A flagship sedan rolls off a line in Hefei carrying two pieces of silicon that did not exist inside this company five years ago. They are not bought from Nvidia. They were designed in-house, taped out during a global chip crunch, and are now the quiet reason a Chinese automaker controls its own assisted-driving (辅助驾驶) roadmap instead of renting it.
Key takeaways
- NIO's Shenji NX9031 (神玑 NX9031) is a 5-nanometre, automotive-grade smart-driving chip with more than 50 billion transistors, designed and owned by the company.
- The chip entered mass production with the ET9 in April 2025 and had shipped in more than 150,000 vehicle sets by February 2026.
- NIO says one Shenji delivers compute comparable to four Nvidia Orin-X chips; two sit in every ET9 for redundancy and fail-over.
- The project started in 2021 after supply shocks; NIO has said it once spent roughly US$300 million a year buying Orin-X parts.
- The chip work now runs as a separate entity, Anhui Shenji (安徽神玑), and is exploring sales beyond NIO plus embodied-robot and agent-inference workloads.
From a procurement line item to a product
NIO (蔚来) is better known for battery swapping and a three-brand line-up than for semiconductor engineering. The Shenji NX9031 changes that narrative. According to NIO's own technical disclosure, the chip uses a 5nm automotive process, packs more than 50 billion transistors (against roughly 17 billion in Nvidia's Orin X), and pairs a self-developed 32-core CPU with an in-house NPU and image signal processor, qualified to the highest automotive safety level, ASIL-D.
The timeline is unusually public for a Chinese automaker. NIO first showed the chip at NIO Day on 23 December 2023, announced a successful tape-out at its NIO IN event on 27 July 2024, and began mass production when the ET9 flagship started deliveries on 23 April 2025. Two Shenji chips are fitted to every ET9; NIO says the dual layout buys both headroom and a backup path as end-to-end driving models grow hungrier for compute.
The "one chip, four Orins" claim
At the 2025 Shanghai Auto Show, founder and CEO William Li (李斌) held up two production samples and made the comparison explicit: a single Shenji performs at the level of about four Orin-X chips, and two of them roughly match a full Nvidia Thor-X. Those are NIO's own benchmarks, not an independent lab result, and they describe system-level equivalence rather than a like-for-like transistor count. What matters for buyers is the implication: NIO no longer needs to stack multiple bought accelerators to reach flagship compute, and it controls the software stack on top.
The strategic logic is cost and certainty. Li has said NIO previously spent about US$300 million a year procuring Orin-X parts at peak, and that self-developed silicon has already saved the company meaningful money while insulating it from allocation risk. With price competition squeezing Chinese car margins, a per-vehicle saving of around 10,000 yuan that NIO has cited is not trivial.
A chip business that wants to leave the mothership
The more interesting move is organizational. Reporting in October 2026 describes the chip effort as a separately run entity, Anhui Shenji (安徽神玑), led by NIO's smart-hardware chief Bai Jian (白剑). The business is exploring supply to outside partners and branching into embodied-robot control and agent-inference workloads — areas where a low-latency, safety-grade accelerator is useful beyond cars.
According to Li An, chief scientist at BrainNet (脑机网), in-house silicon for Chinese carmakers is less a benchmark play than a supply-chain and margin play: the value is controlling roadmaps and cost, not beating Nvidia on a chart.
That framing fits what NIO has actually done. The company is not claiming to out-specify the frontier; it is claiming to own a critical component that it used to import, and to have shipped it at scale. By February 2026 the chip had reached more than 150,000 vehicle sets across the NIO and Onvo (乐道) line-ups, according to public references including Baidu Baike and Securities Times — a volume that turns an R&D project into a manufacturing operation with its own yield, cost and roadmap questions.
What "going external" really means
Selling a safety-grade automotive chip to other automakers is hard. Each customer brings its own sensor suite, operating system and certification burden. NIO's near-term external story is therefore more likely to be licensing know-how, supplying the entity's compute for robots, or acting as a domestic alternative for firms wary of U.S. export limits on advanced accelerators — not a direct Nvidia replacement in foreign cars.
The embodied-robot and agent-inference angles are where the same silicon could find a second life. A chip tuned for real-time perception and safety in a car is plausibly useful for a robot that must sense and act in the physical world. Whether Anhui Shenji can build a real external business remains unproven; the capex and talent bill to stay on a 5nm cadence are steep.
Honest limitations
This article relies on NIO's own disclosures, Chinese state-media reporting (CNR, Securities Times) and public references (Baidu Baike) rather than independent benchmark data. The "one chip equals four Orin-X" and "Thor-X equivalence" claims are NIO's, not third-party measurements. The 150,000-set shipment figure is reported as of February 2026 and has not been audited; later figures citing mid-2026 volumes come from a single state-media source and should be read as directional. Export-control exposure, foundry capacity and the cost of successive nodes are not quantified here.
What readers can do now
- Treat NIO's chip as evidence that Chinese EV makers are verticalizing into silicon; watch whether Onvo and the Firefly (萤火虫) brand also adopt it.
- For suppliers and rivals, track Anhui Shenji's hiring and any outside design wins — that, not a press claim, is the real test of an external strategy.
- For investors, separate the saved procurement cost (real, recurring) from the "chip company" valuation story (unproven, capex-heavy) when reading NIO's margins.