Introduction
When you sit in a car listening to navigation, engaging in voice interaction, or using automated driving assistance, you may not realize that the in-vehicle chips are increasingly supporting AI model functionalities. Meanwhile, in the cloud, the competition for the “strongest AI model” is quietly determining which algorithms can be implemented in vehicles. The future of cars is not only a hardware battleground but also a testing ground for AI inference.
1. Cars are Evolving from Transportation Tools to “Smart Platforms”
In the past, cars were merely tools for transportation; however, in the era of “smart connectivity + autonomous driving,” the boundaries of vehicles are being reshaped. They are transforming from machines into “mobile intelligent terminals” and even “super computing platforms.” Computing power is shifting downwards: in the future, vehicles will need to take on more responsibilities for perception, preliminary decision-making, and real-time responses, rather than relying entirely on the cloud. Reliability is paramount: in-vehicle chips must withstand extreme temperatures, vibrations, and electromagnetic interference, as any failure could jeopardize safety. Ecological collaboration is essential: software, hardware, automotive manufacturers, cloud services, and data must integrate; otherwise, isolated breakthroughs cannot leverage overall efficiency. Thus, the essence of a car is transitioning from a “transportation tool” to a “smart platform.”
2. Automotive Chips: From Functional Devices to Intelligent Hubs
Automotive chips are undergoing a structural transformation: functional chips (MCUs, power devices, sensors, etc.) are responsible for motor control, battery management, and signal processing, serving as a fundamental yet indispensable role. Main control/computing chips are responsible for autonomous driving and intelligent cockpits, acting as the “brain” of the vehicle, determining its perception and decision-making capabilities. Automotive-grade challenges: they must meet stringent standards for high temperatures, low temperatures, vibrations, lifespan, and power consumption, with thresholds far exceeding those of consumer electronics chips.
Some Examples:
The Horizon J6 and Journey series have collaborated with domestic automotive manufacturers for mass production, focusing on NOA and advanced driving assistance. The installation volume of intelligent cockpit chips in China is expected to reach approximately 6.9 million units in 2024, with domestic chips surpassing 10% for the first time. NVIDIA’s DRIVE Thor platform targets computing power for L4 autonomous driving scenarios. In summary: the future of cars will be determined by the capabilities of chips in terms of intelligence.
3. The Competition of AI Models: From Cloud to Vehicle
In the cloud, the “ranking battle” of large models is in full swing. The current mainstream landscape (2025): First tier: GPT-4.5 (OpenAI), Claude 3.7 (Anthropic), Gemini series (Google/DeepMind) are rapidly catching up: DeepSeek R1 (Chinese open-source model), Qwen2.5-Max (Tongyi Qianwen), Doubao (ByteDance), Kimi (notable for long text capabilities) are emerging. Open-source and lightweight models like LLaMA, Mistral, and Grok have advantages in specific fields. However, the strongest models in the cloud may not be suitable for vehicle applications. In-vehicle chips are constrained by computing power, power consumption, and latency, necessitating “lightweight + high reliability” models. Implementation strategies include model compression (distillation, pruning, quantization) to transform large models into “smaller brains.” Edge-cloud collaboration will enable vehicles to handle perception and preliminary decision-making while the cloud manages advanced reasoning. Customized models will be specifically optimized for in-vehicle scenarios, including voice, vision, and interaction tasks. This means that the success of AI in vehicles will not depend on who has the strongest model, but rather on who can adapt large models most appropriately.
4. The Dual Competition: Who Can Outpace the Future?
I believe that the evolution of cars and AI in the next decade will present several directions: Integration of hardware and software: chip manufacturers, automotive companies, and model providers will form a closed loop rather than operate in isolation. Edge-cloud collaboration will become mainstream: vehicles and the cloud will work together, balancing real-time needs and complexity. Safety and explainability will be hard thresholds: especially in the transportation sector, the black box of algorithms is difficult for society to accept. Data and ecological barriers are crucial: those who possess more driving data will be able to train, optimize, and implement faster. Therefore, the future winners will not simply be the “strongest chips” or “strongest models,” but those who can integrate chips, models, and ecosystems into a cohesive unit, creating a virtuous cycle.
