NVIDIA's Thor chip delays mass production, and XPeng considers putting on hold its adoption. | Exclusive from 36Kr.
Text | Angela Li
Editor | Qin Li Xuan Yang
The successive delays of NVIDIA's flagship in-vehicle AI chip Thor are increasing the risk of losing core customers.
According to 36Kr, Thor was originally planned to be mass-produced in mid-2024, but it has been significantly postponed. "It is expected to be installed in vehicles by the middle of next year, and it is still the entry-level version." This affects the new vehicle product decisions of some domestic car companies. According to multiple sources from 36Kr, XPeng, a domestic car company, is considering suspending the use of the Thor chip for its new vehicle next year.
"Thor has been delayed until everyone's self-developed chips have matured," an industry insider told 36Kr directly.
Previously, at NVIDIA's GTC Global Technology Conference in March this year, XPeng announced that it would use the Thor chip as the "artificial intelligence brain" of its upcoming electric vehicle. Including XPeng's newly launched model P7+, which was also originally considered to be equipped with Thor, but due to the delay, the P7+ ultimately uses NVIDIA's second-generation intelligent driving chip Orin.
An informed person told 36Kr Auto that XPeng is accelerating the installation of its self-developed intelligent driving chip "Turing". Currently, the chip has been taped out, and XPeng is testing and verifying the stability and performance of the chip. "The full-stack NGP has already been running on the chip XP5 (the internal code name of XPeng's chip)."
At the same time, according to 36Kr Auto, NIO has not reserved NVIDIA's next-generation chip Thor for next year. In July this year, NIO announced that its self-developed intelligent driving chip "Shenji NX9031" has been officially taped out. Like XPeng, NIO's new vehicle intelligent driving system next year will be equipped with its self-developed chip Shenji, NVIDIA Orin, and Horizon, but not Thor.
In response to the above information, 36Kr sought verification from the official of XPeng Motors and NIO Automobile, and both parties said they would not comment.
XPeng and NIO were the earliest mass-production enterprises of NVIDIA's first two generations of in-vehicle AI chips. In 2018, NVIDIA's first-generation intelligent driving chip Xavier was released. In 2020, XPeng was the first to mass-produce and install it on the new vehicle P7, achieving the high-speed NOA (High-speed Navigation-Assisted Driving) function with 30 Tops computing power.
In 2021, NVIDIA released the second-generation intelligent driving chip Orin, which was launched on NIO's new vehicle ET7 the following year. The total computing power of the integrated 4 Orin chips is 1000 Tops. NVIDIA Orin has become the mainstream choice in the intelligent driving field.
Although it failed to continue to be the first to be adopted by the two old customers, NVIDIA Thor, which is designed for AI and large models, is still the target of car companies. BYD, Zeekr, Li Auto and other car companies have previously announced that they will adopt the Thor chip.
However, according to 36Kr Auto, although Li Auto is also one of the first batch of mass-production car companies of Thor, it is also developing its own intelligent driving chip, with the project code name "Schumacher". An insider said that Li Auto is pre-researching the next-generation solution VLA (Visual Language Action Model) of "end-to-end" intelligent driving, which will have a better comprehensive effect when matched with its intelligent driving chip to be mass-produced in 2026. "The dependence on Thor may gradually decrease."
Many market institutions have high hopes for Thor, expecting it to bring NVIDIA a second growth curve beyond the data center business. However, NVIDIA's recently announced results for the third fiscal quarter of fiscal year 2025 show that the automotive chip business accounts for only 1% of the total revenue of $35 billion.
Currently, the second-generation chip Orin is still the main product of NVIDIA's in-vehicle AI chip shipments. NIO CEO Li Bin said last year that NIO's intelligent driving chip procurement in 2023 accounted for 46% of NVIDIA's global shipments. Based on NIO's sales of 160,000 vehicles last year and 4 Orin chips per vehicle, NIO's annual purchase volume is 640,000, and NVIDIA Orin's rough sales estimate for last year is 1.39 million.
With the popularization of intelligent driving, the second-generation chip Orin will be the key product for NVIDIA to continue to increase sales and occupy the market. However, the delay of the new product Thor and the loss of customers will add challenges to NVIDIA's in-vehicle AI chip business next year.
