A Shanghai AI Lab affiliated scientist's startup, the first publicly visible Chinese AI native chip design company, raised a tens of millions RMB WuYuan Capital seed round to test design as a service.
Novasilicon (芯星元) wants to do for chip design what TSMC did for chip manufacturing in 1987: turn the work into a service that other companies buy by the job. Per a LeiPhone exclusive interview with Li Linyang, Li closed a "tens of millions of RMB" seed round led by WuYuan Capital for his AI-native chip design startup. LeiPhone calls Novasilicon the first Chinese company to surface publicly in a global cohort of ex-DeepMind and ex-Nvidia alumni startups that have collectively raised more than $500 million in the same niche since 2021.
The company is not building chips. It is building a chip-design service, with AI agents as the labor input. Chip design has historically been the most labor-intensive step in a custom-silicon program, the part that decides whether a startup or a hyperscaler hires a design team or rents one. Novasilicon's bet is that the work can be split between human engineers and a fleet of AI agents that plan, call EDA tools, and iterate on the result.
Linyang per LeiPhone. His co-founders bring twenty-plus-year design resumes: Xue Jianxi (COO) in digital chip design, and Jianqiu in analog and mixed-signal, an early core member of ASR Microelectronics (翱捷科技). The Lab's prior model work, MOSS, the reported first Chinese open-source large language model, and InternThinker, a reasoning model with professional Go ability and a transparent chain-of-thought, is the track record he leans on.
Chip-design data, including RTL, verification, netlist, and layout, is held tightly inside the few companies that do this work at scale, and the open corpora are too small to train on. Novasilicon is therefore training its agents with reinforcement learning rather than imitation learning, the same family of techniques that let AlphaZero teach itself Go from self-play rather than from human game records. LeiPhone reports the team calls the approach "SuperLearner-style" autonomous exploration.
Novasilicon sells non-recurring engineering, or NRE, work to LLM firms, smart-robotics companies, cloud players, and consumer-electronics brands that need a custom chip but do not want to staff a full design team. Li's analogy for this stage is "AI version of Broadcom", with Broadcom's custom-silicon work for Google and Meta as the reference shape. The longer claim is structural: a "Designless + AI DesignHouse" model that platformizes chip design the way TSMC platformized fab capacity, splitting "fabless" one more time so that customers consume design on demand.
Novasilicon says it will ship a standardized backend design product within the year. If that product lands, the platform thesis has something to test. If the company stays in NRE through 2026, the "Designless" label stays a deck slide.
Google DeepMind published AI floorplanning for TPU in 2021. LeiPhone also reports that the OpenAI-Broadcom co-design took around nine months from product definition to tape-out. That nine-month window is the demand pulse Novasilicon is selling into: it is roughly how short the design cycle could get if the AI-agent loop works on a hyperscaler ASIC.
Li told LeiPhone he expects AI to participate in 5% of chip-design workflow in 2025 and 80% by 2030. That is a founder forecast, not a market consensus, and should be read as one. The Fortune Business Insights edge-AI market projection, USD 29.85 billion in 2026 rising to USD 107.86 billion in 2034, which LeiPhone cites, is a third-party figure relayed through a single outlet; both are reference points, not anchors.
Watch the standardized backend product, the first named NRE customer outside the founder's network, and any disclosure of design-cycle time per tape-out. Those are the receipts that turn "Designless" from a category into a layer.