The landscape of the global tech industry is undergoing a seismic shift, particularly in the realm of artificial intelligence (AI). At the forefront of this transformation are the US trade restrictions that have placed immense pressure on Chinese companies, especially those reliant on advanced AI hardware. Nvidia, a leading player in the graphics processing unit (GPU) market, has found itself navigating a complex web of export regulations, prompting significant changes in its product offerings and the broader GPU industry.
In response to these restrictions, Nvidia has introduced the H20 GPUs, which are scaled-down versions of its more powerful H100 units. These new chips are designed to comply with stringent export control requirements while still catering to the needs of AI developers. However, the H20 GPUs come with a hefty price tag of around US$10,000 per unit, making them a significant investment for companies looking to leverage AI technology. The limited supply of these GPUs has only exacerbated the challenges faced by Chinese firms, who are in dire need of cutting-edge AI hardware to remain competitive.
The scarcity of high-end Nvidia GPUs has given rise to a burgeoning black market, where products like the H100 and A100 are sold at inflated prices. This underground economy poses a dilemma for companies such as ByteDance, the parent company of TikTok. Given their already tenuous relationship with US authorities, engaging in these illicit markets is not a viable option for ByteDance, which is now seeking alternative solutions to secure its AI hardware needs.
ByteDance Takes Charge: Developing Its Own AI Chips
In light of these challenges, ByteDance has embarked on a bold initiative to develop its own AI chips, aiming to reduce its dependence on Nvidia. In 2024 alone, the company reportedly invested over US$2 billion in Nvidia’s H20 GPUs, highlighting its significant commitment to AI hardware. However, the long-term vision is clear: ByteDance is determined to create its own AI GPUs tailored to its specific needs.
According to reports from The Information, ByteDance is currently working on two distinct AI chips—one focused on AI training and the other on AI inference. These chips will be manufactured using the advanced N4/N5 process technology from Taiwan Semiconductor Manufacturing Company (TSMC), which is also utilized in Nvidia’s latest Blackwell GPUs. This strategic move indicates ByteDance’s ambition to achieve greater self-sufficiency in AI, potentially diminishing its reliance on Nvidia’s offerings.
Broadcom, a company renowned for designing AI chips for tech giants like Google, is set to play a pivotal role in the development of ByteDance’s new GPUs. The chips are expected to enter mass production by 2026, and if ByteDance successfully transitions to its own hardware, it could significantly alter the dynamics of the Chinese tech industry and its dependence on Nvidia.
ByteDance’s Challenges: A Tough Road Ahead
Despite ByteDance’s ambitious plans to develop its own AI chips, the company faces a myriad of challenges. While many Chinese companies have been exploring the development of their own AI hardware, a substantial number still rely on Nvidia for the most demanding computational tasks. Even if ByteDance’s chips prove successful, the company must grapple with the question of whether it can fully sever ties with Nvidia’s technology.
One of the most significant hurdles ByteDance must overcome is software compatibility. Currently, ByteDance utilizes Nvidia’s CUDA platform and software stack for AI training and inference. Transitioning to proprietary AI chips would necessitate the development of a new software platform that seamlessly integrates with its hardware. This endeavor is not only complex but also resource-intensive, potentially taking years to perfect.
However, if ByteDance can successfully navigate these challenges and bring its AI chips to market, it could pose a formidable challenge to Nvidia’s dominance in the AI sector. The involvement of Broadcom in this project is particularly noteworthy, as it signals that major chip designers are closely monitoring the evolving landscape of AI technology and are eager to capitalize on new opportunities.