TikTok parent ByteDance to produce 2 custom AI chips on TSMC’s 5nm process in 2026

TikTok’s parent company ByteDance is developing not one, but two new AI GPUs, which are reportedly being manufactured on TSMC’s 5nm process technology and will go into mass production in 2026.

SEE GALLERY – 2 IMAGES

The information comes from The Information, where their sources claim that ByteDance will reduce its dependence on NVIDIA for AI hardware while also complying with US export regulations. ByteDance’s new AI GPUs are in the design phase, with one being used for AI training and the other for AI inference.

ByteDance’s new AI GPUs are set to be manufactured on TSMC’s 4N/5N process nodes, which are similar to the 4NP process node TSMC uses to manufacture NVIDIA’s new Blackwell AI GPUs. The TikTok parent company has reportedly spent over $2 billion purchasing over 200,000 NVIDIA H20 AI GPUs (which cost around $10,000 per H20 AI GPU) this year alone, with many of the AI ​​GPUs yet to be shipped by NVIDIA.

The shortage of NVIDIA AI GPUs and their higher prices are some of the reasons why ByteDance has decided to produce its own AI hardware, even though NVIDIA is tailoring the DGX H20 and other AI GPUs specifically for the Chinese market after US export controls became increasingly stringent.

The NVIDIA DGX H20 AI GPU offers performance of just 296 INT8/FP8 TOPS/TFLOPS and 148 BF16/FP16 TFLOPS in AI calculations, while the full H100 AI GPU offers performance of 3958 INT8/FP8 TOPS/TFLOPS and 1979 BF16/FP16 TFLOPS in AI calculations… more than 10 times faster.

However, the DGX H20 has 96GB of HBM3 memory with up to 4.0TB/s memory bandwidth and supports an 8-GPU, which is why it’s a popular AI graphics card in China… but ByteDance wants to get more performance out of it and clearly doesn’t need to deal with NVIDIA anymore (even after spending billions on NVIDIA AI hardware).

However, ByteDance can develop its own AI hardware, but now… now it relies on NVIDIA CUDA and a stack of supporting software for AI training and inference. ByteDance will have to build its own software platform, and the working software stack is fully compatible with its new hardware.