddy Jensen. 

More importantly, it gives Sam leverage the next time Nvidia arrives with another invoice with an obscene number of 0’s on the end.

Obviously, OpenAI isn’t the only customer attempting to manufacture its own discount. Alphabet has its TPUs, Amazon has Trainium, and Meta recently agreed to deploy one gigawatt of custom chips built with Broadcom technology.

So if Jensen is worried about anything, it should probably be this. His largest customers keep writing enormous checks while spending billions to ensure they won’t have to forever.

But all of that remains fantasy, at least for now.

Nvidia still owns the software ecosystem and remains essential for training frontier models. Jalapeno, meanwhile, is an engineering sample armed with a flattering benchmark and an ambitious shipping date. 

And based on Nvidia’s price action today, I’ll take a wild guess and assume the market remains more concerned about missing Jensen’s rally than fearing Sam’s pepper.

At the time of publishing this article, Stocks.News holds positions in Alphabet, Amazon, and Meta as mentioned in the article.