Conclusion
Cars are no longer cold machines but rather “deployment platforms for billion-level AI chips”; AI models are no longer confined to the cloud but must extend to vehicles, edges, and the real world. When these two technological curves intersect, perhaps autonomous driving will no longer be about “machines replacing humans,” but rather a new way of “humans collaborating with intelligent brains while in motion.”
A Thought for Readers:
On the future roads, the fastest vehicle may not necessarily be the one with the most horsepower, but rather the one that collaborates most harmoniously with chips and AI.
Postscript:
0.2 nanometers, a channel that can only accommodate nine silicon atoms side by side, has been drawn into the finest chip design blueprint to date by photolithography. It reflects the intertwined fates of semiconductors, operating systems, automobiles, AI, and even the European continent: 1. In the wafer fab, this 1.4-nanometer “draft” (to be mass-produced in 2027) has narrowed the gap between SMIC and TSMC to two generations of nodes, but the cost of missing EUV is reflected in yield; 2. In the packaging workshop, Changdian and Tongfu have combined two 7-nanometer Ascend 910C chips to achieve “near 5-nanometer” performance, using Chiplet technology to turn “non-usable” into “usable,” compressing a five-year gap into a three-year achievable range; 3. On mobile phone motherboards, the “millimeter-level” packaging of 3-nanometer SoCs is casually applied by pick-and-place machines, yet it delineates four racing tracks: Apple’s A18 Pro leads in single-core performance, Xiaomi’s Xuanjie O1 secures a domestic ticket, Huawei’s Kirin 9030 withstands with 7-nanometer equivalence, and Samsung showcases hardware prowess with foldable screens; 4. In vehicle cockpits, Qualcomm’s 8295 holds 77% market share, NVIDIA’s Orin-X commands 60% of intelligent driving computing power, and Huawei’s HarmonyOS cockpit achieves a <100 ms turnover speed between “mobile-phone-car-home,” cornering BBA; 5. In the cloud, DeepSeek-R1 has made it to the cover of “Nature” with $290,000 and 560 H800s, marking the first time Western scholars endorse a Chinese large model, yet it still lags behind Baidu’s Wenxin 4.0 in multimodal and commercial ecosystems; 6. In Europe, BMW has established a digital R&D center with 2,000 personnel in Shanghai, Audi’s Q6 e-tron China version is natively HarmonyOS, and Mercedes-Benz in Stuttgart is calling for self-developed MB.OS while evaluating the Harmony kernel—”the shell belongs to Munich, the core to Shenzhen.”
Thus, we discover:
The so-called “extreme chip” is not the loneliness of nine atoms at 0.2 nanometers, but rather an ecological onion wrapped layer by layer with packaging, HBM, EDA, app stores, OTA, and developer communities. When European car manufacturers yield “technological sovereignty” to “market survival,” when Xiaomi completes the “buy chip—make chip—whole vehicle” lightning loop in three years, and when HarmonyOS captures 19% of the Chinese market, equaling iOS with only 4% global share, the number of nanometers in chips is no longer just a physical scale but a flywheel turned by nations, enterprises, developers, and consumers together. Perhaps the true “extreme” is not in silicon atoms but in the speed of ecological rolling—whoever can first get the “design—manufacturing—packaging—software—scenarios” flywheel spinning will lock in the gap in this generation rather than being left behind in the next generation. Next time we discuss “which nanometer is the strongest,” the answer may not lie under a microscope but in the number of PRs on GitHub, in OTA update logs, in app store reviews, and in the success rate of a car owner’s wake-up command “Hey, Xiao Yi.” 0.2 nanometers, thus far; ∞ nanometers, just beginning.