Author | Wang Ruihao, Editor | Li Yuchen
On July 17, 2026, at the Shanghai World Expo Center during the World Artificial Intelligence Conference (WAIC), NIO's chip subsidiary, Shenji, made its independent debut. It showcased the NX9031 series intelligent driving chips, which have cumulatively shipped over 300,000 units, and launched the "Ruidong" development platform for embodied intelligence. Four days prior, NIO had just announced its participation in the strategic placement of CXMT (ChangXin Memory Technologies), engaging in deep cooperation on automotive-grade DRAM. And a month before that, NIO had just completed what might be the most complex OTA (Over-The-Air) update in the industry's history—simultaneously pushing the latest world model NWM 2.5 to over 700,000 users, spanning two brands, four platforms, two chip solutions, and over a dozen vehicle models.
What the outside world sees is NIO's intensive moves across three fronts: chips, data, and algorithms. However, few notice that these seemingly independent news items actually point to the same underlying logic: as end-to-end and world models gradually become industry consensus, the algorithm itself is no longer the most solid barrier between enterprises. In the second half of intelligent driving, the competition is about who can build a massive infrastructure of engineering, data, computing power, and verification at the foundation, an infrastructure that enables new algorithms to be continuously deployed across brands, vehicle models, and platforms.
01. 2020: From One Person to One Hundred
In August 2020, NIO's intelligent driving team was just getting started, with the goal of achieving mass production of autonomous driving by March 2022. Starting from scratch, it took one year and nine months.
At that time, NIO's first-generation vehicles used Mobileye's solution. The industry was debating two questions: Can L2 really be developed into L4? Should autonomous driving be fully self-developed? Some swore that "L2 can never become L4," while others felt that self-development was too money-burning and preferred using supplier solutions.
However, there was little debate within NIO—they decided on full-stack self-development very early on. The reason was simple: the supplier's perception solution was a black box, and using someone else's stuff was uncontrollable. Mobileye's solution didn't allow algorithm modifications; if a problem occurred, they could only wait for the other party to fix it, and after the fix, they still had to wait for a release. In the smart EV industry, this pace is impossible for fighting battles.
Therefore, the first step was to choose the chip. Before officially launching the self-developed chip process in 2022, when working on the new architecture in late 2020, the team had already anticipated that the future would lead to this step.
At that time, NIO had two core internal predictions: First, in the future, there would definitely be multiple chips to deploy, and the toolchain must possess cross-chip uniformity; Second, the then-industry-standard NVIDIA toolchain, whose development philosophy and architecture design were mainly based on the engineering environment of three to five years ago, could no longer keep up with the evolution pace of algorithms and engineering architectures in 2020 in terms of flexibility and iteration efficiency.
Based on these two points, the team made a key decision—only use NVIDIA down to the CUDA layer, as this is the necessary layer connecting to the hardware, and all software above it would be fully self-developed. This approach could leverage the latest engineering architecture to improve efficiency and pave the way for reusing most of the software stack when switching to self-developed chips in the future.
Thus, NIO started building its self-developed toolchain in 2020, with the goal of making this toolchain universally usable across different chips. But this also meant starting everything from scratch—new chip platforms, new algorithm architectures, new data systems, everything was new.
From August to the end of 2020, in just four months, NIO's intelligent driving team grew from one person to over a hundred. The core members were basically all in place within those four or five months. Their common trait was: this thing had to start from laying the foundation. Subsequently, the Leifeng Network's "Zuolin Youli" official account will launch "The Ren Shaoqing I Know." Welcome to add the author's WeChat: jinyuan-0428 for communication.
At the end of 2020, NIO began working on the first thing—the sensor solution.
At that time, there was no consensus in the industry on how to arrange autonomous driving sensors. Should LiDAR be used? How many lines? Where to place it? How many cameras? How high the pixels? None of these questions had a final conclusion.
