文 | HiEV大蒜粒车研所，作者 |本一，编辑 |德新
本周在北京世界智能网联汽车大会，华为引望 CEO 靳玉志、地平线 CEO 余凯以及 Momenta CEO 三大国内智驾供应商罕见同台，同个圆桌上还有蔚来 CEO 李斌、小鹏汽车 CEO 何小鹏以及比亚迪高级副总裁、新技院院长杨冬生，对谈由清华大学汽车产业与技术战略研究院院长赵福全主持。
在这场对话里，对自动驾驶走到哪一步了，三家智驾供应商的掌门人竟然给出了三种不同的判断。
华为高级副总裁、引望 CEO 靳玉志给出的判断是正从辅助驾驶加速迈向自动驾驶。
地平线创始人兼 CEO 余凯认为，「革命才刚开始」。在他看来接下来，还要跨好几次「太平洋」。行业顶级水平的城区 NOA 刚刚做到 100 公里一次关键接管，而 Robotaxi 需要的是 10 万公里级别。
Momenta CEO 曹旭东的判断则较为乐观。他认为智驾能力的提升速度，已经从过去的每两年十倍，加快到了每年十倍。
三人在一件事上看法一致：下半场要用可量化的安全数据说话。靳玉志拿出的是累计里程和相对人驾的安全数据，余凯用的是 MPCI，曹旭东则把订阅率的增长归因于安全性的提升。
分歧则主要集中在两个问题上：技术进步到底有多快，以及供应商在下半场应该站在什么位置。
一、下半场来没来：一年十倍，还是要再跨几次太平洋
赵福全在开场就定了一个明确的前提基调，2026 年是自动驾驶的拐点，上半场靠技术突破，下半场要解决怎么走进千家万户、怎么让企业赚钱。
他请每位嘉宾谈谈对「下半场」的理解。
引望 CEO 靳玉志对下半场的判断没有落在技术速度，而是放在了产业节奏上。他列了三点变化。
第一是辅助驾驶在加速迈向自动驾驶，L2 及以上的渗透率已经到了 70%，工信部在有序组织 L3 的准入，自动驾驶的强制性国标已经发布，明年 7 月 1 日实施。
第二是安全要求变高。L2 阶段出了事故，责任主体还是司机；到了 L3 以上，责任会转移到系统。
第三是出行方式会被重构。靳玉志描述的场景是，到了自动驾驶时代，你去上班之后，车会自己去停车、充电、洗车、接送小孩，甚至给你买杯咖啡。
余凯的回答则最谨慎。他说自己从事技术研发，每天的感受都是「这么多问题要解决，这个事情真难」，所以第一反应是「革命才刚开始」。
他用 MPCI，也就是平均多少公里出现一次关键接管，作为划分阶段的依据。
按照他的判断，中国行业顶级水平的城区 NOA 刚刚跨过 100 公里一次接管。用户在 4S 店试驾半小时碰不到一次接管，就会觉得可用，十万元级的车也开始把城区 NOA 当作标配，所以他也同意2026 年是一个拐点。
但从 100 公里到 1000 公里、再到 1 万公里，余凯认为没有现成答案。他的疑问是，沿着现在的路线多采数据、多训模型，是不是就能实现下一次跨越。他给出的参照标准是，特斯拉在城区大约 1000 公里接管一次，真正做 Robotaxi、做 L4 和 L5，需要 10 万公里以上。
余凯把每一次十倍量级的提升比作跨越一次太平洋，并给出了一个阶梯式的演进路线：现在是可用，1000 到 1 万公里是好用，10 万公里才是爱用。他希望行业能在未来四到五年里跨过去。
曹旭东也认同 2026 年是上半场和下半场的分水岭，但对技术进化的速度与余凯判断不同。
Momenta 内部有一个「智驾摩尔定律」，原来的说法是每两年提升十倍。
他说 2022 年到 2026 年基本符合这个规律，四年里提升了大约百倍，智驾从几乎不可用的玩具变成了开始可用、渐渐好用。现在，他认为下半场这条定律要更新为每年提升十倍。
根据 Momenta内部对特斯拉 FSD 的测试评估：从 V13 到 V14，一年多的时间里安全性提升了十几倍到几十倍。按照这个速度，他判断 2027 年、尤其是 2028 年，会是 L3 和 L4 的重要拐点。
他还转述了当天中国科学技术协会名誉主席万钢主旨报告中的预测：L3 渗透率在 2028 年达到 15%，2030 年达到 35%。
曹旭东甚至提出一个听起来就很卷的「行业标准」：到 2028 年，即使是法规和责任上仍属 L2++ 的辅助驾驶，产品体验也需要达到 L4 级别的安全。
车企一方，何小鹏的看法更接近余凯，能不能做到万公里或十万公里一次安全接管、能不能每年提高十倍，「我觉得也很难」。
他举的例子是，汽车在雪地、沼泽、草地行驶时，需要把底盘运动控制和自动驾驶完全融合，他认为一两年做不到。
二、下半场的两把尺子：安全数据与订阅
责任转移到系统之后，要论证安全就得拿出「铁证」。
靳玉志说，基于这个考虑，引望今年年初开始向全行业实时公布辅助驾驶里程和相对人驾的安全数据。他提到，到对话当天，乾崑智驾的累计里程应该超过了 170 亿公里。
从用户的使用情况来看，今年国庆七天，乾崑智驾用户每天用辅助驾驶行驶 8000 万公里，七天累计 5.6 亿公里，辅助驾驶的里程占比达到 60%。
另一把尺子是「用户愿不愿意付钱」。
这一点曹旭东讲得最具体。
他预判面向 L3 的订阅模式会更加普遍，理由是用户价值真正创造出来之后，支付意愿也会变强。
他以特斯拉为例：FSD V13 时，北美的订阅率大约在 20% 左右。到特斯拉最新一期财报，这个数字已经到了 55%。他把增长归因于安全性的提升。国内方面，他提到华为和蔚来在订阅模式上做了比较早的探索。
李斌认为，自动驾驶落地看两个指标：多少用户愿意付钱，以及能不能用量化的安全数据，包括保险公司的反馈，让用户感知到自动驾驶确实更安全。
