Wait, now Broadcom makes GPUs? Nvidia could face an unexpected enemy in China, as ByteDance could use rivals bigger than Intel, AMD, Arm and Qualcomm to design its AI chips

By James On Sep 29, 2024

U.S. trade restrictions have created significant obstacles for Chinese companies, limiting their access to advanced AI hardware needed to remain globally competitive.

Nvidia’s H20 GPUs, scaled-down versions of the powerful H100, are designed to meet export control requirements but still carry a hefty price tag of around $10,000 per unit.

Even at that price, the availability of these GPUs is limited, further exacerbating the problems Chinese companies face. This shortage has led to a thriving black market for Nvidia’s high-end chips, such as the H100 and A100, where prices continue to rise due to overwhelming demand. However, global companies, especially ByteDance – TikTok’s parent company, which is already under intense scrutiny in the US – can afford the legal and reputational risks associated with participating in such illegal markets.

Two AI chips

ByteDance has made significant investments in AI, reportedly spending more than $2 billion on Nvidia’s H20 GPUs by 2024. The informationthe company wants to develop its own AI GPUs to reduce dependence on Nvidia.

The report adds that these chips will include one designed for AI training and another for AI inference, and both will be produced using TSMC’s advanced N4/N5 process, the same technology used for Nvidia’s Blackwell GPUs.

Broadcom, known for its AI chip designs for Google, will reportedly lead the development of these GPUs, which are expected to go into mass production in 2026. Although several Chinese companies have developed their own AI GPUs to reduce dependence on Nvidia, most still rely on Nvidia’s hardware for more demanding tasks. Whether ByteDance can completely switch to its own hardware – and whether it would want to – remains to be seen.

The move will certainly not be without problems. If Tom’s hardware notes: “The company now relies on Nvidia’s CUDA and supporting software stack for AI training and inference. Once it gets started with its AI GPUs, it will need to develop its software platform and ensure its software stack is fully compatible with its hardware.”