ByteDance ordered over 100,000 chips this year but received fewer than 30,000 by July, slowing the company’s progress. ByteDance reportedly spent over $2 billion on NVIDIA’s H20 chips in 2024, but it is now determined to reduce its dependence on external suppliers
read more
ByteDance, the parent company of TikTok, is shifting its strategy in response to US trade restrictions that have made accessing advanced AI hardware, such as NVIDIA’s high-performance GPUs, increasingly difficult.
Despite being a major player in the AI race, ByteDance is now avoiding the high costs and limited availability of NVIDIA’s GPUs, including the H20, a scaled-down version of the powerful H100, which costs around $10,000 each.
These restrictions and inflated costs have driven many Chinese companies, including ByteDance, to explore alternatives. With limited legal access to NVIDIA’s cutting-edge chips and the risks associated with engaging in black market trades, ByteDance has set its sights on developing its own AI GPUs.
This decision marks a significant shift for the company, which has historically relied heavily on Nvidia hardware.
Developing AI chips in-house
ByteDance reportedly spent over $2 billion on NVIDIA’s H20 chips in 2024, but it is now determined to reduce its dependence on external suppliers. According to recent reports, ByteDance plans to develop two new AI chips: one designed for AI training and the other for AI inference.
These chips will be produced using the advanced N4/N5 process technology from TSMC, the same that powers Nvidia’s high-end Blackwell GPUs. Broadcom, a company renowned for its AI chip designs, will be leading the development, with mass production slated for 2026.
This move reflects a broader trend among Chinese companies, many of which are attempting to build their own AI hardware to circumvent US restrictions. However, the transition to in-house hardware presents significant challenges. As ByteDance currently relies on NVIDIA’s CUDA software stack for AI tasks, developing an entirely new software platform compatible with its custom chips will be a major hurdle.
Huawei’s surprise entry
In addition to its in-house developments, ByteDance is also turning to Huawei’s Ascend 910B chip, using it primarily for less intensive AI inference tasks. However, sources reveal that ByteDance plans to use the Ascend 910B to train a large-language AI model, despite the chip being less powerful than NVIDIA’s GPUs. The constrained supply of Ascend 910B chips, however, has been a significant obstacle. ByteDance ordered over 100,000 chips this year but received fewer than 30,000 by July, slowing the company’s progress.
ByteDance’s AI technology already plays a central role in its operations, powering its flagship AI model, Doubao, and a range of applications, including a text-to-video tool named Jimeng. Doubao, launched in August 2023, quickly became one of China’s most popular apps, with over 10 million monthly active users. The company’s increasing reliance on AI has made it one of the largest buyers of Huawei’s AI chips.
ByteDance’s AI ambitions
ByteDance’s decision to diversify its chip suppliers and develop its own hardware highlights the growing importance of AI in today’s tech landscape. AI has become a central focus for companies across industries, from gaming to e-commerce, as businesses aim to integrate custom AI models that can enhance their services.
However, ByteDance’s reliance on AI also places it at the forefront of the global competition in this space. While the company continues to be a major buyer of NVIDIA’s H20 chips, it is exploring ways to reduce dependency on US suppliers in light of trade restrictions. Whether ByteDance can fully transition to its own AI hardware remains uncertain, but it’s clear the company is positioning itself to stay competitive in an increasingly AI-driven world.
For now, ByteDance’s AI ambitions face obstacles, with supply issues and technical challenges making it difficult to predict when its new AI model will be fully operational.