何小鹏则提出了下半场的三个新指标：安全的碰撞里程、全场景覆盖率和用户付费率。
李斌也给出了蔚来的数据。按照他的说法，搭载 1000 多 TOPS 算力的蔚来车型已经有 95 万辆，辅助驾驶累计里程超过 91.4 亿公里，人车共驾的安全里程是单独人驾的 8 倍，四家保险公司出具的证明显示，重大事故下降了 30% 多。
赵福全在收尾总结时也指出：消费者已经被免费升级「惯坏了」，但订阅背后是企业的研发成本和利润空间。
三、L3的分歧最大：是过渡，还是跳过？
赵福全在第二轮追问三家车企 L3 的时间表时，三个人的说法差别很大。
何小鹏认为应该跳过 L3，把 L2 和 L4 放在一起考虑。
他举的例子是奥迪在欧洲少数高速公路上开放的 L3，他试了一天，只激活了几分钟。他的判断是，L3 某种程度上把责任转移给了系统，「但是体验没有优化，安全没有提高」。他给出的时间是，L4 的技术落地大概在 2028 年下半年。
杨冬生的看法相反。他认为 L3 是一个大门槛，安全要求「不是倍增，可能是一个几百倍的要求」，限定场景的 L3 有希望在 2028 年落地。
李斌则把技术和合规分开看。
合规上，GB 44721—2026《智能网联汽车 自动驾驶系统安全要求》将在 2027 年 7 月 1 日实施；技术上，他认为世界模型加闭环强化学习已经是行业共识，L2 到 L5「技术是同源的」。至于蔚来什么时候量产 L3，他的回答是「取决于国家什么时候允许」。
曹旭东则给出了第三种解法。他说自己在欧洲体验 L3 的感受，和何小鹏完全一样。
Momenta 的一个欧洲客户在中国做 L3，一开始沿用的是欧洲的思路。
随着中国 L3 的市场和政策逐渐明朗，双方改为用中国 L4 的思路来做 L3：传感器、芯片域控和执行器都按 L4 的架构设计，L3 法规就绪时先上 L3，等 L4 法规成熟，再通过软件升级到 L4。
曹旭东的判断是，L3 的过渡周期会比较短，很快会升级到 L4。按照他的说法，这样客户只需要开发一套 L4 架构的方案，就能同时覆盖 L3 和 L4，开发成本大幅降低。上市的 L3 也会带来和欧洲 L3 完全不同的体验，「更准确地说，是刚才小鹏总说的那个 L4 的体验」。
Momenta 的思路和何小鹏的「跳过 L3」在目标上相近，都是把 L4 的能力放进量产车。区别在于，它没有回避 L3 的法规节点，而是让 L3 成为 L4 方案的第一个版本。
四、供应商的位置：螺丝钉、Tier 2 和共创
在第三轮提问中，赵福全问得很直接。
他说很多自动驾驶初创公司的老板跟他讲，五年前成立的公司好多已经不在了，因为车企吸收了他们的能力，最后自己干了。下半场的商业模式、数据闭环、技术边界和责任到底怎么划分？
靳玉志的回答是，引望的角色从诞生起就没有变，是智能化部件的供应商，「电子螺丝钉」，帮车企造好车、卖好车，不和车企竞争。引望今年的研发投入将近 200 亿元，未来还会保持这样的强度。
他提到的新动作是生态协同。
10 月 19 日，引望与宁德时代、时代电服签署换电合作，目标是让车辆自主寻站、进站、换电、出站，全程 99 秒完成。按照靳玉志的说法，乾崑智驾已经接入了全国百万级的充电桩，再加上换电网络，由引望统一接入，车企不需要各自再接一遍。
余凯的回答从「我们不造车」开始。他把地平线定位为专注计算的公司，只做计算芯片和基础软件，域控制器这类硬件不做，运营「也绝对不是我们的活儿」。地平线 95% 的车型项目通过合作伙伴交付，是典型的 Tier 2。
他用英伟达来说明芯片公司的价值在哪里：英伟达的 GPU 硬件架构和 AMD 差不多，强在上层的 CUDA 软件生态。他的判断是，创新不是规划出来的，而是协同中产生的。在一条这么长的产业链里，一家公司在每个环节都做到最强的概率不高。
余凯举了一个客户案例。地平线和博世合作拿下了铃木的项目，面向的是铃木的印度市场，而地平线在印度一个工程师都没有，项目由一家中国科技公司和一家德国零部件供应商，为一家日本车企完成。他说中国科技企业有研发和人才优势，但应该尊重现有的合作网络，做「价值的创造者」。
赵福全问曹旭东，「你还能活多久啊」，原话。
曹旭东的回答是，上半场是供应关系，下半场会更多是共创和共赢。共创体现在前面那个 L3 架构的案例里，Momenta 和客户一起决定产品架构。共赢和订阅模式有关。他预判 2028 年以后订阅会逐渐成为主流，Momenta 可以和客户在产品、营销、运营上一起提高订阅率，「订阅率越高，大家的饼就越大」，双方可以分享的收益也越多。
在赵福全看来，地平线和 Momenta，两家在和中国车企合作的同时，也在和外资车企、零部件企业合作，中国企业出海去做本地化，离不开高阶自动驾驶的落地。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, HiEV Garlic Cars Research Institute, author, |, editor, |, |
