A rising tide didn’t lift all boats in the chip sector on Friday.

Advanced Micro Devices Inc.’s stock fell sharply Friday to rank among the S&P 500’s worst performers on the day. The selloff came after Broadcom Inc. announced that it was adding a fourth customer for its custom artificial-intelligence chips — a sign of continued demand for AI chips, but perhaps not in the way AMD investors may have hoped.

AMD, like Nvidia Corp., makes graphics processing units that are used for AI applications like training and inferencing. But Broadcom’s growing momentum with customers has some on Wall Street thinking that hyperscalers looking for alternatives to Nvidia might turn to internal offerings in partnership with the likes of Broadcom, rather than opting for AMD’s.

Broadcom Chief Executive Hock Tan said on the company’s earnings call Thursday that, in addition to three major customers ramping production of custom silicon, it has secured a fourth customer that is expected to add $10 billion in revenue next year. Although Broadcom did not name its fourth customer, analysts were suspecting it to be OpenAI. The Financial Times reported on Thursday that the AI startup is preparing to produce its own AI chip that was co-designed with Broadcom and is set to ship next year.

The OpenAI win will likely come “at the expense of Nvidia and AMD,” Melius Research analysts led by Ben Reitzes said in a Friday note. OpenAI CEO Sam Altman had joined AMD counterpart Lisa Su at the latter company’s annual Advancing AI event in June, where it unveiled its upcoming MI350, MI400 and MI4500 series of AI accelerators.

But with current AI spending plans looking “so big by so many” companies, the Melius team said fretting about the impact now “seems premature.” Still, shares of AMD fell more than 6% on Friday.

As hyperscalers, or companies operating large-scale data centers, look to reduce dependency on Nvidia’s GPUs, the second choice is coming down to AMD or internal silicon designs, Jay Goldberg, a research analyst at Seaport Research Partners, told MarketWatch.

Broadcom makes application-specific integrated circuits, or ASICs, which are chips that are custom-built for specific uses. 

“The way this is shaping up is that the contest for AI compute is going to be between Nvidia and Broadcom,” Goldberg said.

Next year, Goldberg sees tech giants such as Microsoft Corp. and Amazon.com Inc. looking more toward developing their own chips, which would be made by Broadcom. Alphabet Inc.’s Google also works exclusively with Broadcom on its tensor processing cores, or TPUs. Earlier this week, analysts at D.A. Davidson said Google is seeing growing interest in its TPUs, which are used for machine learning and AI. The D.A. Davidson team said Google could be sitting on a $900 billion opportunity if it spun out its TPU and DeepMind AI research lab together.

Goldberg said that when looking at chip shipments, Nvidia makes up 70% of that figure, and custom chips about 28%. That leaves about 2% or 3% for AMD.

But looking at how much compute power the chips offer for AI, “it’s actually much more heavily weighted towards Broadcom,” Goldberg said. That’s largely due to Google, which Goldberg said currently holds about one-third of the world’s AI-compute capacity.

Meanwhile, in a note earlier this week, Goldberg’s team said conversations with supply-chain players made them worried that AMD will have a difficult time meeting expectations with its AI-chip roadmap, which some on Wall Street are hopeful will help it close in on Nvidia’s dominance. Progress in AMD’s AI-accelerator business looks to be slowing down, they noted, and there are “signs that AMD is struggling to grow orders from the many customers it announced at its AI event this summer.”

The market for AMD’s AI chips “remains challenging with highly demanding customers,” the analysts said. They pointed to concerns that some of AMD’s “headline customers have only purchased evaluation systems that are unlikely to convert into volume orders for at least one generation of the MI systems.”

Although AMD is still “a viable long-term competitor” in the AI-chip market, it still looks far from winning a more meaningful share of the market anytime soon, Goldberg’s team added.

But while he expects AMD’s MI355 GPUs, which are currently rolling out, to underwhelm customers, Goldberg told MarketWatch that the company’s upcoming MI400 and following series of chips, which are expected later next year, will look more competitive to Nvidia’s next-generation Rubin platform that will launch around that time.

And AMD could have an advantage even as hyperscalers start pushing hard on producing their own silicon, he said. 

“Some of those projects will work and some will fail — it’s not easy to do,” Goldberg said. 

By late 2026 and early 2027, AMD could start looking “a lot more attractive” to hyperscalers when internal silicon efforts don’t end up having the results they want, he added.