It seems that the high-level intelligent driving is finally going to run through. Marked by Tesla's self-developed progressive autonomous driving in 2016, practitioners have been fighting for it for nearly 8 years.
On February 20, Tesla's third-party companion app, TeslaScope, detected that Tesla began to push FSD (beta) v12 to non-specific people2.1。In the U.S. and Canada, vehicles equipped with HW3, HW4, and subscribed to or purchased FSD (S 3 x Y) are eligible for push.
teslascope expects the push ratio to be 0Around 5-2%. Based on 400,000 FSD users in North America, this means that 2,000-8,000 random owners will receive an end-to-end FSD(beta) v122.1。
In China, Xpeng Motors officially announced that "the number of cities covered by intelligent driving, the available mileage and the reputation of users are the first in the industry". Later, Huawei asked the official announcement of OTA, and the navigator function is as high as 99% of the available road sections in the country, "it can be used all over the country, and it can be opened anywhere".
It seems that in a few minutes, "autonomous driving" is about to be completed. But the flip side of the rapid change is the stagnant number of users of Tesla's FSD, the "bottomless" investment in computing power, and the unprofitable new energy vehicle industry.
He Xiaopeng, founder of Xpeng Motors, said in a New Year's letter that more than 40% of the R&D budget will be invested in 2024, of which a total of 3.5 billion yuan will be invested in R&D in "AI technology with intelligent driving as the core". "This year is the first year for Chinese auto brands to enter the 'sea of blood' competition, that is, the first year of the knockout competition."
On the same day, Gaohe Automobile announced that it would suspend production and work for six months, and advised employees to find their own way out. In the plan of the founder Ding Lei, Gaohe Automobile is a model project of the parent company, Human Horizons, for the vehicle-road collaboration business.
High-end intelligent driving and autonomous driving are the key cores of the transformation of car companies from automakers with a net profit margin of less than 7% (Toyota in FY2019) to technology-based companies with a net profit margin of more than 20% (Apple in FY2019).
But before the industry is landed and the transformation is completed, how many people will fall in the darkness before the dawn?
Unified understanding, BEV on the bus
2023 is a year to improve BEV, everyone has shifted from traditional 2D perception to BEV, and the performance of BEV is much better than traditional 2D in all aspects. Eatelephant, an autonomous driving expert on the front line, told the Electric Vehicle Observer.
To put it simply, BEV+Transformer realizes the transformation of the view from 2D to 3D, based on projection mapping, estimating depth information from the image, and generating a unified high-dimensional (3D) feature space representation.
The transformer-based BEV paradigm has been published in the academic community for a long time. However, it was not until AI Day 2021 that Tesla gave the best practices of BEV+Transformer in the field of autonomous driving, and this technical paradigm began to be explored in practice.
Tesla uses BEVs (Bird's Eye View) to fuse information from multiple sensors.
The BEV is slow and fast on the bus.
At the beginning of 2023, the "Electric Vehicle Observer" visited to understandThe only Chinese brands that have truly achieved mass production of BEV+Transformer are Xpeng and Huawei. But by the end of the year, many brands such as Zhiji, Ideal, and NIO have realized this set of technical paradigms on the car. And judging from the test drive experience, the gap with Xiaopeng and Huawei is not very large.
The reason for this is that there have been a lot of BEV-related academic outputs at home and abroad, and all basic solutions are based on open source in academia. The main barrier to mass production lies in the production of automated and semi-automatic annotated data. At the same time, at present, most players' BEV mapping is to fuse lidar and vision in the front, output it with a large model, and then cooperate with high-precision maps, the effect can meet the requirements of mass production.
What's next?
Roll Kaejo is roll engineering ability,"A senior investor looks at the gap between the level of intelligent driving in China," he saidBut engineering capabilities are not as valuable as technological advances
The so-called "technological advancement" is the "end-to-end" of Tesla's FSD V12.
