Author丨Pan Lei
Editor丨Sea waist
Source丨Midjourney
"The Denza Z9GT is the first to use BYD9000 custom chip, using 4nm process and Armv9 architecture, and the Antutu running score is as high as 1.149 million."
Recently, when DENZA released a new car, it suddenly disclosed relevant information about BYD's "core making".
Almost at the same time, Wei Jianjun, chairman of Great Wall Motors, also said through social media that the self-developed "Bauhinia M100 chip" was also successfully lit.
Prior to this, NIO and XPeng had successively demonstrated the developed "Shenji NX9031" and "Turing" autonomous driving chips.
In addition, Geely Xinqing Technology has also launched related chips.
However, some car companies expressed a different view.
Zhu Jiangming, founder, chairman and CEO of Leaprun, once said that Leaprun once made Lingxin 01, which has a computing power of about 4TOPS, but soon realized that chips are an industry that needs to be supported by ultra-large scale.
"The scale is not large enough, resulting in a very low input-output ratio. Now there are many car companies trying to make chips, I think this is inappropriate, industrial division of labor is always necessary."
According to the "Research Report on the Evolution Trend of Autonomous Driving Software and Hardware Integration" released by Chentao Capital, if the annual shipment of chips developed by car companies is less than 1 million pieces, it is difficult to be economical.
And so far, except for BYD, Tesla, Great Wall, Geely and other giant car companies, few new car companies can exceed one million units in annual sales.
At least two industry insiders believe that these car companies are making cores to promote the "integration of software and hardware" in intelligent driving, but the specific effect remains to be seen.
"There's a huge cost involved."
Investing nearly 10 billion may not produce any benefits
"The R&D cost of chips is high, but as long as it is mass-produced, the unit price will be cheap."
This is the view given by Li Bin, founder of NIO, at the "2024 NIO Technology Innovation Day (NIO IN)" at the end of July.
In his view, the volume of production capacity to achieve scale effect, thereby diluting the cost.
From NIO's own point of view, making chips may indeed cost more money.
In 2023, NIO's R&D investment will reach 13.4 billion yuan, and in the first half of this year, it will be 6.083 billion yuan.
How much money is spent on R&D chips, the outside world does not know.
However, according to Chen Tao Capital's "Research Report on the Evolution Trend of Autonomous Driving Software and Hardware Integration", if you take the 7nm process and 100+ TOPS high-performance SoC as an example, its R&D cost is higher than 100 million US dollars (including labor costs, tape-out costs, packaging and testing costs, IP licensing fees, etc.), and if the price is 100 US dollars and the gross profit margin is 50%, its break-even point is 2 million chip shipments.
In the view of Li Ming, an industry person familiar with chip research and development, the actual cost of developing chips is even higher, perhaps tens of billions of yuan.
Based on this, it is still too early to say that self-developed chips can reduce costs.
Among all chips, the R&D cost of high-end intelligent driving chips is quite high.
He introduced that the R&D cost of such chips includes team, operation, computing power construction, etc.
In terms of teams, a hardware team of at least 100 people is needed, and another 100-200 people are responsible for middleware and toolchains, adding up to a team of about 300 people.
The capitation cost is about 2-300 million per year, plus the same operating expenses, the whole calculation is 5-600 million, "the actual number may be higher".
Then there's the cost of time.
It takes about 3 years from the tape-out of this chip to the production of benefits.
In the past three years, the head and operating expenses alone will need almost 2 billion.
The cost of a tape-out is about tens of millions of dollars, plus the need to buy a software IP license like A72 or A78 from Arm, it will also cost about $100 million.
"Overall, it will take about $100 million for chip casting."
Then there is the problem of data, "high-end intelligent driving is data-driven, so you have to build a computing power center".
At the end of last year, Tesla CEO Elon Musk said that he would invest another $1 billion in Dojo supercomputing by the end of 2024.
"So far, billions of yuan have already been spent," Li said, but at this stage it still does not generate any benefits.
Only when based on this chip, the intelligent driving ability that consumers can perceive is formed, and the car is paid for, can it be regarded as generating value.
"Consumers will only be willing to pay if your OTA is like Tesla, and every update brings value to consumers."
Therefore, for car companies, the account of core manufacturing still needs to be carefully calculated.
The financial situation of "Horizon", a supplier of intelligent driving chips, can also verify Li Ming's statement to a certain extent.
According to the data released by Horizon, in the three years from 2021 to 2023, it lost 2.064 billion yuan, 8.72 billion yuan and 6.739 billion yuan respectively, with a total loss of more than 17.5 billion yuan.
In the same period, the total R&D investment of Horizon also reached 5.39 billion yuan, and 1,478 employees were full-time R&D personnel in the 2,066-person employee team.
It is worth mentioning that the shipment of Horizon's "Journey System" intelligent driving solution has exceeded 6 million sets.
In the end, it is the professional people who do the professional things
"In the future, companies that aspire to make a difference at the AI level may have non-general-purpose chips, that is, proprietary AI chips like Xpeng Turing chips."
