In January 2025, Google Cloud and Mercedes-Benz stood on the same stage to unveil Automotive AI Agent: built on Gemini, running on Vertex AI and launching in the new CLA's MBUX assistant. [Google Cloud] In 2026, BYD began adding Google built-in to new European models, while XPENG announced that overseas vehicles would use Google Maps Auto SDK as a navigation foundation. [BYD] [XPENG] The in-car AI assistant, a category that is already two decades old, has entered its third generation.
The debate has heated up with it, but most of the industry is asking the same question: whose model is strongest? Gemini versus GPT versus Chinese foundation models, compared on benchmarks, latency and hallucination rates.
That is the wrong question. Twenty years of in-car voice history point to the same conclusion: this category has never been won by the technology ranking. It is won by the need the product is built around.
Three layers of demand, three very different destinies
Lay out an automaker's needs from external suppliers and the differences in intensity are enormous.
Maps are a monopoly-grade necessity. No automaker can build a global map from scratch: collection, freshness and point-of-interest depth are a multibillion-dollar annual sinkhole. Outside China, consumers' expectations for in-car navigation have been trained by their phones. If the system cannot approach Google Maps quality, users switch to CarPlay and the cockpit experience the automaker spent years building is effectively bypassed. Automakers may resent that dependency, but they cannot route around it.
Voice is a fully competitive product. LLMs have demolished the old entry barrier: today, almost any automaker can connect an API to a microphone array and assemble an assistant that can hold a conversation. Traditional automotive suppliers are transforming, independent companies are chasing contracts, technology giants are bundling products, and automakers are considering in-house stacks. Four camps are fighting at once, and none has a monopoly lever.
Cloud is almost completely substitutable infrastructure. Automakers are naturally multi-cloud, use whichever provider offers the right economics, and show close to zero loyalty.
Those three layers produce three supplier destinies: map vendors retain negotiating power, voice suppliers are continually repriced and replaced, and cloud providers fight on price.
The fate of voice suppliers was written by the category
The dominant supplier of first-generation in-car voice once appeared in nearly every vehicle that offered voice control. Yet anyone who lived through that era remembers that even at the height of its dominance, the business was difficult. Automakers treated voice as an experience feature rather than mandatory equipment. Budgets were cut year after year, customization costs could not be passed through, and the threat of replacement never went away. Local players later won the Chinese market with better Mandarin experiences; LLMs then gave everyone access to a model. The original leader's business has since been split and passed through several corporate structures, and its position in the capital markets is plain to see.
The problem was never that its technology was poor. The problem is that voice itself is not a hard requirement. Remove the assistant and the vehicle can still be sold. Remove navigation and the sale becomes much harder. A supplier living under a cuttable budget line can have superb technology and still face an upper limit defined by gentle, relentless price pressure.
Replacing the old stack with an LLM does not rewrite that fate. Any in-car AI company that still defines itself as "smarter voice" is walking the same road.
Google's real card is not Gemini
Return to the Mercedes-Benz announcement and one detail stands out. The concrete selling point is not simply that Gemini is intelligent. It is that the agent can call on Google Maps Platform data covering 250 million places, with more than 100 million map updates every day. The demonstrations are all navigation questions: "Is there an Italian restaurant nearby?" "Does it have good reviews?" "What is the signature dish?" [Mercedes-Benz]
In plain language: Google is selling the agent, but maps are the ballast.
That is the genuinely asymmetric part of this competition. Models are broadly comparable and the ranking will reshuffle again next year. But only one company can currently offer a conversational assistant natively grounded in the world's strongest global consumer map dataset. Traditional automotive suppliers do not own the map. Independent voice companies do not own it. Automakers attempting to build in-house own neither. Position the same product as an "AI assistant" and it enters a red ocean; position it as "navigation you can talk to" and the field is nearly empty.
Go one layer deeper and this is probably the product's end state. On a vehicle screen, "map" and "assistant" should never have been separate things. Drivers do not really want a car that can chat. They want navigation that can negotiate: "Find a charger on the way, make sure it has coffee, and don't add too much time." The agent is not a new feature sitting beside the map. It is the map's next interaction layer. Whoever truly grows the two together first turns an experience feature into mandatory equipment — and moves itself from a cuttable budget line into a core requirement.