This week, at the World Congress of Smart Nets in Beijing, China's visit to CEO Yong Yuzhi, Horizon, CEO Yucai and Momenta CEO, three of the country's top intelligent drivers, together with CEO Lee Bin, Peng Moo, and Senior Vice President of Biadi and Director of the New Technical Institute, Yang Dongsheng, was chaired by Zhao Fook, Director of the Institute of Automotive Industry and Technology, University of Chinghua.
In this conversation, three different judgments were given as to where autopilot went.
The judgement given by Wah, Senior Vice-President, and looking forward to the CEO, is to accelerate from auxiliary driving to automatic driving.
“The revolution is just beginning” according to the founder and CEO of the horizon. In his view, there will be a few more trips to the Pacific Ocean. The top-level industry city NOA has just managed to take over 100 kilometres a key one, while Robotaxi needs 100,000 kilometres.
The judgment of Momenta CEO Cao Xudong is optimistic. He believed that the increase in intellectual driving capacity had accelerated ten times more than every two years in the past, to ten times more per year.
The three people agree on one thing: the second half of the game is to speak with quantization security data. As a result of the cumulative miles and relative human driving safety data produced by Tsubaki, the remaining MPCI was used, while the increase in subscription rates was attributed to increased security.
Differences have focused on two issues: how fast technological progress is going and where suppliers should be in the second half.
One, the second half of the game, ten times a year, or more across the Pacific.
Zhao has set a clear premise at the beginning of the campaign: in 2026, it was an automatic drive, with a technological breakthrough in the first half, and in the second half, how to get into thousands of homes and how to make money.
He asked each guest to share his understanding of the second half.
The second half of the game has been judged not by the technology, but by the industry rhythm. He listed three changes.
The first is that the auxiliary driving is accelerating towards automatic driving, that the penetration rate of L2 and above has reached 70%, that the Ministry of Industry and Communications has organized access to L3 in an orderly manner, and that the compulsory national sign for automatic driving has been implemented by publish on 1 July next year.