Thor's Repeated Delays, Deepening Car Companies' Chip Anxiety
Currently, the popularity of domestic intelligent driving has been pushed to an unprecedented level. As the chip hardware platform for end-to-end, AI large models and other technical bases, NVIDIA's latest generation of in-vehicle AI chips has attracted the attention of the entire industry.
Thor was first unveiled at NVIDIA's GTC Conference in autumn 2022. It is understood that Thor will have two main versions with computing powers of 750 Tops and 1000 Tops respectively.
Finding potential customers and creating standard products is NVIDIA's consistent strategy when entering the market with new products. At the GTC Conference in 2024, car companies such as XPeng, Zeekr, BYD, and Li Auto all announced their cooperation with Thor.
However, the mass production time of Thor has been repeatedly postponed. A person close to NVIDIA told 36Kr that Thor was originally planned to be mass-produced in mid-2024, but it has now been postponed by at least one year.
Nick, the product manager of XPeng P7+, even said on the community platform that the Thor chip does not have a definite SOP time, and "it would be good to see it in 2026."
A person close to NVIDIA told 36Kr that Thor may have encountered problems with the chip architecture.
A senior person in the chip industry pointed out to 36Kr that Thor integrates NVIDIA's latest generation of high-performance GPU architecture - Blackwell, which is specifically designed for Transformer, large language models (LLM) and generative AI workloads.
However, since this year, the mass production of Blackwell chips has encountered many difficulties. The Blackwell chip is manufactured using TSMC's 4-nanometer (nm) process and contains 208 billion transistors. NVIDIA originally planned to ship in the second quarter of this year, but then the shipment was postponed.
Foreign media reported that before mass production, TSMC engineers found a design defect in the bare die connecting the two Blackwell GPUs, which would lead to a reduction in the chip yield or production volume.
"We have a design defect in Blackwell, it is functional, but the design defect results in a very low yield. This is 100% NVIDIA's fault," Huang Renxun previously said publicly.
Although the defect of Blackwell has been solved with the help of TSMC, as a new product, the mass production challenges of Blackwell are not over yet.
Foreign media also reported that the Blackwell chip also uses TSMC's new packaging technology CoWoS-L, and the yield of this packaging technology faces certain challenges; at the same time, the heat dissipation design of the Blackwell chip also has defects.
This means that the mass production of Thor may also encounter similar challenges. And car companies that plan to install the Thor chip also have to adjust their pace accordingly.
Some car companies are already planning to escape and develop their own chips. Previously, the three new car-making companies XPeng, NIO, and Li Auto have been preparing for self-developed chips for several years. If all goes well, they will enter the mass production and installation stage in 2025.
The higher software and hardware efficiency brought by self-developed intelligent driving chips is obvious. Tesla has also verified this. Its HW3.0 hardware launched in 2019 can still support the upgrade to the end-to-end intelligent driving solution after 5 years of release. With the mass production of the self-developed chips of the new forces, NVIDIA Thor, which is still in its infancy, may face a siege.
On the other hand, the current automotive market competition is fierce. The adjustment of each car model of car companies involves all links such as production materials, sales channels, and production capacity rhythm, and a slight move in one part may affect the whole situation. More importantly, the time window for some car companies' key products to seize the market may be only a few months.
In order to seize the market, the new vehicle development time of car companies has been compressed to 12 - 18 months, and the supply chain is required to reduce costs by 10% every year. How much patience will car companies that are used to leading the supply chain and are facing competitive anxiety have for NVIDIA Thor?
"Some supply chain partners are very powerful. For example, CATL and NVIDIA make a lot of money. We don't have much say. We also call for them to give us some of their profits," Recently, at the face-to-face communication meeting of NIO, CEO Li Bin mentioned the topic of supply chain cost reduction.
Revenue Accounting for 1%, NVIDIA's Automotive Chip Needs Stronger Risk Resistance
NVIDIA Thor's predicament is obvious enough. Due to the delay, the original cooperative car companies may turn into competitors; but from the market next year, even if Thor is successfully mass-produced, the market may not be able to open quickly.
2025 will be a big year for the popularization of intelligent driving. After the fierce intelligent driving technology competition this year, both traditional car companies and new forces are expected to bring intelligent driving to the 150,000-yuan-level market next year.
In this market, car companies have stricter control over the BOM (Bill of Materials) cost of vehicle models, and consumers are also beginning to focus on cost performance. NVIDIA has also prepared three chips to respond to changes in the market.
