NVIDIA to Enter the Robotaxi Market with End-to-End + Reinforcement Learning | Exclusive from 36Kr
NVIDIA expands its business portfolio once again.
36Kr has learned from multiple sources that NVIDIA is incubating a Robotaxi project internally. This decision was announced at a recent All Hands Meeting and will be led by Ruchi Bhargava, a senior director who has been with the company for many years.
People familiar with the matter told 36Kr that the new project will adopt a brand - new one - stage technical route. This technical route only uses a single "end - to - end" neural network. The core is to conduct intensive training on the neural network through a world model formed by simulation technology, which is similar to the route adopted by Tesla's FSD.
NVIDIA released the Cosmos world foundation model in January this year. This platform integrates text, images, videos, and sensor data to generate high - quality synthetic video data that follows physical laws and has been pre - trained with 20 million hours of data.
One of the significances of the Cosmos world foundation model is that it can expand complex data that is difficult to generate in real - world scenarios to improve the upper limit of the capabilities of autonomous driving systems.
This route has received preliminary recognition from the industry. Companies such as Li Auto and XPeng have started to build their own world models.
A source told 36Kr that NVIDIA's foray into Robotaxi is not simply for business expansion but to launch a "technical sample for Robotaxi."
Previously, NVIDIA has cooperated with three automakers: General Motors, Mercedes - Benz, and Toyota. They will jointly develop or build autonomous driving fleets based on NVIDIA's technology. Jensen Huang revealed in May this year that the L4 - level autonomous driving fleet in cooperation with Mercedes - Benz will be launched this year.
However, the Robotaxi project being incubated this time is a brand - new one. "It is expected to invest $3 billion and will be launched in cities in the United States in the future," a person familiar with the matter told 36Kr. In the project team meetings, the project investment and goals are becoming clearer.
Jensen Huang has publicly emphasized on multiple occasions that autonomous vehicles are not only "the first major commercial application of robotics" but also a "trillion - dollar industry."
By focusing on the Robotaxi project, NVIDIA aims to verify its full - link engineering capabilities from GPU chips to physical AI large models through practical projects, so as to more accurately define the infrastructure and ecological standards required for the next - generation "physical AI."
Robotaxi is still in its early stage, and it's not too late for NVIDIA to enter the game
In 2025, the deployment of Robotaxi in the US market is accelerating.
Waymo, a US - based Robotaxi company, added two more cities, including Austin, to its driverless commercial operation scope in 2025 and is preparing for manned testing in six cities such as Denver. As of April 2025, Waymo provided over 250,000 paid rides per week in the United States.
Tesla launched its Robotaxi service to the public in Austin, Texas, and the Bay Area of California in September this year. Some investors pointed out that the number of downloads of Tesla's Robotaxi application on the first day exceeded that of Uber by 40% and was six times higher than Waymo's highest - ever download volume.
In terms of laws and regulations, the National Highway Traffic Safety Administration (NHTSA) plans to propose a revision bill in 2026, aiming to remove the established norms of "having a driver and a physical control area" and allow Robotaxi to have a new vehicle structure without a steering wheel.
Meanwhile, NHTSA also promised to shorten the exemption review cycle for autonomous vehicles from "years" to "months" to accelerate the deployment efficiency of Robotaxi vehicles.
However, the industry is still in a very early stage.
Waymo has about 700 operating vehicles in the United States, and Tesla's initial deployment in Austin is only dozens of vehicles. Moreover, the technical route dispute between Waymo and Tesla has never stopped, and there is still no clear conclusion on how L4 - level autonomous driving technology will develop. The current small - scale boom is more like intensive verification on the eve of commercialization rather than competition in a mature market.
For NVIDIA, its core advantage lies in chips and computing ecosystems rather than directly operating fleets.
"The Robotaxi project can be regarded as NVIDIA's training ground," a person close to NVIDIA's Robotaxi project told 36Kr. The application of large AI models in the embodied intelligence industry has become an industry consensus. "But in terms of know - how, the deployment and tuning of large AI models in autonomous vehicles are still difficult problems, and NVIDIA wants to polish such engineering capabilities."
For NVIDIA, it's not too late to enter the market at the current window period, and it still has a chance to compete for technological dominance.
NVIDIA's autonomous driving is catching up through exploration
NVIDIA entered the field of autonomous driving software development in 2015. Unfortunately, it still doesn't have a high - level intelligent driving software solution that has been successfully mass - produced and installed in vehicles.
In 2020, NVIDIA reached a cooperation with Mercedes - Benz to provide an AI software architecture for Mercedes - Benz's next - generation models, including autonomous driving software solutions and intelligent cockpits.
36Kr previously reported that in June 2024, Mercedes - Benz executives drove vehicles equipped with NVIDIA's and Chinese company Momenta's assisted driving software on a round - trip journey of over a thousand kilometers between Los Angeles and San Francisco. The assisted driving software that NVIDIA had been developing for four years was actually less effective than Momenta's "challenging" software that was tuned in just one month.
As a result, Mercedes - Benz has switched the assisted driving business of multiple models in the Chinese market from NVIDIA to Momenta.
NVIDIA, which has insufficient assisted driving capabilities, also faces challenges in the L4 - level autonomous driving field.
"NVIDIA has been benchmarking against Tesla's FSD internally for a long time," a person familiar with the matter told 36Kr. NVIDIA conducted multiple benchmarking tests this year, and the company was shocked by Tesla's FSD's inter - city operation and the number of take - overs. "On a journey of five or six hundred kilometers, Tesla's FSD was taken over only 1 - 2 times. There is still a large gap between NVIDIA and Tesla at present."
Ruchi Bhargava, an old employee in charge of the Robotaxi project, has rarely appeared in public. From public information, we can only see that she has been an author in many of NVIDIA's autonomous driving papers.
Judging from the current layout, NVIDIA still has obvious gaps compared with leading players such as Waymo and Tesla in terms of talent reserves, mass - production - level autonomous driving algorithm engineering, complex scenario data accumulation, and actual road test experience.
However, as a bottom - layer chip supplier, NVIDIA also has unique advantages in promoting Robotaxi: its self - developed DRIVE Thor chip has a computing power of up to 2000 TOPS, which greatly improves the end - to - end model inference efficiency; in addition, its accumulation in AI training clusters and development toolchains provides a foundation for model iteration.
More importantly, NVIDIA has financial support that most autonomous driving companies can't match.
In the second quarter of 2025, its net profit reached $26.4 billion, while Waymo has cumulatively invested $12 billion to reach its current operating scale. Against the background that AI technology highly depends on capital support, the profits from NVIDIA's chip business can provide sufficient room for trial - and - error and iteration for the long - term development of Robotaxi.