Chinese AI Startup DFSX Releases Chip to Take on the West
DFSX’s investors include state-backed entities and industrial funds such as a venture-capital vehicle co-founded by Alibaba’s Jack Ma
- Chinese startup Dongfang Suanxin introduced its self-designed DF1000 artificial-intelligence chip in Shanghai.
- The company uses a fully domestic supply chain to bypass restrictions on high-end technology that have hindered Chinese semiconductor production.
- Dongfang Suanxin is valued at roughly $1.8 billion following its latest financing, which includes state-backed entities and industrial funds.
A Chinese startup has stepped out of the shadows with the release of an artificial-intelligence chip, aiming to prove that China can power its AI boom entirely on homegrown technology.
Dongfang Suanxin, or DFSX, introduced its self-designed AI chip—the DF1000—on Monday in Shanghai, alongside a product roadmap promising a second-generation chip by the end of this year and a third iteration in 2027, it said in a social-media post.
According to materials published by DFSX, it is using a fully domestic supply chain to bypass restrictions on high-end technology that have hindered Chinese firms’ ability to produce cutting-edge semiconductors.
Founded in 2024, little has been known about DFSX until recently, a decision the company said it made to “keep its head down and focus on core breakthroughs.”
According to Chinese corporate registry data, DFSX is valued at roughly $1.8 billion following its latest financing. Investors include state-backed entities and industrial funds such as a venture-capital vehicle co-founded by Alibaba’s 9988 -0.90%decrease; red down pointing triangle Jack Ma.
While major players like Nvidia chase raw power with tiny transistors and expensive, highly restricted High Bandwidth Memory, DFSX uses a physical workaround. By stacking custom memory on top of the computing layer like Lego bricks, it eliminates data traffic jams—delivering memory bandwidth without foreign-made HBM.
Designed to handle heavy AI training and inference workloads, DFSX says its products overcome three bottlenecks: It doesn’t rely on cutting-edge manufacturing, breaks through memory bandwidth walls and features a full-stack software ecosystem.
The startup uses a legacy 14-nanometer manufacturing, but claims the DF1000 can match some mainstream Western chips built on far more advanced 4-nanometer processes in certain inference workloads. Though it still trails in training, DFSX says its next-generation DF2000 will close that gap.
Led by Wei Shaojun, a senior central government adviser and professor at Tsinghua University in Beijing, the company’s design marks a departure from the scale-out strategies of chip giants like Huawei Technologies, which compensate for outdated manufacturing by clustering tens of thousands of chips into supernodes.
DFSX is instead commercializing 20 years of Tsinghua research by placing data storage right next to the processor’s brain to eliminate delays, allowing the chip to reshape its internal wiring to match the software it is running.
Tech policy experts say success would prove that Chinese companies can build competitive AI despite restrictions, disrupting a Western-led market.
DFSX’s software-defined architecture aligns with a broader shift from raw horsepower to specialized efficiency. David Bennett, chief executive of Japanese sovereign AI firm ai&, says the transition to application-specific hardware is the industry’s defining trend.
“We can see the macro trend playing out,” he said.