Since there is no AI Day in 2023, there is currently no way to know how Tesla will achieve "end-to-end".。The so-called end-to-end of "photon input and execution output" is a complete large model? Or do you eliminate the non-derivable post-processing part and implement a fully derivable multi-model pipeline? Or is there still a small amount of post-processing (e.g. why is SR output?). "Fake" end-to-end?
FSD v12 version after eliminating 300,000 lines of C++*.
There are different opinions. But to be clear, in various tests**, v12 does show quite amazing anthropomorphism and generalization capabilities.
At the beginning of February, "Late Auto" reported that in the direction of end-to-end, the ideal new model will be launched in the first half of this year, and the new architecture will not limit the scope of the road and can be opened nationwide. NIO will launch an end-to-end-based active security feature in the first half of this year.
Compared with BEV, it may be more difficult to catch up end-to-end.
On the one hand,At present, Tesla is the only end-to-end high-end intelligent driving (autonomous driving), and there is too little experience and verification for the industry to bet on. And even Tesla's V12 still has a lot of uncertainty. It is also committed to end-to-end autonomous drivingAI said that V12 may have more security regulations than V11;
On the other hand,In the field of BEV alone, there are still many gaps to be crossed between Chinese players and Tesla.
One is the "radars". The importance of millimeter-wave radar is decreasing as the practice of pre-fusion of vision and lidar becomes more and more pre-convergent and shows more and more capabilities. So it's different from the sound of "de-lidar" in the ** field,The first to be "killed" may be millimeter-wave radars such as corner radars.
The main driving force for lidar is the pressure to reduce costs.
Second, it is "no picture".
Tesla's mapping capabilities.
Tesla's V10 version in early 2022 removed the post-processing part of the lane markings. However, at present, the road topology of Chinese players still requires a lot of rule-based post-processing processes.
In practice, static BEVs converge more slowly than dynamic BEVs. The model needs to be trained on data above a certain order of magnitude before the output static BEV can have basic performance. The scale and accuracy of data are very large, and a perfect data closed-loop system and continuous investment are required.
In 2024, entering the no-map and improving the performance of no-map will naturally be the new battlefield for top players, but there are still differences in methods. One is to strive to improve the output capacity of the model through data training, and gradually replace the post-processing, which is slow but develops towards end-to-end. One is to constantly iterate the rules of post-processing and regulation to adapt to the needs of no map, which is quick to achieve, but the accumulated post-processing** is useless in the end-to-end era.
"Expensive" tracks
Although it seems extremely close, even Tesla has not yet reached the L3 stage of value realization.
In March 2023, Tesla announced for the first time that the number of FSD North American users was 400,000. At the earnings communication meeting at the end of January, Musk still reported 400,000. Compared to the current Chinese rivals, 400,000 units is still a far ahead number, but it also means that FSD sales have almost stagnated in the past year. In 2023, Tesla's sales in the U.S. alone increased by 25% year-on-year4% to 64580,000 units.
There are very few North American users who are willing to spend $12,000 or $199 per month to subscribe to FSD, but there is no upper limit to the investment.
Tesla's computing power construction plan.
In 2023, Tesla will spend $2 billion to expand the scale of computing power. In January, Musk said on social ** that $500 million is a lot of money, but it is only equivalent to 10,000 Nvidia H100 chips. "Tesla will spend more than that number on Nvidia hardware this year. At the moment, the leverage to remain competitive in the field of artificial intelligence is at least billions of dollars a year. "Tesla expects capital expenditures to exceed $10 billion in 2024, a record high.
From a global point of view, the scale of "burning money" of large models and general artificial intelligence is even more amazing. A few days ago, OpenAI's CEO Sam Altman said at the World ** Summit,The plan is to raise $7 trillion to reshape the global semiconductor industry;Nvidia CEO Jensen Huang said: ".The total value of global data centers is currently about $1 trillion, and that number will grow to $2 trillion in the next four to five years. ”
Some of China's top players have already started to pull up their chips even further.
Xpeng Motors Fuyao Intelligent Computing Center.
