NovaSilicon's AI Agents Automate Chip Layout, Backed by Three Funding Rounds in Five Months
Shanghai Startup's CDK Platform Uses RL-Trained Agents to Handle Analog and Digital Physical Design

NovaSilicon, a Shanghai-based AI chip design startup founded in May 2026, has closed a third funding round in five months — this one led by Shunwei Capital, with GGV Capital and Matrix Partners China also participating. The company simultaneously published two open benchmark papers on arXiv, reporting that its purpose-built AI agents outperform general-purpose models, including OpenAI's GPT-6 Astra, on analog layout generation and digital chip floorplanning tasks.
The back-to-back funding rounds and simultaneous research output signal an unusually compressed timeline for a startup aiming to disrupt a field that the semiconductor industry has long treated as largely unautomatable: the physical design phases that convert a circuit schematic into a manufacturable chip layout.
The CDK Concept: A Process Recipe for Chip Design
NovaSilicon's core product is what the company calls a Circuit Design Kit, or CDK — a deliberate riff on the Process Design Kit, or PDK, that foundries like TSMC supply to describe a manufacturing node's rules and device models. Where a PDK tells designers how a process works, a CDK, in NovaSilicon's framing, tells AI agents how to perform design work: take a chip specification as input, produce a GDS layout file as output.
The company offers two CDK products at launch. Vega targets analog design, covering the full workflow from topology selection through schematic conversion, device sizing, simulation, and final layout. Sirius addresses digital physical back-end, handling floorplanning, place-and-route, clock-tree synthesis, physical verification, and timing closure, while also incorporating front-end RTL generation and verification.
Each CDK is powered by specialized AI agents trained using reinforcement learning rather than imitation learning. NovaSilicon says the RL approach allows its agents to learn from EDA tool outputs used as reward signals — sidestepping the need for large proprietary design datasets that would be difficult and expensive to license. The agents do not simply predict what a human designer would do; they optimize toward measurable outcomes verified by the same EDA tools used in production flows.
Benchmark Results: Beating General Models on Specialist Tasks
NovaSilicon's first benchmark paper, "AnalogLayoutWorld: An Open Benchmark for Agentic Analog Layout Design", introduces a 40-task evaluation suite and reports results across 13 AI model configurations, including several general-purpose large language models.
The paper's headline result: NovaSilicon's Novara agent — the analog layout specialist inside Vega — scored 0.8154 on the AnalogLayoutWorld benchmark, compared with 0.7483 for OpenAI's GPT-6 Astra. Novara achieved a 100% LVS (Layout Versus Schematic) pass rate, confirming that every generated layout was structurally valid. On engineering quality, Novara scored 0.816 against GPT-6 Astra's 0.566. Human engineers accepted 72.5% of Novara's outputs directly, the paper reports.
The second paper, "Evaluating AI Agents for Industrial-Level Chip Floorplanning: A Framework and Empirical Study", evaluates floorplanning — the step in digital physical design where designers position large circuit blocks on the chip before detailed routing. The benchmark uses a dataset of 235,727 floorplan instances, 99 hard macros, and an L-shaped design region of 14.123 square millimeters.
NovaSilicon's Nova-SI agent, the floorplanning specialist inside Sirius, achieved a mean score of 57.97 on the benchmark, against a human engineer baseline of 91.88 and GPT-6 Astra's 31.17. Nova-SI's best single output scored 81.09. The agent also demonstrated efficiency gains: Nova-SI used an average of 36.13 million tokens per task, compared with 62.11 million to 286.50 million for general-purpose AI agents evaluated in the same framework. All eight Nova-SI outputs passed geometry verification checks.
From Fabless to 'Designless'
NovaSilicon frames its ambition through an analogy to TSMC's 1987 founding, which pioneered the fabless semiconductor model by separating chip manufacturing from chip design. The company argues that AI agents capable of automating layout work could enable a comparable separation — a "designless" model in which companies commission chip designs without maintaining large engineering teams for the most labor-intensive physical design steps.