The team was discussing the placement of the LiDAR, with three options: the bumper, behind the windshield, and on the roof. The bumper was the safest, not affecting the styling or aerodynamics, but it had the worst performance—low position, short viewing distance, easily damaged in minor collisions, and high maintenance costs. The solution behind the windshield was technologically immature, requiring special glass, and still suffered performance loss. The roof offered the best performance; standing high and seeing far, it was also immune to minor scratches, but it was devastating to styling and aerodynamics.
For electric vehicles, aerodynamics and styling are almost the lifeline. An extra 0.01 in the drag coefficient means a few kilometers less in range. As for styling, needless to say, whether a car sells or not depends on the first impression, which is the styling.
Internal discussions were very intense. Colleagues in the design department faced the most pressure—traditional automobiles had never grown a "bump" on the roof. Some spoke bluntly, saying the design was "ugly" and would ruin the temperament of the entire car.
But William Li said in that discussion, "You have been designing cars for so many years, and the design of the entire automotive industry is too mature. How to place the LiDAR is a rare opportunity for innovation in the industry. If you can do this point well and make others follow you, you might change the entire industry." This also became an encouragement to the team later on.
The weight of these words is best understood by those who do design. For over a hundred years in the automotive industry, it is very difficult to make a breakthrough design, and even harder to make a design that people still think is right and can be preserved ten years later. It must be valid in terms of aesthetics, functionality, technology, and user value.
After that, the progress went smoothly.
Everyone knows what happened later. The watchtower-style LiDAR has basically become the industry mainstream today. From the stunning release of the ET7 in early 2021 to now, when everyone is following suit, this design has been proven right.
This matter deeply touched NIO's intelligent driving team. At that time, the team had only been formed for two or three months, but they truly felt for the first time: when a company decides something is right and sees the direction for the next five to ten years, even if there are problems and conflicts today, they dare to do it, and time will prove everything. This is not just a slogan, but a real decision that happened.
The camera solution was similar. NIO insisted that NT2 should have seven 8-megapixel cameras. There were also different voices internally—more cameras mean higher costs and greater difficulty in data processing. But the team's judgment was: as the industry moves forward, urban functions, traffic light recognition, and side/rear vehicle monitoring will all require higher pixels and more perspectives. If we save money now, we will regret it in two years. Having persisted, looking back today, this decision was correct.
Therefore, on NT2, NIO deployed hardware ahead of time. 8-megapixel cameras, high-line-count LiDAR, and four Orin X chips—these seemed excessive at the time, but today, over four years later, the ET7 can still push the latest version of software and use the world model. The head of NIO's intelligent driving R&D still frequently uses the 2022 NT3 platform ET7, and feels the car getting better with every OTA. This makes him feel that the decision at that time was right.
02. One Year and Nine Months: From Zero to Mass Production
2021 was the year of greatest pressure for NIO's intelligent driving team.
The team had to mass-produce the Orin X solution by March 2022. But they didn't receive NVIDIA's chips until August or September 2021, leaving only six months until mass production.
As the second-generation dedicated intelligent driving chip, Orin's computing power increased by over eight times compared to the first generation, but its architecture and supporting features were all newly designed. At the hardware level, it faced greater challenges in heat dissipation, power consumption, and system stability, and even basic signal issues like camera frame drops were hard to completely avoid. Moreover, NIO had no prior practical experience in mass production based on the NVIDIA system. The upper-level AI toolchain, CPU scheduling system, and supporting supply chain all needed to be built from scratch. More critically, due to the limited project cycle, the team could not follow the traditional layered and progressive development path. They had to simultaneously advance the deployment and debugging of the upper-level system under the premise that the underlying hardware was not yet fully stable and anomalies like frame drops occurred frequently.
During the 2022 Spring Festival, the team basically released versions every day, testing and solving problems late into the night every day. The entire team worked in both Shanghai and Beijing, operating 24/7. That was not overtime; it was going to war.