A hard-demand anchor does not guarantee victory
Three buckets of cold water are necessary here, or this becomes advertising.
First: the trust record. Google has repeatedly sunset driver-facing experiences, including Google Assistant Driving Mode, which disappeared after years of changes. [9to5Google] Every cockpit procurement team remembers that consumer-internet companies can cancel products. Automotive programs need supply commitments measured in a decade or more; Google's product-sunsetting habit is one of its largest psychological barriers to entering the vehicle.
Second: the early evidence was not flawless. Rivian's first weeks after adding Google Maps features included owner reports of frozen navigation and repeated reboots. [Carscoops] Heavy EV users also remained divided over the maturity of charging planning. A better data foundation does not remove rollout risk.
Third: integration depth is the dividing line. Google itself describes EV routing as being tailored to a vehicle's specific needs, including proactive charging stops. [Google built-in] That promise only works when the automaker exposes state of charge, consumption, charging and vehicle-control data and then integrates the outputs into the cockpit. Renault and Polestar have shown what deeper integration can do; a brand that merely installs the Google layer can still leave drivers reaching for CarPlay. The hard-demand anchor is an entry ticket, not immunity.
Ask everyone the same question
This framework is not only for understanding Google. It is a mirror for the entire industry.
For Chinese cockpit and voice suppliers, the question is: what hard requirement does your assistant grow from? In China, the answer can be local digital ecosystems and assisted driving. Once the product goes overseas, however, global languages, global POIs and local services are all missing. That gap is exactly why leading Chinese exporters are increasingly turning to Google. For traditional mapmakers, the question is: when an agent takes over the map's interaction layer, which hard requirements remain yours? Regulatory data and the ADAS chain — areas the agent cannot cover — are the positions worth defending. For automakers, the question is simpler: when you cut the budget, who goes first? That answer reveals every supplier's real industry status.
The table for third-generation in-car assistants has just been set. Do not watch only whose model posts the best benchmark. Watch who has embedded itself in a need that cannot be cut.
The next decision point comes in the fourth quarter of this year, when Mercedes-Benz, BYD and XPENG implementations will face real owners at the same time. What will be judged is not AI's IQ, but the quality of the demand beneath it.
