ByteDance’s AI Chip Chief Resigns as Its TPU-Style Chip Falls Short, LatePost Says

Shi Yunfeng, who built ByteDance’s AI chip and DPU teams from 2019, is handing over his work as a chip taped out this year falls short of internal goals.

WHY IT MATTERS

ByteDance planned over RMB160 billion of AI purchases this year, over half for Nvidia chips and its own chip; a stalled accelerator keeps Doubao and Seedance reliant on outside suppliers.

Shi Yunfeng, who leads ByteDance’s AI chip and DPU (data processing unit, a chip that offloads networking and storage work in data centers) teams, has submitted his resignation and is handing over his duties, LatePost reported as an exclusive on Weibo. The report says the custom accelerator his team designed, built on a pure ASIC architecture modeled on Google’s TPU, completed tape-out in 2026, but that its actual performance fell short of expectations. The team is now working on a next-generation chip.

TechNode carried the report in English on November 13. None of the coverage opened for this story includes a ByteDance statement, and Shi’s next employer has not been disclosed.

The departure matters because the chip program is one of ByteDance’s main levers on compute costs. Reuters reported in February that the company planned to spend more than RMB160 billion on AI-related procurement in 2026, with more than half allocated to buying Nvidia chips, including H200 models, and to advancing its own chip.

ByteDance’s AI chip team has changed hands repeatedly since 2019

Shi joined ByteDance in 2018, before the company had an in-house chip effort. He graduated from Peking University and had worked at MediaTek, Marvell and 华芯通 (Huaxintong), according to LatePost. It says the AI chip team was formed in mid-2019 inside the machine learning systems (MLSys) department of ByteDance’s AI Lab, and that Shi reports to Yang Zhenyuan, a ByteDance vice president who heads the Volcano Engine cloud business.

LatePost and a piece by the self-media account 星河商业观察 that 21财经 republished on November 13 trace the same history. They say the chip group first sat under Guo Chuanxiong in AI Lab, moved in 2022 to the internal technical platform Data under technical director Wang Jian, and in February 2026 began reporting directly to Yang, along with Yu Hongbin’s CPU team. ByteDance’s AI chip team has reported to three different executives and been reorganized twice in about seven years, and the chip it taped out this year missed its targets.

The group is large. Chinese tech outlet 智能涌现 reported in February that ByteDance’s chip staff had grown to more than 1,000 people, with over 500 working on AI chips and about 200 on CPUs.

A custom design leaves ByteDance with a software problem

The pure ASIC route was chosen to suit ByteDance’s own workloads and to save power, but such a chip does not run on Nvidia’s CUDA software, so models, frameworks and low-level code must be reworked after tape-out, the 21财经 piece explains. It says two projects, Ada for the Doubao model family and SeedChip for the Seedance video model, are still at the stage of adapting operators, the basic software routines a model calls, with limited progress, and that the in-house chips have not been widely deployed inside the company.

Reuters reported in February that SeedChip was designed for inference, that ByteDance aimed to receive samples by the end of March, and that it planned to produce at least 100,000 units this year, possibly rising to 350,000, in talks with Samsung Electronics as manufacturer. A ByteDance spokesperson told Reuters at the time: "The information about ByteDance’s in-house chip project is inaccurate." ByteDance has not published production figures for the chip.

The 21财经 piece also lists ByteDance’s bets outside its own team: a 9.5% stake taken in March 2024 in ReRAM memory maker 昕原半导体 (Xinyuan Semiconductor), ties to in-memory computing firm 知存科技 (Witmem) through Source Code Capital, and a planned merger of an invested chip project into its Picoheart team that has not closed. One person close to the chip team told the outlet the in-house effort was partly meant to strengthen ByteDance’s hand in price talks with outside suppliers.

Volcano Engine’s leadership bench is thinning at the same time

Shi is not the only senior technical figure reported to be leaving Yang’s organization. AASTOCKS reported on November 13, citing Chinese media, that Zhang Xin, a Volcano Engine vice president who also runs governance and operations for ByteDance’s open-source committee, is about to leave. Zhang joined ByteDance when it bought his cloud-software startup Caiyun Technology in July 2020.

The exits come as ByteDance leans harder on AI. EastFrontier reported last week that the company plans a Muse-style AI assistant and does not expect frontier parity before 2029, a timeline that depends on securing compute.

Three things will show where the chip program goes next: whether ByteDance names a successor to Shi or folds his teams into another unit, whether the next-generation chip reaches tape-out, and whether SeedChip or Ada appear in ByteDance’s own data centers in volume. Until then, with its in-house chips not yet widely deployed, ByteDance’s Doubao and Seedance workloads depend on chips bought from outside suppliers such as Nvidia.