Google’s AI Chips Are Hot. Why Partner Broadcom Is Talking Down Their Success.

Google has used its Tensor Processing Units to train and operate its own Gemini large language models. (Justin Sullivan/Getty Images)

Key Points

• Broadcom’s CEO played down the long-term dominance of its jointly developed AI chips with Google, despite a new order from Anthropic.
• Broadcom aims to avoid overreliance on Google, as Google’s program currently accounts for over 80% of its AI compute sales.
• Broadcom’s CEO believes major customers prefer developing their own AI accelerators rather than relying on Google’s hardware.

Google and Broadcom have scored a big hit with their jointly-developed artificial-intelligence chips. So it might be a surprise to hear Broadcom playing down the prospects for the Tensor Processing Units, that they have collaborated on, to become the dominant AI processor.

Broadcom celebrated receiving another major order for TPUs to AI start-up Anthropic in its earnings report on Thursday. But CEO Hock Tan was eager to state that not all of its customers are enthusiastic about using hardware which ultimately depends on Google and its parent company Alphabet.

“That does not mean our other two customers are using TPUs. In fact, they prefer to control their own destiny by continuing to drive their multiyear journey to create their own custom AI accelerators, or XPU racks, as we call them,” Tan told analysts on an earnings call.

So why is Broadcom talking down its major success with Google?

Broadcom is walking a thin line. While its partnership with Google has been a lucrative one, it doesn’t want to become overdependent on a single partner. At the moment Google’s program probably accounts for more than 80% of Broadcom’s AI compute sales, according to BofA Securities.

That’s a potential vulnerability should Google ever decide that it wants to renegotiate its agreement with Broadcom. In fact, Tan was asked about the threat of “customer-owned tooling”—where a customer like Google takes over the majority of chip production and reduces Broadcom’s role—on the earnings call.

“This concept of customer tooling is an overblown hypothesis, which, frankly, I don’t think will happen,” Tan said.

Tan argued that Broadcom’s big customers—which include Facebook-owner Meta Platforms and TikTok-parent ByteDance—would struggle to compete with the pace of innovation of rival chip companies like Nvidia on their own and don’t want to be beholden to Google’s hardware.

“Moving from GPU [graphics-processing units] to TPU is a transactional move. Going into [an] AI accelerator of your own is a long-term strategic move, and nothing would deter you from that,” Tan said.

That raises questions over recent reports that Meta may use Google’s TPUs in future. The Wall Street Journal reported last month that talks were ongoing between the two companies, even as Meta works on its own AI processors.

Google and Meta didn’t immediately respond to requests for comments early on Friday.

Write to Adam Clark at adam.clark@barrons.com