5Y News | Using AI to Disrupt Chip Design, Novasilicon Secures Tens of Millions of RMB in Exclusive Funding from 5Y Capital

Will chip design become the next platform-ized capability in the semiconductor industry, following wafer manufacturing?

Chip design services are approaching an inflection point with AI.

Leifeng Tech has learned exclusively that NovaSilicon (芯星元), a domestic startup, has become the first to enter this space, recently completing its Series A funding — a multimillion RMB round led solely by 5Y Capital.

Unlike traditional design houses, NovaSilicon believes AI is more than just a tool. "We want to fundamentally rethink chip design from the perspective of large language models themselves," founder and CEO Linyang Li told Leifeng Tech.

To truly realize this vision, expertise in AI alone or chips alone won't suffice.

Public records show that Li currently serves as a young scientist at the Shanghai Artificial Intelligence Laboratory. He previously contributed to the development of MOSS, China's first open-source large language model, and led the team behind InternThinker, the country's first model combining professional Go-playing capability with natural-language explanatory reasoning. Since last year, he has also overseen AI for Chip Design initiatives.

The two other co-founders round out the chip engineering expertise: COO Jianxi Xue brings over two decades of digital chip design experience, having worked with multiple prominent domestic AI chip companies; CTO Jianqiu Chen was an early core member at ASR Microelectronics, with more than twenty years of analog and mixed-signal chip design experience.

NovaSilicon aims to serve companies with growing demand for in-house chip development — including large model firms, intelligent robotics companies, cloud service providers, and consumer electronics brands. The company is currently advancing partnerships with foundries, IDMs, and other industry players through NRE (non-recurring engineering) service contracts.

Per its roadmap, NovaSilicon plans to launch standardized back-end design products within the year, further validating its path from engineering services toward productization.

Notably, using AI to design chips is hardly a new proposition. In 2021, Google DeepMind publicly demonstrated AI-completed floorplanning for TPU chips to accelerate design cycles.

Five years later, large model capabilities have advanced dramatically. Abroad, a wave of companies founded by former Google and NVIDIA employees has emerged, collectively raising over $500 million to apply AI-driven approaches to chip design.

As the first such company to surface domestically — and an AI-native one at that — what does NovaSilicon intend to do, and how does it plan to do it?

AI DesignHouse: Targeting the Incremental Market for AI-Designed Chips

Over the past six months, chip giants, large model companies, and cloud providers have all accelerated their bets on cloud computing power: Microsoft launched its second-generation custom inference GPU after a two-year gap; NVIDIA began introducing its new LPU architecture; OpenAI partnered with Broadcom to develop ASICs optimized for its own model inference.

Meanwhile, Fortune Business Insights projects that the global edge AI semiconductor market will grow from $29.85 billion in 2026 to $107.86 billion by 2034, with diverse hardware application scenarios driving demand for customized chips.

Li came to realize that the human-centric chip design model can no longer keep pace with the production rhythm of small-batch, multi-category, rapidly iterating chips — and this unprecedented incremental market represents a massive opportunity for AI-driven design.

NovaSilicon takes a broader view: AI's participation across the entire chip design flow will jump from 5% in 2025 to 80% five years from now.

Is NovaSilicon trying to become an AI EDA company? Li first dispelled this most natural assumption: "We are not an EDA company."

EDA remains indispensable infrastructure for chip design, he explained, and current EDA vendors are primarily incorporating AI into their existing tool frameworks.

Rather than replacing these tools, NovaSilicon seeks to directly break through the established paradigm, deploying multi-AI agents to assume human responsibilities — handling design, planning, EDA invocation, and continuous iterative optimization.

If forced to draw an analogy, Li sees NovaSilicon as more akin to an AI-powered Broadcom, serving markets that have historically been difficult to reach or are currently experiencing explosive growth: enterprises with multiple chip design needs but constrained by lengthy cycles and high costs, or companies that never imagined they could have their own custom silicon.

