NVIDIA's Super Evolution! Grasping both hardware and software, a major shift in intelligent driving

Last night in Washington, NVIDIA officially evolved into a comprehensive supplier of automotive-grade inference chips, training servers, and integrated hardware-software solutions for intelligent driving.

On one hand, it launched the brand new NVIDIA DRIVE AGX Hyperion 10 autonomous driving platform. The highlight this time is that NVIDIA has integrated a hardware-software solution for the first time.

On the other hand, it confirmed collaborations with Stellantis, Lucid Group, Mercedes-Benz, Foxconn, and Uber, aiming to create "the world's largest L4 autonomous driving fleet."

NVIDIA also provided a relatively clear goal: to deploy 100,000 Robotaxis starting in 2027.

In addition to the hardware-software solution, through close cooperation with mobility platforms and automotive companies, NVIDIA's goal is to grasp both passenger and commercial vehicles, similar to Tesla's Cybercab and FSD layout.

Tesla's FSD is currently open in multiple cities in the United States, gradually advancing towards unmanned Robotaxis by the end of the year, with further expansion of operational scale.

By 2026, Cybercab will go into mass production.

NVIDIA's promotion of L4 autonomous driving is undoubtedly driven by the increasing pressure from major global intelligent driving forces, with Tesla being just one of them.

Moreover, brands like Mercedes and Stellantis are intensifying their efforts in the autonomous driving sector and have achieved some success, such as Mercedes' L4 test vehicles accumulating over 16,000 kilometers in testing in China.

If NVIDIA does not timely bind with automotive companies, the space for whole vehicle cooperation may become even smaller.

From an industry perspective, NVIDIA's entry will inevitably lead to changes in the competitive landscape of Robotaxis, which is good news for automotive companies looking to develop autonomous driving businesses.

So, what is NVIDIA's latest technical architecture? What upgrades does it have compared to the previous generation? What benefits does it bring from the perspective of automotive companies?

Let's take a look together.

Hardware-Software Integration

As we all know, autonomous driving platforms heavily rely on computing performance. The upgrade of the NVIDIA DRIVE AGX Hyperion 10 platform primarily focuses on computing power.

According to official information from NVIDIA, the Hyperion 10 is equipped with two Thor chips connected via NVLink-C2C interconnection technology.

The computing power of a single Thor chip exceeds 1000 TOPS at INT8 precision, which is double that of the previous generation (Atlan) At the same time, the DRIVE Thor platform can achieve "cabin and driving integration" computing, supporting simultaneous cabin entertainment and autonomous driving calculations, and can allocate all 2000 TOPS for autonomous driving.

The perception hardware has been streamlined compared to Hyperion 9, reducing 2 LiDARs and 8 ultrasonic radars.

The simplification of perception hardware is beneficial for cost control. It is evident that NVIDIA has optimized overall costs, resulting in less pressure on vehicle integration.

In the past, NVIDIA was purely a hardware solution provider, but with Hyperion 10, the situation has changed, integrating the data center hardware, software, and workflows needed for the autonomous driving development process, achieving software-hardware synergy.

For example, in the cloud, the DGX supercomputer generates high-fidelity simulation data using DRIVE Sim for training the DRIVE AV model;

On the vehicle side, the sensor data acquisition system of Hyperion 10 seamlessly interfaces with the debugging interface of the Thor chip.

With Hyperion 10, NVIDIA has not only achieved software-hardware synergy but also opened up a broader collaboration network.

According to official statements, companies such as Lucid, Stellantis, as well as Momenta and WeRide will be among the first to adopt it.

From the perspective of automakers, a full-stack solution is beneficial for cost reduction and efficiency improvement, mainly by compressing R&D cycles and costs, accelerating large-scale implementation.

For NVIDIA, a closed-loop technology ecosystem facilitates the establishment of autonomous driving standards, forming a technological moat to gain a dominant voice in the industry.

Of course, in addition to the absolute dominance in computing power, NVIDIA has also accumulated a sufficiently deep data foundation.

As of October 2025, NVIDIA's cloud platform has accumulated over 5 million hours of real road data, far exceeding Waymo's 2.5 million hours, providing an irreplaceable advantage for the continuous evolution of AI models.

Speaking of AI models, although NVIDIA did not officially announce it at the conference last night, the "Green Giant" known as VLA has officially come to the world— and it seems to be quite different from all other VLAs.

New Chapter for VLA

In addition to the world's largest autonomous driving fleet and the latest generation hardware-based Hyperion 10 intelligent driving solution, NVIDIA quietly opened a new chapter for its self-developed VLA large model last night Alpamayo-R1 (hereinafter referred to as AR1) is the new VLA (Vision-Language-Action) large model recently released by NVIDIA Labs last night.

The core difference between AR1 and other VLAs is reflected in the title of the NVIDIA Labs blog:

"Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail."

In summary, AR1 aligns closely with other VLA large models in terms of core technological direction.

However, in the complex and variable "long tail" scenarios, AR1 has comprehensively optimized practical performance in reasoning, trajectory generation, alignment, safety, latency, and other aspects through new technological innovations.

NVIDIA's official term is "state-of-the-art."