This year, Xpeng plans to invest a total amount of R&D in "AI technology with intelligent driving as the core".3.5 billion yuan, which is exactly about 500 million US dollars needed for 10,000 Nvidia H100. Previously, the "Fuyao" data center built by Xpeng based on Alibaba Cloud has a computing power of 600 pflops (6 billion floating point operations per second), which is already in the leading position in China.
Li Auto's cash and cash equivalents at the end of Q3 were 885200 million yuan, it is expected that the remaining 100 billion yuan at the end of the year will not be a problem in the account. But the above-mentioned investors said that 100 billion cash feels like it is not enough for the ideal. “On the one hand, they internally judge that the future of intelligent driving will cost a lot of money; Another judgment is that in the next two or three years, everyone may not be profitable, and it is necessary to fight for the future of pocket depth.
"Money" comes from?
The story of intelligent driving cannot be realized yet, but the story of electrification growth is almost over.
At the beginning of the new year, BYD started with 7The Qin PLUS DM-i, which started at 980,000 yuan, fired the first shot of the ** battle in the Year of the Dragon, and Wuling, Chang'an, Nezha, and even Beijing Hyundai quickly followed suit. Whether this is a charge to the base market of fuel vehicles, or a price reduction coerced by fuel vehicles, as strong as BYD, it cannot ignore the consumption of the first war.
At the end of January, BYD released its 2023 performance forecast, with an estimated net profit attributable to the parent company of 29 billion yuan and 31 billion yuan. Combined with the third quarter report, the net profit in Q4 is expected to be 763.3 billion - 963.3 billion, a year-on-year increase of 440%-31.76%, down 267%-7.5%。
With 120,000 more units sold in Q4 than in Q3, BYD ended its profit growth momentum over the past year or so. Obviously, even the "strong" are not immune at the moment.
Ideal for very good at selling cars,Bloomberg terminal sales are generally expected to be 600,000 units in 2024. 800,000 units can be sold ideally, it is nothing more than a matter of profit. The above-mentioned investor said.
For Huawei, although the question world has formed an upward momentumHowever, it is still uncertain whether the alliance with Chery, BAIC, and JAC can replicate the success of Wenjie, and Huawei's project adaptation scale also has an upper limit. What's more important is when you can make money?In 2023, Huawei's semi-annual report disclosed that the revenue of the automobile business was 1 billion yuan, and the loss in 2022 was about 8.4 billion yuan (based on Yu Chengdong's public statement on input and output).
Yu Chengdong personally went to the Cialis factory to supervise the battle.
The undertone of growth is "bloody war".
According to the data of the Passenger Car Association, in 2023, the penetration rate of new energy passenger vehicles will be 3565%, and in December it reached 4016%。This year, it is expected to reach 40% to 11 million units. The above-mentioned investors believe that the upper limit of the new energy penetration rate in China's passenger car market is 70%. In 2023, 2,169Based on the overall sales of 90,000 units, there is still room for growth of 7.45 million units in the future.
The problem is that your 8 million growth is not interesting if it is achieved at the expense of profits. The above-mentioned investors said that China's new energy vehicle companies are also changing their previous following strategy, in order to fight a "bloody road" in the knockout round.
One is the return to diverse SKUs。This year, Ideal will launch 8 new cars, while Xpeng has planned a new product plan of 30 models in three years.
Looking at it today, the single product explosion may only be suitable for Tesla. If you don't have Tesla's brand power, it's hard to replicate this approach. Therefore, whether it is an automobile group or a single brand, this year's product series are more and more like Toyota and Volkswagen, and the SKUs are very detailed and cover different market segments
The second is the imagination of intelligence. Under the expectation of the prospect of fully autonomous driving, the trend of concentration of car companies seems inevitable. ButOnce a company is the first to run through fully autonomous driving, how much room for latecomers to survive is becoming more and more blurred in the current era of chaos
Anything is possible until probability collapses into reality, though.
"The reason FSD can't be sold today is that it's too expensive. So we can observe that if one day FSD lowers the price or pushes a new subscription model, it may be when Tesla has strong confidence in the product and is ready to push it on a large scale. ”said the above-mentioned investor.