In a significant move, ByteDance, the company behind the popular video-sharing app TikTok, is reportedly developing not one, but two new AI GPUs. These custom-designed processors are expected to be manufactured using TSMC’s advanced 5nm process node and are slated for mass production in 2026. This development signifies ByteDance’s intention to reduce its dependence on NVIDIA for its AI hardware needs.
According to sources from The Information, ByteDance is actively designing two distinct AI GPUs, one specifically for AI training and another for AI inference. The company’s decision to develop its own AI hardware stems from the increasing challenges it faces with NVIDIA’s pricing and supply constraints. This year alone, ByteDance is said to have spent over $2 billion acquiring more than 200,000 NVIDIA H20 AI GPUs, each costing around $10,000. Despite this hefty investment, many of these GPUs are yet to be delivered due to ongoing shortages.
The shortage and high prices of NVIDIA AI GPUs have prompted ByteDance to pursue its own hardware solution. While NVIDIA has made efforts to cater to the Chinese market by tweaking its DGX H20 and other AI GPUs, the US export controls continue to tighten, further motivating ByteDance’s independent development.
ByteDance’s new AI GPUs are reportedly designed to be built on TSMC’s 4N/5N process nodes, similar to the 4NP process node used by TSMC for NVIDIA’s new Blackwell AI GPUs. However, it’s important to note that ByteDance currently relies on NVIDIA’s CUDA and supporting software stack for its AI training and inference operations. The company will face the challenge of building its own software platform and ensuring compatibility with its newly developed hardware.
While ByteDance’s move to develop its own AI GPUs represents a strategic shift, it’s crucial to acknowledge that the company still heavily depends on NVIDIA’s software ecosystem. Building a fully functional and compatible software stack will be a key aspect of ByteDance’s success in this endeavor. This development marks a significant step for ByteDance as it seeks to establish greater control over its AI infrastructure and potentially challenge NVIDIA’s dominance in the AI hardware market. It remains to be seen how successful ByteDance will be in developing and deploying its own AI GPUs, and whether it can overcome the challenges associated with building a complete software ecosystem.