According to public information, He Xiaopeng, the head of Xiaopeng Motors, recently stated that the self-developed chip is related to Xiaopeng's desire to be in AI.
Li Bin, founder of NIO, also said that self-developed chips are related to AI.
He emphasized that AI will become the core basic capability of smart electric vehicle companies, so NIO has engaged in the Shenji NX9031 chip.
But he also said that research and development for the sake of research and development cannot be done. "If you want to settle the general ledger, you must have a return".
Overall, these car companies believe that the "universal chips" provided by Nvidia or Horizon cannot meet their needs.
Li Ming also has his own views on this.
He said that many car companies are now using NVIDIA Orin chips for intelligent driving solutions, ideal is in use, Xiaopeng is in use, and Zeekr, Bosch, etc., are also using the same chip.
However, for the same chip, the focus of each car company is not the same.
"For example, ZEEKR, now everyone generally believes that its intelligent driving ability is getting better and better", he believes, this shows that the differentiation of intelligent driving solutions does not lie in the chip itself, but in the software.
He said that the differentiation of high-end intelligent driving solutions is generally reflected in three aspects, including the software development capabilities of car companies, data management capabilities, and iterative closed-loop capabilities of intelligent driving solutions.
"If you achieve effective data management, then the effect of intelligent driving can be differentiated," he says, and this also includes closing the data loop.
"The vehicle-end data can be uploaded, and after uploading, it can be trained in the cloud, and after the training, a new algorithm can be formed, and at the same time it can be issued, so as to continuously promote the iteration of intelligent driving functions," he said, adding that these capabilities are very powerful and also test the iterative cycle ability of car companies.
As for Li Bin and He Xiaopeng's self-developed chips related to AI, Li Ming believes that the reasonable explanation is that they may see the "endgame" of AI applications.
"For example, the application of AI includes VR, embodied intelligence and other different application fields, of course, different AI chips are also needed."
However, he pointed out that so far AI has not been "refined" to specific scenarios at the application level.
"If you want to make a chip, then you must do a deep combination with the application scenario. Because the scene will affect your design," he says, adding that it should be deeply integrated with your own algorithms.
For car companies to make cores, he believes that the reason is that Tesla has run this road through.
"But Tesla runs through, not necessarily other car companies can run through", he said, this is like running a marathon, some people can run, some people may not be able to run.
"In the end, it is still a professional person who does professional things".
The core is inspired by Tesla, but not all of them can be Tesla
Liu Yudong, executive general manager of Chentao Capital, said that car companies must have obtained "inspiration" from Tesla to follow the trend of self-developed intelligent driving chips.
But he stressed that the essential reason is to see why Tesla is doing this.
He believes that new car companies betting on intelligent driving chips must first hope to achieve the fastest product iteration and ensure the best performance of their intelligent driving products.
"Because third-party chips will inevitably bring a lot of waste (such as computing power), or there is no way to meet some of your customized needs. So deeper vertical integration in terms of chips is a big trend."
In this case, especially when "intelligent driving has become the core competence in the second half of the competition", car companies still want more control, and continue to expand this control upstream.
"If you want to ensure the best product experience, you must control the upstream chips."
He believes that the factors that car companies will control must be those things that are related to selling cars or strong user experience.
Because intelligence has become the "anchor" of the product definition of car companies.
The second is cost.
He revealed that the profit margin of the autonomous driving chip provided by the third party is still quite high from the perspective of BOM cost, and the gross profit margin is basically around 50%.
In the eyes of car companies, this part of the value can be held in the hands of self-developed chips when the shipment volume is large enough.
"For example, BYD, with such a large sales volume, it is natural to make chips by itself."
The third reason is that when car companies start core manufacturing projects, there will be some supply chain security considerations more or less, because of the lack of car cores in the first two years, many car companies have lingering fears.
"Especially for state-owned enterprises, this weight will be greater."
Of course, there may be another reason, that is, these car companies make chips, which may also have something to do with market value management.
But he also stressed that not all car companies are Tesla.
To make cores, car companies first need to have the ability to develop their own algorithms, as well as the ability to customize a chip, because the chip serves the software.
"If you don't master the software, you don't have the ability to make this chip yourself," he said, adding that only companies that do a good job of self-developed software and resource algorithms have the ability to customize a chip.
In addition, he does not think that this wave of car companies will become the winner or loser of a new round of market competition.
"At present, the best-selling models on the market may basically have no intelligent driving functions."
As for whether the self-developed smart driving chips of car companies will steal the business of third-party chip suppliers, Liu Yudong said that it is still difficult to judge.
"After all, the core manufacturing of car companies has not been fully verified", he said, although the chip has been taped out, but the car companies have not really made it.
"How economics is it? Can it be sustained in the long term? These things are inconclusive, and the pattern of the division of labor in the industry is also in dynamic change."
This article is the original work of ENTREPRENEUR and shall not be reproduced without authorization, otherwise ENTREPRENEUR will reserve the right to pursue legal responsibility. If you need to reprint or have any questions, please contact [email protected].