NIO's Ren Shaoqing, co author of ResNet — a landmark 2015 deep learning paper and the most cited 21st century AI paper — says one AI driving model, not a Lite version, runs on 2022 NIOs running NVIDIA Orin because memory bandwidth, not raw chip
NIO head of intelligent driving Ren Shaoqing, a ResNet co-author, parked his own 2022 ET7 outside a Beijing test loop last week and ran the latest world-model OTA on the same car that left the factory four years ago. The drive, he said, was meant to show that a 2022 NIO can still run the 2026 stack.
In a single wave that ended this month, NIO pushed one unified world-model build to more than 700,000 vehicles across its NIO and ONVO (乐道) brands, four electrical/electronic platforms (Banyan, Cedar, Cedar S, Coconut+), two chip families (NVIDIA Orin-X and NIO's in-house Shenqi NX9031), and four sensor configurations. The same model code reached an early 2022 ET7 and a Coconut+ car delivered last week. There is one codebase and one push (量子位, 新浪科技).
Tesla's parallel move, FSD V14 Lite, ships a roughly 15%-size distilled model to HW3 vehicles, some nearly seven years old, because full-scale world models are widely assumed to be too compute-intensive for older automotive-grade silicon. NIO's rollout runs the same full model on hardware built before ChatGPT's release. The company is positioning the result as the payoff of a 2020-vintage bet: that a Transformer-native autonomous-driving stack would fit a four-year-old car, if the silicon and the compiler were designed for it from the start (量子位, 36Kr).
Convolutional neural networks, the previous generation of driving perception, can tolerate relatively low memory bandwidth because their weights are reused many times over a feature map. Transformer self-attention works differently: every token attends to every prior token, and the model keeps a growing "KV cache" of those keys and values, with low data reuse. As context length grows, the bottleneck stops being raw compute and becomes the rate at which the chip can stream those keys and values in and out of memory. Ren framed this as the reason NIO prioritized memory bandwidth over raw tera-operations per second when it designed Shenqi, before ChatGPT's 2023 breakout made Transformer-style driving models the industry default (量子位).
The published Shenqi NX9031 figure, 546 GB/s of memory bandwidth, is described in the company's framing as "industry-high" and as delivering performance comparable to four Orin chips. Both claims are sourced to a NIO-aligned industry report and have not been independently measured in a teardown; the 4× parity comparison rests on a single paper's framing rather than a head-to-head benchmark. They are defensible positions, not settled ones (量子位).
On the software side, NIO uses NVIDIA's stack only down to the CUDA layer; the deployment framework, inference engine, and AI compiler are all self-developed. The in-house compiler does auto operator generation, graph optimization, and multi-layer fusion. Ren said it has cut model deployment cycles from one to two weeks to one to two days, with about a 20% inference-efficiency gain over generic toolchains. The cross-sensor layer is also abstracted: a single network equalizes different camera resolutions, and LiDAR and millimeter-wave radar are treated as "hot-pluggable" interfaces that can be present or absent without changing the model (量子位, 新浪科技 6/23).
The 700,000 deployed cars also form a shadow-validation fleet. NIO positions them as dispatchable idle compute, with weekly proactive-safety validation mileage above 40 million kilometers and cumulative mileage above 100 million kilometers, roughly the equivalent of a thousand test vehicles running for a year. Ren, whose 2015 ResNet paper is the most-cited AI paper of the 21st century by Nature's count, drives one of the oldest cars in the fleet and uses it as a regression check on whether the unified model still serves its original hardware (量子位, NIO news).
The first NWM build was pushed in May 2025, and a major driver-assist update followed in January 2026. The unified 700,000-car wave landed in June (NIO news, CnEVPost). The rollout's most quoted numbers (546 GB/s, 4× Orin parity, 100 million km of validation) all rest on NIO and NIO-aligned industry media; none has been re-measured by an outside analyst, a teardown shop, or an owner-side road test. A teardown that puts a 2022 ET7 and a current Coconut+ car on the same highway in the same weather would settle the question faster than the next paper.