In recent years, major tech companies like Google, Meta, and Bytedance have been working to reduce their dependence on Nvidia for AI hardware. This shift is evident in the business activities of semiconductor company Broadcom and its discussions with a major Nvidia supplier.
Broadcom’s CEO, Hock Tan, revealed in a financial results conference for the fourth quarter of 2024 that three hyperscalers among Broadcom’s clients plan to deploy one million XPU clusters in their data centers by 2027. Discussions are also underway with two more hyperscalers developing their own AI XPUs. Although specific names were not mentioned, speculation suggests these companies could include Google, Meta, Bytedance, OpenAI, and Apple.
The rising revenue of Broadcom in the AI sector supports these speculations. According to The Next Platform, Broadcom’s revenue from semiconductor products in the AI sector was $1.94 billion in 2022. This included network chips and compute engine chips, developed by Broadcom’s clients and manufactured by a contractor. Broadcom collaborates with Taiwan Semiconductor Manufacturing Company (TSMC), the largest chip manufacturer globally, for production.
In 2023, Broadcom’s revenue in this area nearly doubled to $3.8 billion. By 2024, the revenue skyrocketed to $12.2 billion, more than three times the previous year. This surge in revenue has led many to speculate about the potential clients behind these numbers.
While Broadcom keeps its client names confidential, referring to them only as “three hyperscalers,” it’s clear these clients have the substantial budgets needed to develop their own chips. Broadcom then manufactures these chips on their behalf. Given the limited number of companies capable of such feats, speculation often points to Google, Meta, and similar companies. Currently, these companies frequently rely on Nvidia for their needs.
If the speculative interpretation of Broadcom’s numbers holds true, Nvidia’s clients could soon become its competitors. Instead of using Nvidia’s chips, these hyperscalers are opting to develop their own, with manufacturing handled by Broadcom and TSMC.
This development is not favorable for Nvidia. Additionally, Broadcom has reportedly approached SK Hynix, a South Korean chip giant, to supply high bandwidth memory components (HBM) for AI chips. SK Hynix, Samsung’s largest competitor in the memory chip market, also supplies Nvidia with HBM. To fulfill Broadcom’s order, SK Hynix may need to adjust its production capacities, potentially causing delays in the next generation of its 1c-DRAM.
Microsoft, Nvidia’s largest buyer of AI accelerators, is also reportedly shifting its focus. According to The Next Platform, Microsoft and Amazon Web Services are increasingly turning to the US chip manufacturer Marvell. This trend could increase pressure on Nvidia, which is also dealing with US export restrictions to the Chinese market, potentially intensifying by 2025.