NVIDIA colleagues also worked overtime with NIO during the New Year—they knew NIO was helping them mass-produce the chip ahead of schedule. In the end, NIO's Orin solution was the first to be mass-produced globally.
But the price of early mass production was that NIO stepped into many pits that others hadn't. Netizens say, "The pits NIO hasn't stepped into, others will have to step into." Actually, the reverse is also true: the pits NIO has stepped into, those who come later don't have to step into. These things were later encountered by peers doing Orin solutions, but NIO was the first to face them, with no prior experience to refer to.
In March 2022, the NT2 Orin solution was officially mass-produced.
Half a year later, NIO brought this solution to Europe. NIO was the first in China, apart from Tesla, to mass-produce a self-developed solution on another continent. The significance of this was not fully realized by the team at the time. Looking back later, they realized that the number of players capable of mass-producing a fully self-developed solution across continents can be counted on one hand. Tesla did it, NIO did it, and the rest are basically traditional supplier models. Chinese technology will inevitably go global in the future, and NIO has already walked a long way on this path.
But mass production was just the beginning.
After March 2022, NIO began to think about where to go next. Orin was already mass-produced, but they couldn't stop there. NT2 was done, what about NT3?
NIO began to realize that the chip matter must be done by themselves. There were three reasons: First, although Orin was good to use, the hardware iteration pace was subject to the control of others; Second, Transformers have increasingly high requirements for memory bandwidth, and general-purpose chips might not be able to keep up with the evolution speed of intelligent driving algorithms; Third, cost and control—self-developed chips are definitely a better solution in the long run.
In early 2022, everyone started discussing the architecture of the next-generation chip. At that time, neural networks were still mainly CNNs (Convolutional Neural Networks). Transformers had just become popular in academia, but no one in the industry dared to say it would become the mainstream. The team began to discuss: if the next-generation chip needs to run on cars for five or eight years, what will the algorithms look like in these five years?
Ren Shaoqing, the head of NIO's intelligent driving R&D, answered: It will move towards a pure Transformer approach.
This judgment seemed a bit radical at the time. Transformers require 8 to 70 times more memory bandwidth than CNNs. If it wasn't the mainstream, spending a high price to do high bandwidth would be a waste. But if it was the mainstream, not doing it now would lead to being choked off in three years, and if doing it, it had to be high bandwidth. A single chip with over 500GB of memory bandwidth, higher than all chips on the market.
In July 2024, the Shenji chip reached BringUp. In March 2025, it was mass-produced. This is among the earliest batch of self-developed intelligent driving chips in the industry, taking the shortest time from tape-out to mass production.
One Shenji can deliver the performance of four Orins. The reason is simple: its architecture is tailor-made for Transformers. The bandwidth is large enough, so running large models is effortless.
But the chip is just the foundation. What truly makes these things work is the organization and algorithms.
03. 2023: When the Industry Retreats from the Third Stage to the Second Stage
In 2023, NIO realized one thing: the industry had changed.
From 2016 to 2021, the intelligent driving industry competed on engineering efficiency. Whoever had more engineers, more test cars, and stronger architects could iterate faster. Huawei relied on this system to catch up from behind. This stage can be called the "Third Stage"—the goals and evaluation criteria were clear, the technical routes had converged, and everyone competed on execution and resource investment.
But starting in 2023, the situation was different. End-to-end emerged, world models appeared, and the technical routes became uncertain again. Ren Shaoqing analyzed that the industry had retreated from the Third Stage back to the Second Stage—everyone knew what the goal was (better intelligent driving) and how to evaluate good or bad (safety, takeover rate), but no one knew which technical route was the correct answer.
The characteristic of the Second Stage is: innovation is more important than stacking people. A small team of 5 to 10 people can produce in one or two months what used to take hundreds of people half a year. The cost of innovation has decreased, but the impact of innovation remains unchanged.
