AI accelerators: Why Broadcom could bring Nvidia new competitors

Contents

Big tech companies such as Google, Meta and Bytedance are apparently increasingly trying to make their supply chains for AI hardware independent of chip supplier giant Nvidia. What many experts now suspect is based on the business of the semiconductor company Broadcom – and its talks with a major Nvidia supplier.

As Broadcom CEO Hock Tan reported in the conference on the financial results for the fourth quarter of 2024, three hyperscalers among Broadcom's customers are each planning to deploy one million XPU clusters in their data centers by 2027. Talks are also underway with two other hyperscalers that have their own AI XPUs in development. Specific names were not mentioned at the conference –, but there is speculation that these are Google, Meta, Bytedance, OpenAI and Apple.

AI chips: Broadcom sales skyrocket

At least that would be one explanation for Broadcom's massive increase in sales of semiconductor products in the AI sector. The portal The Next Platform has looked at the development over the past three years: In the 2022 financial year, sales here amounted to 1.94 billion, including network chips and compute engine chips that Broadcom's customers developed themselves and then had manufactured by a contract manufacturer. Broadcom in turn works together with Taiwan Semiconductor Manufacturing Company (TSMC), the world's largest chip manufacturer.

In 2023, Broadcom already generated almost twice as much turnover in the same area: 3.8 billion dollars. And the figure Tan quoted in the most recent results conference made many people sit up and take notice: Broadcom earned 12.2 billion dollars, more than three times as much, in the AI chip segment in 2024.

Figures suggest client

Broadcom is keeping the names of its customers a secret, instead only referring to "three hyperscalers", for example. However, it is clear that these customers must have such a massive budget that they can develop their own chips, which Broadcom then manufactures to order. As the circle of these companies is not too large, speculation quickly falls back on Google, Meta and Co. Many of them are still supplied by Nvidia in this area.

According to the speculative reading of the Broadcom figures, however, Nvidia customers could soon become Nvidia competitors. Instead of using the graphics card manufacturer's chips, hyperscalers prefer to develop their own and then have them manufactured by Broadcom and TSMC.

Broadcom talks with SK Hynix

This would be anything but a welcome development for Nvidia. Neither is the fact that, according to a media report, Broadcom has already approached SK Hynix, which is to supply the high bandwidth memory (HBM) components for the AI chips. The South Korean chip giant is Samsung's biggest competitor in the memory chip market and also supplies Nvidia with HBM. Accordingly, SK Hynix already has to adjust its production capacities for the Broadcom order, which could lead to delays in the next generation of SK Hynix's 1c DRAM.

Nvidia's largest customer for AI accelerators is Microsoft, but according to The Next Platform, there is a similar development here too. Microsoft and Amazon Web Services are increasingly turning to the US chip manufacturer Marvell. This could significantly increase the pressure on Nvidia, which is also struggling with US export restrictions for the Chinese market, in 2025.

(nen)