ByteDance’s 2026 AI Chip Missed Expectations as Team Lead Seeks Exit

ByteDance’s AI chip that completed tapeout in 2026 fell short of performance expectations and has not entered broad use inside the company. The setback comes as AI chip and DPU team lead 施云峰 seeks to leave and hands over work, while ByteDance develops a next-generation product.

ByteDance began assembling its AI chip team in 2019 and chose an application-specific integrated circuit (ASIC) approach, similar to Google’s TPU. The company aimed to tailor the chips to its own AI work and improve computing efficiency and energy use. Its Ada chip is aimed at Doubao, while SeedChip is intended for Seedance. Both remain in operator adaptation: work to make the operations used by AI models run correctly and efficiently on the chips.

Software support remains a hurdle. Nvidia’s GPUs come with an established CUDA toolchain, while ByteDance’s chips require additional work across models, computing frameworks and lower-level software. ByteDance’s business teams need more computing capacity, but have struggled to spare staff for repeated testing of less mature chips. Buying mature GPUs is an easier way to meet that demand.

ByteDance changed the chip division’s reporting structure in February 2026. 王剑 stopped overseeing the business, and the AI chip and DPU team led by 施云峰, along with the CPU team led by 余红斌, began reporting directly to Vice President 杨震原.

The company is also pursuing chip technology from outside its own development teams. A plan to fold an invested chip project into its Picoheart team has not been completed. Separately, ByteDance is working with 昕原半导体 on an AI inference chip that uses ReRAM memory. The latest sample uses a multi-chiplet design intended to bring together more computing capacity, but manufacturing costs and yield remain unresolved.