Therefore, at the end of 2023, NIO made a decision: change the organizational structure to a "4x100 relay" model.
The first leg is the pre-research team. They explore multi-directional technologies—as long as the field is valuable and has potential, they can try it without setting boundaries. Failure is not a big deal; the key is to be fast.
The second leg is the mass production team. Relying on the already built data capabilities, mass production experience, and user closed-loop system, they quickly and with high quality deliver the directions validated by pre-research. Their task is not to innovate, but to turn the results of innovation into products.
The third leg is the solution team. They ensure that after the main line is developed, it can be landed on different platforms—NT2, NT3, NIO, ONVO. The sensors and chips for different vehicle models are all different; they need to adapt one set of algorithms to over a dozen vehicle models.
The fourth leg is the delivery team for specific vehicle models. They are responsible for the final tuning and acceptance of each car.
The logic of this model is: one person cannot be good at pre-research, mass production, and cross-platform adaptation at the same time. The pre-research people just run forward; once they figure it out, they hand it over to the mass production team, and then they go run the next leg. The attribution of results is clear, and the team's acceptance is very high.
Some asked NIO: Is this more efficient?
NIO's answer is: This is not a matter of efficiency, but doing what is believed to be right. Technology is the key to survival and requires organizational support. Through multiple attempts and scientific management, the uncertainty of innovation is transformed into certainty.
In early 2024, NIO used this model to create many new features and technologies. Each technology tried different connection methods—some were handover modes, some were single-team modes. The final result was that the handover mode is easier to meet in terms of organizational requirements, has lower requirements for personnel capabilities, and is easier to generalize.
04. 2023-2024: World Models, Why We Need Them
In 2023, the industry started working on end-to-end.
The term "end-to-end" is very interesting. You will find that apart from intelligent driving, other fields—large language models, facial recognition, object detection—have long been practicing "end-to-end." But the intelligent driving industry didn't start a large-scale shift until 2023.
Why? Because the complexity of intelligent driving is too high.
Around 2010 to 2013, about three people in the industry could handle the full stack of facial technology—from detection, key point recognition, feature extraction, to search, login, and beautification, all code was written by themselves. Facial technology is relatively simple; the input is a face, the output is an identity, and the dimensions of change in the middle are limited.
But intelligent driving cannot do this. The input is heterogeneous data from multiple cameras, LiDAR, and millimeter-wave radar; the output is steering wheel angle and throttle/brake force; what needs to be processed in the middle is the infinite variation of the real world. Therefore, intelligent driving must be divided into multiple modules such as perception, planning, and control, with each module breaking through independently. This has been the mainstream practice in the industry for the past decade.
But in 2022 and 2023, the industry realized: the original architecture can be broken, and modules can be integrated to improve efficiency and effectiveness. This is end-to-end.
NIO's start time was around 2023. But in internal discussions, everyone increasingly felt that end-to-end was just the first step; it didn't solve a more essential problem: autonomous driving is an Agent in the real world, and it needs to understand the physical world. And all existing frameworks do not answer this question well enough.
Let's look at language models first. Their input and output are both language, and language is a symbolic system of abstract concepts. Therefore, a language model is essentially a conceptual cognitive model centered on language, and it can only process information at the conceptual level.
But "a picture is worth a thousand words." A person's posture at this moment, subtle facial expressions, cannot be clearly described in ten thousand words of text, but a single picture has it all. The details of the real world are often impossible to describe with concepts.
Humans interact with 3D space and 4D time in the real world. Spatial relationships and temporal changes—these things are very difficult to describe in language, but they are the foundation for the survival and action of intelligent agents.
Therefore, beyond language models, there must be a parallel system, which can be called "spatiotemporal cognition." Its foundation is pan-imagery—2D, 3D, 4D, all are acceptable. Language models process concepts, while spatiotemporal cognition processes the physical world. The two together constitute complete intelligence.
