A Shanghai-based semiconductor startup, Dongfang Suanxin, unveiled its first commercial AI chip, the DF1000, on July 13 in Shanghai, ahead of the World Artificial Intelligence Conference (WAIC) opening in the city on July 17. The company says the chip runs on an entirely domestic supply chain, placing it among a growing crop of Chinese chipmakers trying to work around US export controls on advanced lithography and high-bandwidth memory.

What the chip claims

According to company statements reported by the South China Morning Post and semiconductor research firm TrendForce, the DF1000 delivers 520 teraflops of BF16 compute on a mature 14-nanometer process node, well short of the cutting-edge processes used in Nvidia’s data-center chips. To compensate, Dongfang Suanxin uses 3D hybrid-bonding and wafer-level stacking to place memory directly on top of compute logic, an approach it calls “software-defined near-memory computing.” TrendForce estimates this gives the DF1000 roughly half the raw compute of an Nvidia H100 or H200, but about a third more memory bandwidth than the H200 — 6.4 terabytes per second, plus 900 gigabytes per second of chip-to-chip interconnect. The company says the design also sidesteps the high-bandwidth memory that Washington has restricted China’s access to.

Chairman and CEO Wei Shaojun, a semiconductor industry veteran and Tsinghua University professor, framed the approach as a deliberate break from copying Western chip architecture. “We have to forge a path of our own,” he said at the launch, calling for “independent architecture, original technology, a self-sustaining ecosystem and a secure, controllable supply chain.”

Funding and roadmap

Founded in May 2024, the company has grown to more than 500 employees and closed a Series A+ round in April at a valuation of roughly 12.3 billion yuan (about $1.7 billion), according to Chinese financial media. Backers include the state-run National AI Industry Investment Fund alongside venture arms of Meituan, Xiaomi, JD.com and Didi.

Dongfang Suanxin says mass production of the DF1000 will begin by the end of 2026. It has laid out a roadmap for a DF2000 chip in the fourth quarter of 2026, which it claims will double the DF1000’s performance and surpass the H200, followed by a DF3000 in late 2027 aimed at matching Nvidia’s next-generation B300 accelerator.