There’s been a lot of talk about artificial intelligence’s current phase: infrastructure building. That has obvious benefits for chipmakers such as Nvidia and hyperscalers like Amazon , Microsoft and Alphabet . Hyperscalers provide the massive cloud computing power needed for AI applications, and analysts have predicted the growing need for more data centers as they house the vast amounts of computing power needed for AI workloads. But the next bottleneck in AI infrastructure — and one to invest in — is networking, according to tech analysts. Networking in normal tech terms refers to a network of devices that can transmit and share information over physical or wireless communications. In AI, however, the requirements are higher because of large language models and other AI applications that require very high bandwidth and low latency. “While Nvidia and its graphics processing units get most of the headlines for generative artificial intelligence, we see networking as a critical companion in the hardware that undergirds models and applications such as ChatGPT,” Morningstar analysts said in a June 2024 report. “The focus largely up until now has been on the [graphics processing units], the actual AI chips. These are the most important part of this puzzle, of course. But networking is where we see the next bottle neck playing out,” Clare Pleydell-Bouverie, portfolio manager at Liontrust Asset Management, told CNBC Pro Talks in May. That’s because the “large-scale systems” that are coming to the market, such as Nvidia’s rack scale systems, require “vastly more” infrastructure content such as networking, she said. Morningstar’s equity analyst for technology, William D. Kerwin, and technology equity strategist, Brian Colello, said they believe the need for fast networking in generative AI will directly translate to “strong, long-term growth for well-positioned networking vendors.” An increase in investment in generative AI model training and inference will drive the AI networking spending growth of 34% over the next five years, said the firm. That translates to $34 billion in spending in 2028, up from Morningstar’s 2023 estimate of $8 billion. “Networking creates a performance bottleneck for generative AI model development,” Morningstar’s analysts said. “Well-positioned networking firms are a great second derivative play to invest in generative AI,” they added. “The majority of generative AI spending will go toward GPUs, but networking is critical infrastructure to enable GPU performance.” Stocks to play the trend Marvell Technology is Morningstar’s top pick to play the generative AI networking trend, with the firm saying it is “attractively undervalued” currently and giving investors an “immediately opportunity” to tap rising generative AI networking investment. Other key winners in this networking trend are Arista Networks , Nvidia and Broadcom , Morningstar said. However, its analysts believe that the generative AI opportunity is “largely priced” in for these three stocks, as their share prices have already gone through “robust” appreciation. “However, patient investors can wait for a pullback, as the long-term fundamental opportunity is strong,” said Morningstar. It added that it’s bullish on Ethernet adoption in generative AI networks, referring to a type of networking standard. Arista would be the primary beneficiary of the transition to Ethernet, according to Morningstar. The current technology commonly used is InfiniBand. “There are very few players that are able to really step up to provide this infrastructure,” Pleydell-Bouverie added, referring to the networking infrastructure. She named Meta and Broadcom as stocks to play the trend. Broadcom is the “leader” in networking chips, and it’s set to benefit as Ethernet emerges, she added. “Ethernet networking is emerging to be the sort of de facto standard for scaling these AI workloads. And Broadcom has got the best in class sort of chips that power this Ethernet network,” Pleydell-Bouverie said.
