A trillion-dollar tech company becomes a key partner to help Google, Meta and other hyperscalers build an Nvidia-free AI future
- Broadcom has reportedly secured a major HBM order from SK Hynix
- These chips will likely be used in AI hardware for at least one major CSP
- This is part of a broader trend to reduce dependence on Nvidia
Hyperscalers such as Google, Meta and TikTok owner ByteDance are increasingly seeking to diversify their AI hardware supply chains, reducing their dependence on Nvidia, which has long dominated the sector.
Broadcom is playing an increasingly important role in this shift, and in its recent earnings conference call, CEO Hock Tan told investors that the company has three hyperscale customers that each plan to deploy one million XPU clusters by 2027, and that it has been approached by two additional hyperscalers that are also in advanced development for their own AI XPUs.
While Broadcom didn’t name any names, it is widely believed to be working with Google, Meta, ByteDance and OpenAI on custom AI chips. The company is also thought to be working with Apple to develop the iPhone maker’s first artificial intelligence server chip, codenamed ‘Baltra’, which will provide the advanced networking technologies essential for AI processing.
Bad news for Nvidia
According to The ElectricBroadcom has now approached South Korean memory giant SK Hynix to supply the HBM it wants to use in custom AI chips for a “large” (but predictably unnamed) tech company.
The ElectricBroadcom’s sources say that Broadcom has actively pursued SK Hynix to supply its verified HBM solutions and has secured a large order of the in-demand memory, with shipments expected to begin in the second half of 2025.
SK Hynix, Samsung’s biggest memory rival, is a major supplier of HBM to Nvidia, which is unlikely to be happy with this latest news.
To meet increased demand from Broadcom, SK Hynix is reportedly adjusting its production capacity. The Electric says the company will increase its 1b DRAM wafer production (used as a core chip in its HBM) from 140,000–150,000 units to 160,000–170,000 units by 2025. However, there is a downer to this expansion, namely that it could delay production rollout of SK Hynix’s next-generation 1c DRAM as the company prioritizes its immediate production needs.