Language models have been unable to solve the problem of spatiotemporal cognition since their inception. This is also why the industry later began to explore world models—essentially to establish spatiotemporal cognitive capabilities parallel to language models.
Spatiotemporal cognition can be broken down into three levels: physical laws (prior knowledge), spatiotemporal understanding, and reasoning (including violation judgment). Only by mastering these three can one understand the operating logic of the real world.
In the second half of 2023, NIO began to develop world models from this direction. In July 2024, it was released to the public for the first time—the first in China, and among the first globally.
When NIO's world model was just mass-produced, it was once called the "King of Small Roads" by the media. This nickname is related to technical choices—the scenarios handled by small roads most need spatiotemporal cognition. For example, when meeting a car, the judgment is not "whether to yield," but "how much to yield"; or whether to stop completely or brake slightly? Move forward five meters or ten meters? This requires both perceiving the dynamic changes in the time dimension (whether the car in front is moving or stopped) and perceiving the distance relationship in the space dimension (how much space is left).
Previously and even now, the majority of industry practices separate the two: the perception or end-to-end module processes short-term information within 3 seconds, while long-term information is maintained by an outer state machine (essentially if-else rules). This leads to the model only being able to process short time sequences; long time sequence scenarios—10 seconds, 20 seconds—will experience fragmentation, making the car's actions intermittent and stuttering.
Whereas the world model can learn and deduce over longer time sequences, almost completely cutting out the outer state machine and achieving integrated processing. This is the key to its good performance on small roads.
Over the past decade, the industry has done only two things to solve the long time sequence problem: first, relying on maps (maps essentially transmit long time sequence information, such as long-term road status and congestion data); second, manually writing if-else rules (presetting "what to do after 5 seconds, 10 seconds").
But this is the biggest difference between humans and machines. When humans drive, they have cross-day memory; if they make a mistake on a certain road once, they basically won't make it again next time. While early neural networks could only process information from one frame, later they could process 3 to 5 seconds, and beyond that, they were overwhelmed. Long time sequences of over 10 seconds remain a blank in all frameworks to this day.
VLA (Vision-Language-Action) and CoT (Chain of Thought) are also trying to solve long time sequences, but they share a common problem: they can only process conceptual information and cannot handle fine-grained spatiotemporal interactions. Ultimately, it still relies on the world model to learn spatiotemporal cognition to achieve intelligent decision-making in the real physical world.
This is something NIO has been doing since 2023.
05. 2020-2022: Data System and the Lie of "Zero-Cost Copying"
Beyond algorithms, there is something more fundamental. NIO thought about it for a long time at the beginning of laying out intelligent driving, and that is data.
In 2019 and 2020, the industry generally believed: data is zero-cost to copy. Suppliers said, "Copy a set of data for me, and I will help you train the model." Car companies felt this was not a loss; after all, data is just sitting there, and copying a set doesn't cost much.
But NIO believed that copied data has no value. The value of data depends on whether it constitutes a Corner Case for the current model. When the model changes, the value of the data changes. The same data is gold for Model A, but garbage for Model B.
Therefore, data is not zero-cost; its cost lies in computing power. To make the model run in different scenarios and find out the Cases it handles poorly, this process consumes computing power. And computing power is a finite, non-renewable resource with high marginal costs.
Based on this cognition, when NIO started building its data system in 2020, the first thing it did was not to store data, but to build a "vehicle-side computing power scheduling system"—it must first be able to schedule the computing power on the vehicle side to get truly valuable data. This system covers not only the cloud but also the vehicle side.
Why is the vehicle side so important? Because data is generated on the car. To find Corner Cases, the model must run in real scenarios. And 99% of the data in real scenarios are Common Cases, which are of no help to model improvement. These invalid data must be filtered out on the vehicle side, and only the 1% that is truly valuable should be sent back.
To this end, NIO built three sets of systems.