As for deliverables, co-founder Xue explained that clients can either receive design blueprints for their own production, or obtain fully usable chips directly.

The "AI Broadcom" framing answers NovaSilicon's present business model. But when discussing long-term vision, the team prefers another company as reference point.

In their view, 1987 marked the semiconductor industry's first major specialization, as IDMs gradually bifurcated into fabless companies and foundries — with TSMC, the latter's dominant player, making wafer manufacturing an infrastructure accessible to the entire industry.

If TSMC platformized "manufacturing capability," then NovaSilicon — seeking to drive the industry's second great specialization in the AI era — hopes to platformize "chip design capability," pushing fabless companies to evolve further toward "Designless + AI DesignHouse": an increasing number of enterprises would no longer need to build in-house chip design teams, but instead access design capabilities on demand.

No Data Dependency: Reinforcement Learning from the Back End

In Li's view, what truly makes "AI-designed chips" viable as a service is the large model itself.

When agent frameworks like OpenClaw demonstrated AI capabilities in task planning, tool invocation, and multi-agent coordination, it signaled that large models were ready to assume complex engineering tasks — and chip design happens to provide an ideal training ground.

Li believes that every optimization in chip design can be validated through extensive physical simulation, while the EDA toolchain further compresses verification and iteration cycles. Throughout this process, large models can not only participate in design but continuously learn and improve through trial and error.

The OpenAI-Broadcom partnership is a paradigmatic case. The roughly nine-month timeline from product definition to tape-out was underpinned by OpenAI bringing its own large model capabilities into the chip R&D workflow. To some extent, Xue noted, they too are building vertical-domain models.

For this reason, NovaSilicon deliberately avoided the industry's more mainstream imitation-learning approach.

"That path trains or fine-tunes models on historical RTL code, verification data, netlists, layouts, design documentation, and similar materials. It does yield results in text-heavy stages like RTL generation."

Li explained that chip design data is inherently highly confidential, while high-quality open-source data is extremely limited. Sustaining access to sufficiently large datasets becomes a bottleneck constraining large model capabilities instead.

Based on this, NovaSilicon has chosen a reinforcement learning route similar to SuperLearner — building models with general decision-making capabilities that autonomously explore to discover optimal strategies, rather than relying on massive amounts of human-generated data.

NovaSilicon is hardly alone on this path; Google, NVIDIA, and other giants are already pursuing similar approaches.

Meanwhile, a wave of new startups focused on AI-driven chip design has rapidly emerged: core members of Google's AlphaChip team and NVIDIA's chip design AI lead Haoxing Ren have both struck out on their own; Cognichip, targeting foundational models for chip design, attracted Intel CEO Lip-Bu Tan to its board.

Capital is concentrating on this direction as well. Leifeng Tech's analysis shows that just three overseas AI chip design startups — Ricursive, ChipAgents, and Cognichip — have raised over $500 million in disclosed funding.

In Li's assessment, NovaSilicon differs from these companies in both technical approach and business model. His team aims to further integrate general-purpose models, multi-agent coordination, and chip design expertise to ultimately build a platform capable of continuously delivering chip design outcomes.

Yet chip design flows are lengthy and complex; achieving full-process automation remains a high barrier. For now, NovaSilicon is starting with analog back-end layout design and digital back-end place-and-route tasks.

Li told Leifeng Tech that the back end determines "whether it can be manufactured, and whether it will function after manufacturing." Compared to the front end, it offers more explicit engineering constraints and quantifiable feedback — making it the most viable entry point for establishing commercial viability.

"Many people think the front end is simply writing code, but they're only seeing the tip of the iceberg above water. From the perspective of what large models need to solve, the difficulty is fundamentally different," Li said bluntly.

For NovaSilicon, however, the back end is only the starting point. Li revealed that in the coming months, the team will advance construction of a multi-agent cluster covering the full front-to-back flow of AI chip design — and this is precisely the future NovaSilicon intends to capture.