AI data centres have become a new constraint on the automotive chip supply chain, and Chinese carmakers are responding by designing more silicon in-house. That was the consensus at the sixth Intelligent Vehicle Chip Ecosystem Conference in Shanghai on 16 September, hosted by Gasgoo and reported by D1EV. With AI data centres consuming huge volumes of advanced process capacity, high-bandwidth memory (HBM) and advanced packaging, high-end automotive chip capacity is under pressure. Meanwhile, computing platforms are passing 1,000 TOPS, and the industry's focus has shifted from buying chips to building them, from raw TOPS to usable compute and ecosystem fit.
Chip value is rising faster than chip count
Gasgoo partner and executive vice-president Gu Xiaoying set out the market context. Auto-grade chips are moving from volume growth to value growth, she said: an L3 battery-electric vehicle carries roughly 200 to 300 more chips than an L1 petrol car, and the value of those chips rises by 124%. The global market was worth $67.7 billion in 2024 and is expected to reach $96.9 billion by 2029, with China the core demand market. In the MCU segment the top five suppliers hold 92%, with Infineon at 36%; Nvidia holds around 40% of intelligent-driving domain controllers; Qualcomm about 70% of cockpit SoCs. Local producers are gaining ground in storage and communications chips, Gu added.
Changan's senior deputy chief engineer for chip development, Ding Ke, said Chinese brands have already overtaken on the new technology track: their combined market share reached 75% in the first half of 2026, L2 assisted-driving penetration exceeded 70%, and high-level L2+ fitment reached 29%. He called the intelligent car the next biggest ICT ecosystem after PCs and smartphones, and said the vehicle is becoming an evolvable intelligent agent. Future architectures, he said, will centre on an AI compute brain above 1,000 TOPS with high-bandwidth central computing.
On software models, Ding said 7-billion-parameter models are already deployed in cars, with 30-billion-parameter models under consideration using mixture-of-experts. End-to-end technology is moving from transformer-plus-bird's-eye-view architectures to VLM (vision-language models), VLA (vision-language-action) and world models. He also outlined Changan's RISC-V work, using custom instruction sets and BF16+INT8 mixed-precision acceleration to improve accuracy and reduce cost. The company's stated goal is to push down CPU and GPU prices, open up cooperation and extend chip control down to the IP-core level.
Axera automotive business senior product director Yang Quanshen described his company as a pure chip supplier that leaves integration to Tier 1 suppliers and returns technology sovereignty to carmakers. Axera has shipped more than 1.6 million intelligent-driving chips and has them in over 40 mass-production models. Its products span an entry-level M57 chip for integrated parking and driving, a 5nm M9 series aimed at one-stage end-to-end, urban NOA (city advanced driver assistance) and VLA architectures, and an X6 edge-side accelerator for cockpit large models.
Trench SiC, ignition chips and a single-chip cabin
Bosch power semiconductor product manager Wang Junyue, who also heads Bosch's global silicon carbide (SiC) data-centre business, said the company has stuck with trench SiC. Since starting SiC production for traction inverters in 2021, Bosch has shipped more than 70 million automotive SiC chips, with the cumulative figure now around 80 million, and has recorded no field failures. It operates two 8-inch SiC wafer fabs in Germany and the US, with the US plant due to start mass production at the end of this year, and plans to expand global SiC capacity 20-fold between 2024 and 2030. Third-generation products are scheduled for 2027. Wang said Bosch's AI data-centre business, launched on 1 July, will absorb significant SiC capacity, with solid-state transformers an almost all-SiC application.
Shanghai Belling's automotive electronics market manager Zhang Fei covered engine ignition, noting that the application extends beyond petrol cars to hybrids, EREVs (range-extenders), motorcycles and generators. Belling's SAQ3100 dedicated ignition IGBT gate driver is AEC-Q100 Grade-0 qualified, withstands 150°C under-bonnet temperatures and integrates four protection functions. The company says its ignition solution has gone through eight years of development and ten product iterations, with cumulative shipments of 40 million units, an ISO 26262 ASIL-D functional safety certification and more than 200 million automotive-grade devices shipped in both 2024 and 2025.
Black Sesame Technologies product head Erite said the mainstream market, not flagship models, is where chip costs get out of control: dozens of cockpit and ADAS chip combinations force carmakers to repeat the same work for every new chip. Black Sesame's Wudang C1296 cross-domain fusion chip collapses that choice into one platform and has gone into mass production on Dongfeng's Tianyuan intelligent-driving platform. The company runs two product lines, Huashan for ADAS and Wudang for cross-domain fusion, and says it is one of the few suppliers offering a matrix of chip combinations.