The first is the flexible computing power invocation system integrating vehicle and cloud. This is the foundation of the entire data system—allowing it to schedule computing power across platforms and vehicle models, whether NT2 or NT3, whether Orin or Shenji, all in one pool.
The second is the open-loop A/B test on the vehicle side. NIO imitated the Internet's approach, running two complete systems on the vehicle—one for users to use, and the other running the next-generation algorithm, with the user unaware. The traditional approach is one model plus two sets of post-processing code, but it was found that this could not solve the iteration problem of active safety. The industry access indicator for active safety is one false trigger per 100,000 to 200,000 kilometers; testing it once requires 100 cars/day. The AES (Automatic Emergency Steering) system is even more extreme, requiring one false trigger per 2 million kilometers; testing it once requires 10,000 cars/day. Traditional road testing simply cannot iterate this.
With the open-loop A/B test, even systems like AES that require tens of millions of kilometers of testing can achieve iteration once every three days.
The third is automated analysis for takeover data. Now NIO's intelligent driving has over 700,000 cars on the road, with daily takeovers ranging from millions to tens of millions. Manual analysis simply cannot keep up. First, filter out over 99% of invalid data on the vehicle side, only sending back 1% or 0.5% to the cloud, then perform secondary screening in the cloud, and finally only a very small number of samples are left for R&D personnel to confirm. Fully automated, no humans need to look at the data.
These three sets of systems were built from 2020 to 2022. But the effects have only been felt by users in the last two years—because the early stage was laying the foundation, and the building only started being constructed in the last two years.
This data closed-loop system is leading in the industry. Colleagues from other companies, or people who have left NIO, all acknowledge this system when discussing it, with no controversy. Some even say it is more complete than Tesla's—NIO does not evaluate this itself, but at least it shows that what they have built can withstand peer inspection.
06. Safety, and the People Behind It
NIO's values are "liberate energy, reduce accidents." Reducing accidents is a high-priority goal, so NIO's technical investment always prioritizes active safety.
Active safety used to be a thankless job. Users don't feel effective triggers—the car brakes once, and the user might not even notice. But when there is a false trigger, users will complain. Engineer feedback is also slow; seeing no results for half a year, no one wants to do this job.
But now it's different. With the world model and those three sets of data systems, iteration has become faster. Engineers can see the algorithm regression test results every week and know the optimization ratio for the real world. After release, they can also see after-sales data—how many times AEB (Autonomous Emergency Braking) was triggered in this version, and how many people were helped to avoid accidents.
Positive feedback overrides negative pressure.
Data shows that NIO's current active safety testing exceeds 40 million kilometers per week. Different versions, each version tests several million kilometers. If relying on physical cars, it is completely impossible. This is also why active safety can continue to advance; it's not just the performance seen on the test track, but the Corner Cases in the real world are continuously expanding the boundaries.
The safety mileage on large scales is now 6.79 million kilometers; this figure refers to the mileage of airbag deployment under human driving. In the second half of this year, the safety mileage will reach over 8 million kilometers. Insurance claim payout data has dropped by nearly 40% compared to 2024. There are more cars, but major accidents are going down.
A user posted a video online: the ES9 encountered a tow truck on the road, with the tow truck's rear facing backwards. The system triggered AEB, and it was rear-ended by the car behind. This was indeed a false trigger, and NIO saw it. The ES9 is a new platform; the vehicle size is large and the height is high. In some scenarios, other cars wouldn't misjudge, but it precisely would.
But as soon as this Case appeared, the data returned to the cloud and directly entered the training set. The next version of the model knows that it shouldn't brake for this kind of tow truck scenario. This is the value of the data closed loop—every problem encountered will not be encountered in vain. It will make all cars better.
Behind these trainings is the combination of language models and intelligent driving models, with a very clear direction: humans communicate with the car using natural language, and the car understands the intention and executes it.
