How Nvidia's stock can benefit from this emerging trend in AI models

By Britney Nguyen

As the industry shifts toward AI models with better memory and understanding, Citi expects a new Nvidia offering to best its competitors

Citi sees Nvidia benefiting from the rise of AI reasoning models next year.

The rise of reasoning models that can recall and understand longer conversations could boost demand for artificial intelligence and memory chips next year, according to analysts at Citi. Nvidia looks especially poised to capitalize, thanks to a forthcoming offering.

Last week, Amazon (AMZN) Web Services announced an "episodic functionality" for its AI-agents platform that allows them "to learn from past experiences and apply those insights to future interactions," therefore improving how the tools make decisions. The update reflects "the recognition that memory is the key element to agent intelligence," Citi analyst Atif Malik said in a Monday note.

In order for AI agents to prove worthy for customers, their memory capabilities have to go beyond the short and long term to "understand the context of a current interaction," Malik said.

Google also unveiled a new approach in November to improve continual learning in large language models based on how humans learn and remember information. The concept of Nested Learning was used to develop Google's Hope model, which outperformed other advanced models "in reasoning, language modeling and memory management," Mailk noted.

Reasoning models need much larger context windows to understand and keep relevant conversations going, and Malik said Nvidia's (NVDA) new graphics processing units, which it teased in September, will be the answer to that.

See more: Nvidia's stock can run higher as Citi analysts find plenty of reason for optimism

Rubin CPX, which is expected to launch at the end of next year with its Vera Rubin platform, is purpose-built "for the lowest cost token revenue for ultra-large context processing," Malik said, referring to running AI models, or inference. Tokens are the pieces of data processed by AI models, and developers charge for usage by the number of tokens put in and generated out.

The new Rubin chip "enables AI systems to handle million-token software coding and generative video with groundbreaking speed and efficiency," Malik said, and works with Nvidia's Vera central-processing units and Rubin GPUs. That means companies that will transition to the Rubin platform can "monetize investments at an unprecedented scale," Malik said. He modeled that customers could get a 50x return on investment, or $5 billion in revenue from tokens for every $100 million that is invested.

Another benefit for Rubin CPX is that it uses GDDR7 memory, which is more cost-efficient than high-bandwidth memory components used in Google's tensor processing units and Amazon's Trainium chips. Malik said he sees Rubin CPX giving other AI processors "a run for their money next year," as he modeled the total cost of ownership being three times less thanks to the use of GDDR7 instead of HBM.

Citi has a buy rating on Nvidia's stock, with a price target of $270.

Don't miss: Nvidia's stock is almost historically cheap - and that's a good sign for bulls

-Britney Nguyen

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

(END) Dow Jones Newswires

12-08-25 0958ET