We will have to wait for actual testing to see the effects, but today we will briefly explain the innovations of AR1 from a theoretical perspective.

First, let's review a term: CoT, Chain of Thought. This was the core term that shocked the AI community earlier this year, introduced by Chinese scientist Wei Xuesen.

CoT, which gained widespread recognition in the academic community in early 2022, allows large models to automatically provide reasoning steps, significantly enhancing the interpretability of sufficiently large LLMs.

NVIDIA's innovation on AR1 is also a "chain," officially named the CoC (Chain of Cause) dataset.

NVIDIA believes that effective reasoning for autonomous driving must be based on "causal relationships," and that reasoning itself "must maintain structural alignment with driving tasks."

This requires that reasoning trajectories do not produce lengthy and disordered narratives, as human thinking is inherently low-latency and efficient.

In applying the CoT Chain of Thought, NVIDIA identifies flaws such as "ambiguous behavior descriptions, superficial reasoning, and causal confusion."

Thus, the CoC causal chain of AR1 can be simply understood as NVIDIA enforcing a "clear causal structure" during the reasoning trajectory process.

This enforcement consists of labeling methods such as "selecting key frames" and "separating historical and future segments."

Therefore, it can be seen that the preliminary labeling work for the causal chain is quite important, and the causal chain dataset generated from this can ultimately provide clear supervisory data for decision learning.

Another interesting point is that the large model evaluator used by CoC is GPT-5. Whether the evaluator itself affects the globalization progress of AR1 needs further observation.

Returning to the model itself, AR1, defined as a "modular VLA architecture" large model, is positioned by NVIDIA as "compatible with any existing VLM backbone network."

Currently, AR1 has expanded to a maximum scale of 7B (three tiers: 0.5/3/7B), which has clearly surpassed the mainstream scale of 4B for assisted driving large models from major manufacturers.

With all these innovations, NVIDIA ultimately announced the following results:

AR1's trajectory planning performance in complex scenarios improved by 12%, near-collision rates decreased by 25%, reinforcement learning post-training improved inference quality by 45%, and inference-action consistency enhanced by 37%.

However, NVIDIA also mentioned another point in the paper:

At this stage, if AR1 wants vehicle-grade latency, it needs to run inference on a card at the RTX A6000 ProBlackwell level—this card has an INT8 computing power of up to 4000T, which is about six times that of Thor-U.

Nevertheless, the birth of AR1 signifies an important step forward for NVIDIA's vehicle-grade hardware-software integration strategy.

From being a pure chip supplier a decade ago, it has leapfrogged to a complete closed loop of chips, models, and training servers, currently the only supplier in the world.

All in Robotaxi

Whether it's the Hyperion 10 platform or the VLA large model included within it, NVIDIA is participating in the autonomous driving race in a different manner.

NVIDIA is no longer just a hardware supplier; instead, it is pushing forward autonomous driving through a full-stack solution, collaborating with automakers to accelerate Robotaxi development.

This Tuesday, the new American car manufacturer Lucid announced the adoption of the DRIVE AV platform and its hardware. Interim CEO Marc Winterhoff stated that this system will first be applied to the upcoming mid-size vehicle and will gradually expand to other products.

He also emphasized that due to the rhythm of market launch and funding issues, the decision was made to "not start from scratch," which may be understood as a deep binding with NVIDIA's autonomous driving solutions.

The collaboration between Stellantis and NVIDIA is even more direct.

The former provides vehicle engineering and manufacturing capabilities, while the latter offers autonomous driving solutions and AI computing power support, jointly developing driverless taxis.

The partnership between Mercedes-Benz and NVIDIA marks the transition of L4 technology from the driverless taxi field to private passenger vehicles.

According to the latter, Hyperion 10 will be integrated into the Mercedes-Benz S-Class, and perhaps "boss cars" will come with "private drivers" in the future On the other hand, for Chinese car companies looking to develop Robotaxi businesses overseas, NVIDIA's Hyperion 10 platform will also bring some positive significance.

Currently, most smart cars in China are equipped with NVIDIA's autonomous driving chips. It will likely be easier to integrate NVIDIA's DRIVE AV solution when aiming to create Robotaxis overseas.

A series of layouts and actions clearly demonstrate NVIDIA's determination in the field of autonomous driving. Professional institutions such as Kings Research predict that from 2025 to 2030, Robotaxis will enter a period of explosive growth, with a market size reaching at least $44 billion.

Not only are Tesla and Waymo applying pressure, but there are also Robotaxi companies from China, such as WeRide and LoGo, among others.

Additionally, numerous domestic and foreign car companies will compete for a voice in the Robotaxi arena, with some setting their mass production timelines for Robotaxis around 2026-2027.

With such a vast growth market, NVIDIA, the benchmark for intelligent driving, will clearly not miss this opportunity.

So, do you think NVIDIA has the chance to define the rules of the Robotaxi arena in the future?

Article authors: Mu Yu, Yu Fei, Source: Electric Planet, Original title: "NVIDIA's Super Evolution! Grasping both Software and Hardware, A Major Shift in Intelligent Driving"

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investing based on this is at one's own risk