The second is the heightened security requirements. The subject of the L2 phase had been the driver in an accident; at the L3 level and above, responsibility would be transferred to the system.
And the third is that the way of travel is re-engineered. In the case of the autopilot, when you go to work, the car stops, recharges, washes, picks up kids and even buys you coffee.
Yu Kai's answer is most cautious. He said that he was engaged in technological research and development and that he felt "so many problems to solve, this is so difficult" that the first reaction was "The Revolution is just beginning".
He used MPCI, i.e., the average number of kilometres, as the basis for the demarcation phase.
According to his judgement, the top-level city of the Chinese industry, NOA, has just taken over 100 kilometres a time. Users won't be able to take over in less than half an hour at the 4S store, and 100,000 dollar-sized cars are starting to match the city NOA, so he agreed that 2026 was a turning point.
But from 100 km to 1,000 km, to 10,000 km, Yu Kai thinks there is no ready answer. His question was whether the next crossing could be achieved through multiple data and training models along the current route. The point of reference he gave was that Tesla took over the city about 1,000 kilometres once, and actually did Robotaxi, L4 and L5, which took over 100,000 kilometres.
Yucai compares every ten-fold increase across the Pacific Ocean and gives an evolutionary trajectory: Now it's available, from 1000 to 10,000 kilometers is good, and 100,000 kilometres is good. He hopes that the industry will cross over in the next four to five years.
Cao Xudong agrees that 2026 is a watershed for the first half and the second half, but that the rate of technological evolution is different from that of Yu Kai.
Momenta has an internal "Advanced Moore's Law" which is said to be ten times higher every two years.
He said that from 2022 to 2026, it was largely consistent with this pattern, with an increase of about 100 times in four years, and that intellectual driving had changed from a little useless to a little more operational. Now, he believes that the second half of the game will be updated to a tenfold increase per year.
According to a test assessment of Tesla FSD within Momenta: Safety has increased by more than a dozen to dozens in over a year from V13 to V14. At this rate, he judged that 2027, especially 2028, would be an important turning point for L3 and L4.
He also referred to the prediction in the Honorary Chairman of the Chinese Association for Science and Technology, Mangjik, in his keynote report on the same day that the L3 penetration rate had reached 15 per cent in 2028 and 35 per cent in 2030.
Tsao Xudong even proposed a "industry standard" that would sound so much like it: by 2028, the product experience would need to be safe at the L4 level, even if it was an auxiliary driving L2+.
On the side of the car, Peng's views are closer to that of Yu Kai, and it's hard for me to be able to take over ten kilometres or 100,000 kilometres at a time.
He cited the need for cars to fully integrate chassis control and automatic driving when travelling in snow, marshes and grasses, which he considered impossible for a year or two.
II. Two feet in the second half: security data and subscriptions
When the responsibility is transferred to the system, the proof of safety will have to be presented.
On the basis of this consideration, it is hoped that real-time release of safety data on auxiliary driving miles and relative human driving will begin early this year. He mentioned that, by the day of the dialogue, the cumulative miles of dry and intelligent driving would have exceeded 17 billion kilometres.
In terms of user usage, this year, on the 7 days of the National Day, dry-driving users travel 80 million kilometres a day using auxiliary driving, with a cumulative 7 days of 560 million kilometres, representing 60 per cent of the mileage of the auxiliary driving.
The other ruler is "Will users pay for it?"
This point is most concrete.
He predicted that L3 subscription patterns would become more common, on the grounds that the willingness to pay would grow when the value of the user was actually created.
He cited Tesla as an example: when FSD V13, North America had a subscription rate of about 20%. By Tesla's latest edition, this number has reached 55%. He attributed growth to increased security. On the domestic front, he referred to the earlier exploration of the pattern of subscriptions by China.
According to Li Bin, self-driving depends on two indicators: how many users are willing to pay, and whether it is possible to use quantization security data, including feedback from insurance companies, to make automatic driving feel safer.
For his part, He Xiao Peng proposed three new indicators for the second half of the game: a safe collision mileage, a full scene coverage and a user paid rate.
Li Bin also gave the data. According to him, there are 950,000 motor vehicles carrying more than 1,000 TOPS calculators, with a cumulative driving distance of more than 91.40 billion kilometres, with a person driving eight times as safe as a person driving, and certificates from four insurance companies showing a drop of more than 30 per cent in major accidents.
In his closing summary, Zhao Fook also pointed out that consumers had been "unusual" to upgrade free of charge, but that the subscriptions were backed by business R & D costs and profit margins.