OrinX is the flagship product with a computing power of 254 Tops. Industry-leading players generally use 2 OrinX chips, which can achieve urban NOA and also support OTA upgrade to the end-to-end software solution;
Orin N is the entry-level product with a computing power of about 80 Tops. The expected ceiling of capabilities is urban memory driving, high-speed NOA, etc.
In addition, NVIDIA has also launched an Orin Y version, which is mainly a substitute for Orin X. An informed person told 36Kr Auto that the performance of Orin Y is 80% of Orin X, with a computing power of approximately 200 Tops, but the price is only half of NVIDIA OrinX. "This is the version that NVIDIA is strongly promoting," an industry insider said.
According to public information, Orin is a product that NVIDIA spent 4 years and invested billions of dollars to develop. If Orin is regarded as the main source of revenue, the automotive business revenues in 2022 - 2023 are $566 million and $903 million respectively. Combined with this year's revenue, NVIDIA Orin's revenue may barely cover the research and development costs.
With the popularization of intelligent driving next year, the N/Y in NVIDIA's Orin series will further occupy the market share. This is a common phenomenon in the semiconductor industry. After the release of a new product, the previous-generation product can still maintain a high shipment volume due to various factors such as price, performance, market demand, and supply situation.
Lowering the price can also bring greater shipment volume and revenue figures. In November this year, NVIDIA announced the results for the third fiscal quarter of fiscal year 2025 (August 3, 2024 - October 27, 2024). The total revenue for this quarter is $35.1 billion. The revenue of the data center business accounts for nearly 90%, while the automotive chip business department only accounts for 1% of the total revenue.
Fortunately, the automotive chip business in the third fiscal quarter increased by 72% year-on-year and 30% quarter-on-quarter. This growth is mainly due to the increased demand for self-driving car chips and the chips NVIDIA sells for robots.
However, in this market, domestic supply chain players such as Horizon have been deeply rooted in this market for a long time. It may not be easy for NVIDIA to obtain high gross margins in the strict cost control of car companies.
Moreover, for NVIDIA, the situation Thor faces is quite different from when Orin was launched.
When Orin was launched, it was at a stage when domestic car companies were fully developing intelligent driving, and the domestic players were relatively blank; the emergence of Thor actually caters to the needs of the era of large computing power in vehicles, but obviously Thor faces more uncertainties and more domestic alternative solutions. If the intelligent driving chips of the new forces are really successfully installed, the probability of the intelligent driving industry's disenchantment with NVIDIA Thor will increase.
Another challenge for NVIDIA is that its intelligent driving software solution capability is relatively lacking. In the more than one year since the former intelligent driving head of XPeng, Xinzhou Wu, took office, he has also been leading the team to catch up with the progress of intelligent driving. According to 36Kr Auto, the first-generation solution of NVIDIA's intelligent driving focuses on being able to drive nationwide without relying on high-precision maps.
"It is still in the stage of running through 0 - 1. Xinzhou Wu is very hardworking and gets on the car to test the demo almost every day, and the demo needs to be updated every day," a person close to NVIDIA told 36Kr. But this is the path of the industry's leading intelligent driving players a year ago, not the currently popular "end-to-end" intelligent driving solution.
The relatively backward progress of intelligent driving has made NVIDIA's intelligent driving customers start to consider more solutions. Its customer Mercedes-Benz has already simultaneously enabled the intelligent driving solution of the domestic intelligent driving company Momenta. The end-to-end intelligent driving solution of this intelligent driving company has been implemented in the models of BYD Denza, SAIC IM 智己, and GAC Toyota.
36Kr previously exclusively reported that The new Mercedes-Benz CLA model will be equipped with Momenta's end-to-end intelligent driving solution. "The number of models that have now obtained cooperation has expanded again," the source said.
Of course, NVIDIA Thor is still a typical representative of the end-side computing power base in the era of AI large models, and the AI landing experience it has obtained in the AI large model wave can still be reused. Car companies without self-developed chip solutions that want to catch up with Tesla will still consider Thor.
But if NVIDIA has a more closed-loop software and hardware integration capability, it will undoubtedly bring greater automotive business revenue and a deeper insight into the domestic intelligent driving market. NVIDIA's automotive business segment, which accounts for 1% of revenue, still needs a stronger risk resistance ability.