07. What We Learned in These Six Years
Six years ago, NIO's intelligent driving team started with one person. Today, NIO has a team of over a thousand people, with four platforms and over a dozen vehicle models delivering the same version simultaneously.
Looking back on these six years, Ren Shaoqing also shared three important moments on-site.
The first is about choices. In 2016, when the head of NIO's intelligent driving team got his Ph.D., he had two choices: Microsoft's first high-level offer in China, or starting a business. He chose autonomous driving, which was still very niche at the time, but he judged that the cycle for this matter was 10 to 15 years, and it was worth doing. Now ten years have passed, and it has been changing all the time—technology upgrades, product upgrades, business model upgrades; not a single day is boring.
Joining NIO in 2020 was also based on two judgments: wanting to be closer to users and get product-level feedback; wanting to solve the problem of data value. From 2016 to 2020, progress in the vision field stagnated, and public datasets had problems with poor scenario generalization and high costs for adding new categories. When doing detection in 2016, it was found that models performing well on public data dropped rapidly in real-scenario performance. Adding a new object category required re-labeling all data, with extremely high costs. Whereas humans learn through 3D interaction, which is much more efficient.
He believes that "data volume plus data quality" is the key to breaking through the bottleneck. Only mass-production scenarios can provide sufficient data. This judgment was verified in 2023 and 2024; the increase in data volume indeed broke through the performance bottleneck and ushered in a new AI peak.
The second is about getting things done. The LiDAR selection at the end of 2020, the Shenji chip architecture in 2022, and the organizational change at the end of 2023—every decision was not "safe" at the time. The reason for choosing the "unsafe" path was seeing a further direction.
William Li has had a great impact on NIO's intelligent driving team. His long-term vision and strategic determination taught the team to stick to their own judgments under pressure. The words William Li said—"If you can do the LiDAR point well and make others follow you, you might change the entire industry"—are actually not just about LiDAR, but about how to treat every innovation opportunity.
The third is about value. When Tesla's FSD enters China, everyone welcomes it. At the technical level, it can prove that Chinese companies can go toe-to-toe with the best in the world; at the commercial level, if Tesla can prove that charging for intelligent driving is viable, then the domestic environment will become better. NIO's subscription data is already very close to Tesla's data in the US. This shows two things: the performance has caught up, and users are willing to pay for a good experience.
The value of technology must ultimately fall on the users. Customers at ease, employees at ease, partners at ease—this is the best state of career in NIO's mind.
Some asked NIO what they think of L3 and L4. NIO's answer is: This is not a purely technical issue, but also a regulatory and social issue. NIO is among the first batch of enterprises on the L3 and L4 access list of the four major ministries and commissions, leading the way at the regulatory level. At the technical level, based on the world model and data-driven architecture, safety is still continuously improving. When the regulations are relaxed, the functions will also be complete.
But even if the regulations are not relaxed, what NIO is doing has value—helping users reduce accidents and liberate energy.
Returning to the time point of July 2026. The independent exhibition of the Shenji chip at WAIC, the strategic cooperation with CXMT, and the OTA covering 700,000 users a month ago—these things strung together point to the same thing: NIO has completed the complete puzzle from algorithms, chips, toolchains to data systems in six years. Algorithms will change, terms will change, but the underlying engineering and data capabilities are more enduring things. BEV, OCC, end-to-end, world models, VLA—the names keep changing, but the data system has been stable since day one. It is a closed loop above computing power scheduling.
What truly determines how far a company can go is not how many new terms are released, but whether the underlying system can support these new terms to truly land and allow old users to use them as well.
In these six years, NIO's intelligent driving R&D team has transformed from an academic organization doing algorithms into a comprehensive team that must balance technology, products, and organization. The role has changed a lot, but the underlying drive hasn't changed—doing a long-term challenging thing, and verifying where the boundaries of AI are in the real world.
This exploration continues.
The author of this article has been following NIO for a long time. Welcome to add the author's WeChat: jinyuan-0428 for communication.