III. L3 is most divided: transition, or skip?
In the second round of questioning the schedule of the three car companies L3, Zhao Fu all had very different accounts.
He thinks I should skip L3, put L2 and L4 together.
He gave the example of the L3 that Audi opened on a few highways in Europe, and he tried for a day and only a few minutes. His judgement was that L3 shifted responsibility to the system to some extent, “but the experience was not optimized and security was not improved”. He gave the time that L4 technology landed around the second half of 2028.
The view of Yang was the opposite. He thinks that L3 is a big threshold, that security requirements are "not multiplying, probably a few hundred times" and that the L3 that limits the scene has hope to land in 2028.
Li Bin looks at technology and compliance separately.
For compliance, GB 44721-2026 Safety Requirements for Smart Networked Autodriving Systems will be implemented on 1 July 2027; technically, he considers world model closed-ring intensive learning to be an industry consensus, with L2 to L5 "technology being co-sourced". As for the timing of L3 production, his answer was "depending on when the state allowed it."
Cao Xudong gave a third solution. He said he felt exactly the same way about L3 in Europe.
A European client in Momenta made L3 in China, initially following European thinking.
As the market and policy of China L3 became clear, the parties moved to the Chinese L4 approach to L3: Sensors, chip domain controls and implementers were designed as L4 structures, L3 regulations were in place first L3 and L4 regulations matured and then upgraded to L4 through software.
The judgement of Cao Xudong is that the transition cycle of L3 will be shorter and will soon escalate to L4. According to him, this would require only the development of an L4 framework programme to cover both L3 and L4 and significantly lower development costs. The listed L3 will also have a completely different experience with the European L3: "More precisely, the L4 experience that Peng has just described."
Momenta's thinking and Ho's "jumping L3" are close to the target, putting L4's capabilities in a mass-producing car. The difference is that it does not circumvent the L3 regulation node, but makes L3 the first version of the L4 scheme.
IV. Location of suppliers: screws, Tier 2 and co-generation
In the third round of questions, Zhao Fu asked all questions directly.
He said that many of the owners of the self-pilot start-up companies told him that many of the companies that had been established five years earlier were no longer in existence because the car companies absorbed their capabilities and ended up working on their own. How does the second half of business model, data closure, technical boundaries and responsibilities fit?
The answer is that the desired role has not changed since its birth. It is the supplier of intelligent components, the "electronic screw" that helps build cars, sell them, and not compete with them. This year's investment in research and development is expected to be close to $20 billion, and it will be sustained in the future.
The new action he mentioned was ecological synergy.
On 19 October, an exchange of electricity with the Ningdé era and with the age-old uniforms was requested, with the aim of enabling vehicles to find their own stations, to enter, to exchange electricity, to exit the stations and complete 99 seconds. According to Xi Jin Yuzhi, the dry-taiji driver has been connected to a million-grade charger throughout the country and, together with the power exchange network, has been connected to the single-in-visibility network, which does not need to be repeated separately.
Yu Kai's answer starts with "We don't build cars." He's positioned the horizon as a focused computing company, only computing chips and basic software, not hardware like domain controllers, and operating "it's definitely not our job." Horizon 95 per cent of car-type projects are delivered through partners, typical of Tier 2.
He's using Weedda to show where the chip company's value is: the GPU hardware architecture in Weedda is similar to the AMD, the CUDA software ecology in the upper layers. His judgement is that innovation is not planned, but is the result of synergy. In a chain of such a long industry, a company is less likely to achieve the highest probability of each link.
Yu Qai presented a client case. The project, which was carried out in cooperation between Horizon and Bo, aimed at Suzuki ' s Indian market and none of its engineers in India, was completed by a Chinese technology company and a German spare parts supplier for a Japanese car company. He said that Chinese science and technology enterprises had a research and development and talent advantage, but that they should respect existing cooperation networks and become “value creators”.
Zhao Fu asked Cao Xudong, "How long will you live?"
Cao Xudong replied that the first half was a supply relationship and the second half would be more co-creation and win-win. In the case of the previous L3 architecture, Momenta and the client decided the product architecture. The win-win and subscribe mode is relevant. He predicted that subscriptions would gradually become mainstream after 2028, that Momenta could increase subscriptions with clients in terms of products, marketing and operations, that "the higher the subscription rate, the greater the pies" and that the more benefits could be shared by both parties.
From Zhao's perspective, horizons and Momenta, while working with Chinese companies, are also working with foreign-owned companies, parts and components, and Chinese enterprises are going to sea to localize, without having to drive on top of their own.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.