Shares of Broadcom Inc. briefly declined by more than 4% this morning after a report claimed its lucrative chip partnership with Google LLC might be scaled back in the coming years.

The chipmaker’s stock recouped about half of its losses later in the trading session. The reason is that Google issued a statement reaffirming the companies’ partnership. “Our work to meet our internal and external Cloud needs benefit from our collaboration with Broadcom; they have been an excellent partner and we see no change in our engagement,” the search giant told Reuters.

Google and Broadcom jointly develop the former company’s TPU, or Tensor Processing Unit, chip series. TPUs are specialized processors optimized to run artificial intelligence models. They can be used both to train AI models and to perform inference, or the task of running an already trained neural network in production.

Google publicly detailed its first TPU in 2016 and has since developed four newer iterations of the chip together with Broadcom. The search and cloud giant uses TPUs to power AI-optimized virtual machines in its public cloud.

The report that set off Broadcom’s stock price decline was published by The Information in the early morning hours. The publication cited a source as saying that Google plans to stop sourcing TPUs from Broadcom as soon as 2027. The report claimed that moving to AI chips developed fully in-house could save the search giant billions of dollars per year.

According to The Information’s source, Google executives “set a goal earlier this year” to end the AI chip partnership with Broadcom. The move was reportedly motivated by a disagreement between the companies over the AI chips’ price. Google executives, the report continued, have “extensively discussed” dropping Broadcom.

Google’s reaffirmation of the partnership today not only lifted the chipmaker’s stock, but may have also contributed to the 1.9% share price decline Marvell Technology Inc. experienced in conjunction. The latter company is one of Broadcom’s top rivals. In addition to saying Google may switch to in-house AI processors, today’s report claimed the search giant hopes to replace Broadcom chips it uses to link together servers with Marvell silicon.

According to Reuters, JPMorgan Chase & Co. analysts stated in May that Google has accelerated TPU chip orders. The bank estimated that Broadcom could generate $3 billion in revenue from the partnership this year.

Google’s newest TPU, the TPU v5e, debuted last month. It promises to provide twice the performance per dollar of its predecessor for AI training workloads. Additionally, Google says that the TPU v5e can perform generative AI inference tasks 2.5 times more cost-efficiently than the previous-generation chip.

The basic building block of Google’s TPUs is a processing module known as an MXU. It’s a set of circuits that can perform over 10,000 multiply-accumulate operations per processor clock cycle. A multiply-accumulate operation is a calculation that AI models perform frequently while crunching data.

MXU modules account for the bulk of a TPU’s processing power. According to Google, the rest is provided by two other sets of circuits known as vector units and scalar units. 

The TPU’s vector units are designed to speed up AI models’ activation functions. Those are components of a neural network that decide which neurons should participate in a given calculation. The TPU’s scalar units, in turn, help manage the memory an AI model uses and also perform other auxiliary computations.

In its May note about Google’s AI collaboration with Broadcom, JPMorgan stated the companies are already working on a successor to the latest TPU v5e chip. The next-generation processor will presumably be offered via Google Cloud like earlier processors in the series. Thanks to a recently introduced technology called Multislice, Google Cloud customers can deploy up to tens of thousands of TPUs in a single cluster to support large-scale AI projects. 

Photo: Google