Complex data center workloads like training machine learning models and running artificial intelligence (AI) applications would take a very long time if powered only by central processing units (CPUs). To that end, specialized semiconductors are used to speed up computationally intensive AI tasks.
In that semiconductor vertical, Nvidia (NVDA 0.39%) graphics processing units (GPUs) have emerged as the industry standard. In fact, the company has between 80% and 95% market share in AI accelerators, according to analysts. But Nvidia shareholders recently got some worrisome news from rival Broadcom (AVGO 3.15%).
Nvidia shareholders just got some bad news from Broadcom
Broadcom sells a range of semiconductor products, including combined Wi-Fi and Bluetooth chips in Apple and Samsung smartphones, as well as networking chips in Arista switches. But Wall Street is particularly fascinated by its leadership in application-specific integrated circuits (ASICs.) ASICs are chips purpose-built for specific use cases like accelerating artificial intelligence (AI) workloads.
Analysts estimate Broadcom has approximately 60% market share in custom AI chips due to its relationships with three hyperscalers, a term that refers to companies with vast data center footprints. While Broadcom has not identified its hyperscale customers, analysts generally believe they are Google parent Alphabet, Meta Platforms, and TikTok parent ByteDance.
Broadcom estimates revenue from its three existing hyperscale customers will range from $60 billion to $90 billion in 2027, up from $12.2 billion in 2024. In other words, the company anticipates that custom AI chip sales will increase at least 70% annually in the next three years, but perhaps as quickly as 95% annually.
That is disappointing for Nvidia shareholders because it means Broadcom will likely gain market share in AI accelerators. In fact, analysts at Morgan Stanley estimate that ASICs will account for 13% of AI accelerator sales in 2027, up from 11% in 2024. They also think that figure could hit 15% in 2030. But there is more bad news for Nvidia shareholders.
Broadcom CEO Hock Tan told analysts on the fourth-quarter earnings call that Broadcom has selected two new hyperscalers that are likely to be revenue-generating customers by 2027. That means revenue from custom AI chips could actually grow faster than 95% annually in the next few years. Importantly, while Broadcom did not identify the customers, analysts believe they are Apple and ChatGPT creator OpenAI.
Image source: Getty Images.
Nvidia shareholders still have good reason to be optimistic
I mentioned that Broadcom selected two additional hyperscalers as potential customers. CEO Hock Tan himself used that word because Broadcom will not develop ASICs for small companies, nor would small companies have an interest in using custom AI chips.
There are two reasons for that. First, designing ASICs is expensive, so the customer must have a data center footprint large enough to warrant the expense. Piper Sandler analyst Harsh Kumar recently told CNBC that each chip costs about $500 million to design, so it would not be financially sensible to work with clients that can only order a few thousand. Instead, orders must be in the range of 250,000 to 500,000 units.
Second, custom chips are less flexible because they are designed for specific workloads and lack supporting software development tools. Nvidia offers a robust ecosystem of code libraries and pretrained models that streamline GPU application development. No such tools exist for ASICs. That means deploying custom chips requires a high degree of technical expertise, which means Broadcom’s customer base is limited.
Additionally, companies that experiment with ASICs may ultimately decide the costs outweigh the benefits. Antoine Chkaiban at New Street Research says only two companies have deployed custom AI silicon at scale: Google and Amazon. So Nvidia is well positioned to maintain its leadership in AI accelerators. Bank of America analysts estimate it will capture 75% market share in 2030, down only slightly from 80% in 2024.
Looking ahead, Wall Street thinks Nvidia’s adjusted earnings will increase at 34% annually through fiscal 2027, which ends in January 2027. That consensus makes the current valuation of 53 times adjusted earnings look very reasonable. Prospective investors can buy a few shares with confidence, and current shareholders have good reason to optimistic going forward.
John Mackey, former CEO of Whole Foods Market, an Amazon subsidiary, is a member of The Motley Fool’s board of directors. Suzanne Frey, an executive at Alphabet, is a member of The Motley Fool’s board of directors. Bank of America is an advertising partner of Motley Fool Money. Randi Zuckerberg, a former director of market development and spokeswoman for Facebook and sister to Meta Platforms CEO Mark Zuckerberg, is a member of The Motley Fool’s board of directors. Trevor Jennewine has positions in Amazon, Arista Networks, and Nvidia. The Motley Fool has positions in and recommends Alphabet, Amazon, Apple, Arista Networks, Bank of America, Meta Platforms, and Nvidia. The Motley Fool recommends Broadcom. The Motley Fool has a disclosure policy.
Add Comment