This article is based on public information. The argument and conclusions are those of TopChinaCar.
2025年1月,Google Cloud和梅赛德斯站上同一个讲台,发布Automotive AI Agent:Gemini打底,跑在Vertex AI上,首发装进新CLA的MBUX助手。[Google Cloud] 2026年,比亚迪开始在欧洲新车型中内置Google built-in,小鹏则宣布海外车型采用Google Maps Auto SDK作为导航底层。[比亚迪] [小鹏] 车载AI助手这个二十年的老品类,进入了第三代。
行业讨论随之热闹起来,但绝大多数讨论都在问同一个问题:谁的模型强?Gemini对GPT对国产大模型,跑分、延迟、幻觉率。
这是个错误的问题。车载语音二十年的历史反复证明:这个品类的胜负手从来不在技术排名,在你的产品长在什么需求上。
三层需求,三种命运
把车厂对外部供应商的需求摆开,强度差得远。
地图是垄断级刚需。没有任何车厂能自建全球地图——采集、鲜度、POI深度是每年十位数美金的无底洞。而在中国以外的市场,消费者对车机导航的期待被手机养刁了:达不到Google Maps的水准,用户直接CarPlay投屏,你的座舱等于白做。车厂可以恨它,但绕不开它。
语音是充分竞争品。LLM把语音助手的门槛砸穿了:今天任何车厂接个API、配上麦克风阵列,就能攒出一个能对话的助手。传统车规玩家在转型,独立公司在抢单,科技巨头在打包,车厂自建派在观望——四路人马混战,谁都没有垄断性筹码。
云是完全可替换的基础设施。哪家便宜用哪家,车厂天然多云,忠诚度约等于零。
三层需求,对应三种供应商命运:卖地图的有谈判地位,卖语音的一路被压价被替换,卖云的打价格战。
语音供应商的宿命,是品类写好的
第一代车载语音的霸主,巅峰期几乎装进了每一辆有语音功能的车。但熟悉那段历史的人都记得,即便在垄断期,它的日子也不好过:车厂把语音当“体验项”而不是“必配项”,预算年年被砍,定制化成本转嫁不出去,替换威胁常年悬在头顶。后来本地化玩家用更懂中文的方案切走中国市场,再后来LLM让所有人都有了模型——第一代霸主的业务几经拆分辗转,如今在资本市场的处境,业内有目共睹。
问题从来不是它的技术不行。问题是语音这个品类本身不是刚需:砍掉语音助手,车照样卖;砍掉导航,车卖不动。挂在“可砍预算”科目下的供应商,技术再好,命运的上限也就是被温柔地压价。
这个宿命不因为换了LLM就改写。今天的车载AI助手玩家,如果继续把自己定位成“更聪明的语音”,走的还是同一条被写好的路。
Google真正的牌,不是Gemini
回头看梅赛德斯那份发布通稿,有个细节值得玩味:真正的卖点不只是Gemini多聪明,而是这个agent“可调用Google Maps Platform的2.5亿地点数据,地图每天更新一亿次”。演示场景也全是导航:“附近有意大利餐厅吗?”“评价好不好?”“招牌菜是什么?” [梅赛德斯-奔驰]
翻译一下:Google卖的是agent,压舱的是地图。
这才是这场竞争里真正不对称的地方。比模型,大家半斤八两,而且明年就会洗牌;比“原生长在全球最好地图数据上的对话式助手”,地球上只有一家能做。传统车规玩家没有地图,独立语音公司没有地图,车厂自建派两样都没有。同一个产品,定位成“AI助手”进的是红海,定位成“会说话的导航”是无人区。
再往深一层想,这其实是产品的终局形态:车机屏幕上,“地图”和“助手”本来就不该是两个东西。用户要的从来不是一个能聊天的车,是一个能商量的导航——“找个顺路的充电站,要有咖啡,别绕太远”。Agent不是坐在地图旁边的新功能,agent是地图的下一代交互层。谁先把这两个东西真正长在一起,谁就把“体验项”变成了“必配项”——把自己从可砍预算,挪进了刚需科目。
刚需锚点不等于自动赢
写到这里必须泼三盆冷水,不然就成软文了。
第一盆:信任前科。Google曾多次终止面向驾驶场景的产品,其中包括经历数年调整后消失的Google Assistant Driving Mode。[9to5Google] 每个座舱采购都明白,消费互联网公司可能说砍就砍。车规产品要的是十几年的供给承诺,这种产品终止习惯,是Google进车最大的心理障碍之一。
第二盆:早期实证并不完美。Rivian加入Google Maps能力后的最初几周,出现了车主反馈导航冻结、需要反复重启的问题。[Carscoops] 重度电动车用户对充电规划成熟度的评价也依旧分化。更好的数据底座,并不会自动消除量产落地风险。
第三盆:集成深度才是分水岭。Google自己对Google built-in的定义也强调,EV路径规划需要针对车辆的具体需求,并主动规划充电停靠。[Google built-in] 这只有在车厂真正开放SOC、能耗、充电和车控数据,并把结果集成进座舱时才成立。雷诺和Polestar已经展示了深度集成能做到什么;如果一个品牌只是把Google那一层装上去,用户仍然可能转向CarPlay。刚需锚点给的是入场券,不是免死牌。
同一个问题,问所有人
这个框架不只是用来看Google的。它是一面镜子,照所有人:
对中国的座舱和语音玩家,问题是——你的助手长在什么刚需上?在国内,答案可以是本地生态和智驾;可一旦出海,多语言、全球POI、当地服务,一样都没有,这正是头部出海车企纷纷倒向Google的原因。对传统图商,问题是——当地图的交互层被agent接管,你的刚需还剩哪一块?法规数据、ADAS链路,这些agent覆盖不到的地方,才是能守住的阵地。对车厂,问题最简单——你砍预算的时候,先砍谁?那个答案,就是每个供应商真实的行业地位。
第三代车载助手的牌桌刚摆好。别盯着谁的模型跑分高,盯着谁把自己长在了砍不掉的需求上。
下一个判决点在今年四季度:梅赛德斯、比亚迪和小鹏的Google方案将同时接受真实车主的检验。届时见分晓的不是AI的智商,是刚需的成色。
本文基于公开信息写作,观点与结论由TopChinaCar独立提出。