Whether that analogy holds depends on how far AI agents can advance into the full design stack. NovaSilicon's current CDK products target the physical back-end — layout, floorplanning, verification — which represents a significant but not complete portion of total chip design work. Front-end tasks such as architecture, micro-architecture, and RTL design remain substantially human-driven across the industry.
The startup's nearer-term revenue model is straightforward: NRE (non-recurring engineering) fees for custom chip design services. Its first product target is UCIe PHY IP — the physical layer of the Universal Chiplet Interconnect Express standard, which defines die-to-die interconnects for multi-chiplet packages and counts AMD, Intel, TSMC, Samsung, Arm, Google, Meta, Microsoft, and Qualcomm among its founding consortium members. Sirius is being applied to 3D Compute-In-Memory designs.
A Crowded Race With Well-Funded Competition
NovaSilicon is not operating in an empty market. ChipAgents, a Santa Clara, California-based competitor founded in 2024, announced in July 2026 that it had expanded its Series A financing to $134 million — bringing in B Capital alongside existing backers including Bessemer Venture Partners, Micron, MediaTek, and Ericsson. The company reported deployments at more than 120 semiconductor companies and 6× ARR growth in the first half of 2026, with its AI agents focused on RTL design, verification, and debugging workflows.
The major EDA incumbents — Synopsys, Cadence, and Siemens — are themselves accelerating AI integration. In September 2026, Synopsys and OpenAI unveiled their AgentEngineer collaboration, with the two companies developing GPT-Synopsys, an AI system targeting full-flow chip design automation within Synopsys's existing EDA toolchain. That partnership directly overlaps with the analog and digital layout territory where NovaSilicon is staking its claim.
The competitive pressure is intensifying precisely because the layout automation problem is finally yielding to AI methods. For years, analog layout in particular resisted automation because its constraints — matching, symmetry, parasitic sensitivity — required the kind of spatial reasoning and iterative judgment that neither rule-based tools nor early machine learning approaches handled reliably. NovaSilicon's approach of training RL agents on reward signals from EDA verification tools, rather than on human-curated training data, represents one of several bets being placed across the industry on how to close that gap.
Funding Trajectory
NovaSilicon's three rounds in five months began at seed stage shortly after the company was incorporated in May 2026, with 5Y Capital leading. A pre-A round followed, backed by Guoxiang Capital along with the Shanghai Future Industry Fund and Shanghai Science and Innovation Group. The third round, which the company has not publicly sized, was led by Shunwei Capital with participation from GGV Capital and Matrix Partners China.
The speed of the fundraising reflects both investor appetite for China-based AI semiconductor tools and a specific window of opportunity. US export controls on EDA software sales to China were imposed in late May 2026 — briefly disrupting access to tools from Synopsys, Cadence, and Siemens, which together hold roughly 70% of the Chinese EDA market — before being lifted in early July 2026 as part of a US-China trade agreement under which Washington rolled back semiconductor software restrictions in exchange for Beijing easing rare earth export approvals. That brief episode reinforced among Chinese investors the strategic case for domestically developed chip design AI tools, even as the immediate market disruption resolved.
Technical Foundation
NovaSilicon's CDK architecture separates specialist agents for each phase of the design flow. Within Vega, the Novara agent handles the layout phase after earlier agents have completed topology selection, sizing, and simulation. Within Sirius, Nova-SI handles floorplanning after front-end steps are completed by companion agents.
The RL training framework gives each agent a structured reward signal based on EDA tool verification outputs — DRC (Design Rule Check) and LVS results, timing sign-off, power analysis — rather than requiring labeled examples of expert-designed layouts. This approach reduces the data dependency that has limited supervised learning approaches in chip design, since most high-quality commercial layout data is proprietary to the companies that created it.
Whether CDK-class agents can extend further up the design stack — into schematic capture, circuit sizing, and eventually architectural exploration — will determine the scope of the designless model NovaSilicon is describing. For now, the company's publicly released results focus on the layout and floorplanning phases where the benchmark frameworks exist to measure progress objectively.



