A little known Shanghai startup just left stealth mode with a chip built to dodge the exact restrictions Washington designed to stop it.
A little known Shanghai startup spent two years building in near total silence. Then it surfaced. This summer it put up a corporate website, opened a social media account, and told the world what it had been working on: an AI chip that sidesteps the exact machines Washington has cut China off from, and leans instead on stacking memory directly on top of the processor.
The company is Dongfang Suanxin, founded in 2024 and led by Wei Shaojun, a Tsinghua University professor and vice president of the China Semiconductor Industry Association. Its flagship product, the DF1000 series, is a 14 nanometre AI processor that the company says delivers 520 teraflops of BF16 performance, according to a report from the South China Morning Post. That is a modest process node by global standards. Nvidia and TSMC work several generations ahead of it. Dongfang Suanxin is not trying to out engineer them on transistor size. It is trying to route around the problem entirely.
The strategy has a name: software defined computing paired with a 3D stacked near memory architecture. Instead of chasing smaller transistors through extreme ultraviolet lithography, a class of machine the Dutch firm ASML still cannot legally sell to Chinese customers, Dongfang Suanxin stacks memory directly onto the compute die. Data travels a shorter physical distance. The company argues that shortens the same bottleneck EUV was built to solve, without needing EUV at all.
You do not build that kind of operation quietly. Dongfang Suanxin already employs more than 500 people, with research and development branches spread across seven Chinese cities, including Beijing and Shenzhen. It closed a Series A+ round in late April 2026 at a post money valuation of roughly 12.275 billion yuan, or about $1.7 billion. The investor list is telling. The National Artificial Intelligence Industry Investment Fund, a state backed vehicle, sits alongside Yunfeng Capital, the fund co-founded by Jack Ma, and the venture arms of Xiaomi and JD.com. That is not speculative money. That is Beijing and China's biggest consumer tech names betting on the same architecture at once.
Not the Only One Trying #
Dongfang Suanxin is not alone. It is not first either. Huawei has already gone public with LogicFolding, a technique that folds 2D circuits into vertical stacks rather than shrinking them, and used it to push the Kirin 2026 smartphone chip to a 55 percent increase in transistor density over last year's Kirin 9030 Pro, on the same process node. Huawei has said that by 2031 it expects its high end chips to reach transistor density equivalent to a 1.4 nanometre process, entirely without EUV.
Others are attacking the same wall from different angles. Shanghai Atomic Technology activated a pilot line in January 2026 aimed at sub-1 nanometre processes, targeting small batch production by December and larger scale output by 2030. Yuanjiwei, also Shanghai based, says it wants to match 5 nanometre silicon performance by 2029 without EUV. Prinano claims it has validated mass production of photonic chips using nanoimprint methods that skip deep ultraviolet lithography altogether, cutting production costs by as much as 90 percent, according to reporting from Tom's Hardware.
The Real Problem for Washington #
None of these companies can yet match Nvidia's most advanced accelerators head to head. Huawei's Ascend line remains the closest thing China has to a credible Nvidia competitor at scale, and Dongfang Suanxin is entering a field Huawei already dominates domestically. But the direction is consistent across every one of these firms, and that consistency is the real story. Washington's export controls were built around a specific chokepoint: extreme ultraviolet lithography machines that only ASML makes, and that the Netherlands, under US pressure, will not ship to China. Chinese firms are no longer trying to acquire that chokepoint. They are trying to make it irrelevant.
Frankly, that is a harder problem for US policy than it looks. Sanctions built around denying a single piece of equipment work well when there is no other path to the same performance. They work far less well once a country's engineers start redesigning the destination instead of trying to reach it the old way. Dongfang Suanxin's DF1000 will not outrun an Nvidia Blackwell chip on paper. That's not the point. Whether that matters depends on how many AI workloads in China can be served by a slower chip stacked cleverly enough to close the gap, and how much state and private capital keeps arriving to find out.
Also read: Amazon Bets On TwelveLabs To Turn Video Into The Next AI Battleground • Google's Money Just Made Proxima Fusion Europe's Best Funded Fusion Bet • Monzo Cofounder Tom Blomfield Leaves Y Combinator to Join Anthropic's Compute Team