Gasgoo Munich-Halfway through 2026, a stark contrast has emerged. Launch events hype "AI large models" and "computing power exceeding 1,000 TOPS." However, the system that truly determines winter EV usability, prevents chip throttling, and manages charging wait times is a long-overlooked "supporting actor"—thermal management.
This is no exaggeration. Nvidia's Thor automotive chip has a heat flux density of 250W/cm², with local hotspots exceeding 500W/cm². Starting July 1, 2029, passenger vehicles using refrigerants with a GWP over 150—specifically R134a—will be banned. The industry has a window of less than three years. Winter heating dominates an EV's power usage. Traditional rule-based control cannot handle complex, coupled heat sources.
The slogan "AI-defined cars" has echoed for over a year. Thermal management is the first "non-intelligent driving" system redefined by this shift. It is evolving from a passive "nanny" dissipating heat into an active "butler" allocating vehicle energy.
AI-Driven Predictive Control: From "Rule Books" to "Neural Networks"
Traditional thermal control relies on engineers presetting thousands of "if-then" rules. For example, turn on the PTC heater if the temperature is below X. But with five or six heat sources to coordinate, the rulebook collapses. These include the battery, electric drive, cabin, autonomous driving chip, and onboard fridge. "The smarter the system, the more sensitive and prone to error it becomes," said Zhou Damei, a thermal technology manager at Pan Asia Automotive Technical Center. "Frequent component switching challenges reliability and stability."
The solution is to rewrite control logic using AI. Zeng Like, Valeo's chief product and technology officer for platforms, outlined a practical path.
• Multi-zone AI Airflow Distribution: The air conditioning box involves complex variables like circulation, flaps, and defrosting. Traditional forward algorithms cannot maintain precision. Valeo uses neural networks trained on bench data. This is transferred to updated models, reducing airflow fluctuation to under 1%, which is imperceptible to users.
• Sensor Virtualization: Neural network models predict outlet and cabin temperatures. This replaces some physical sensors, cutting costs.
• Automated Calibration: Engineers feed calibration logic to the AI. This allows the controller to tune parameter maps itself, compressing months of work into weeks.
• Failure Prediction: Models train on system characteristics at different leak levels (20g/50g/80g). The system outputs a precise leak value between 0 and 200 grams. This is far more granular than traditional threshold alarms, which trigger only when half the fluid is gone.
Liu Jun's team at Dongfeng Motor pushes further with vehicle-cloud collaboration and MPC. Based on trip predictions and user habits, the system stores heat or pre-cools in advance. For short trips, it avoids blindly heating the battery. For long journeys, one-tap preparation optimizes powertrain and battery water temperatures. It also preheats water pipes before charging to reduce fast-charging time. "High and low-temperature range for hybrids and EVs can be optimized here," Liu said.
Industry benchmarks offer a clue. AI-based predictive thermal architecture can boost winter range by 20%. One leading startup achieved an 8% increase via a February OTA update without hardware changes. By 2028, AI thermal management adoption is expected to top 90%. The question is not "whether to adopt," but "if you're late, you'll be outpaced."
Valeo's Zeng Like offered a sobering assessment. AI is inherently probabilistic and prone to hallucinations. "We cannot let AI-generated strategies run directly," he said. "They must be filtered through human rules." Explainability, ISO 26262 functional safety, and SOTIF are three barriers AI thermal management cannot bypass.
System High Integration: From "Scattered Piping" to "Dual Indirect Architecture"
The industry has championed integration for five years. The new consensus for 2026 is clear. Integration is not about stuffing parts together; it is architectural reconstruction.
After Tesla introduced its Octovalve in 2020, Chinese competitors followed quickly. Wang Tianying, thermal management director at IM Motors, clarified the roadmap. The industry is moving beyond "integrated heat pumps" to "dual indirect systems." The refrigerant is sealed in a compact unit in the engine bay. Only three water pipes enter the cabin, coupling the battery, electric drive, and cabin on the water side.
IM Motors validated this logic with its ITMS 3.0 EV platform and ITM 4.0 range-extender platform.
• The flow channel plate weighs just 0.82kg, the lightest in the industry.
• Refrigerant charge is compressed to the 120-gram level. This significantly reduces leakage risks in collisions.
• OEMs no longer handle refrigerant filling. Suppliers deliver sealed units for installation. "Overall manufacturing costs actually go down," confirmed Wang Tianying.
• In terms of efficiency, CATARC testing rated the air conditioning at Grade 1 for both +40°C and -20°C conditions.
Wang Yimin, R&D director at Helladd Asia Pacific, offered a sharp critique. "Refrigerant-side five-way valves have hit a ceiling," he said. "The fight is moving to the water side." Helladd's two-way and three-way valves have been in mass production for over two years. Its five-way valve B-sample is compatible with R290. It boasts a ceramic valve core internal leakage of 2mL/min and a lifespan of 15 years or 300,000 km. The refrigerant-side five-way is the peak. The water-side nine-way is the next battle.
Integration has side effects. Li Jingyuan, a senior thermal integration engineer at Changan Automobile, admitted to stumbling blocks. "Indirect heat pumps consume about 100W more than direct systems," he said. "Component sharing leads to 'cross-leakage heat loss.'" Higher integration means greater system-level heat loss and maintenance costs. These are challenges OEMs and Tier 1 suppliers must solve together.
Tao Hong of Haili New Energy offered a sharp take. "New energy vehicle thermal management systems will converge toward simplicity and reliability," he said. "The massive customization compressor makers now do does not help cost reduction. Standardization is something the industry must do together."
Cooling High-Compute Chips: The "Invisible Bottleneck" at 250W/cm²
This section is likely to be overshadowed by the "AI-defined car" narrative in 2026. Yet it remains the biggest headache for engineers.
First, the data. L4/L5 autonomous driving computing power has broken the 1,000 TOPS barrier. Nvidia Thor hits 2,000 TOPS. Using 4nm or 5nm process nodes, these chips face heat flux densities of 250W/cm². Local hotspots spike over 500W/cm². 3D packaging complicates thermal coupling. With the harsh vehicle environment, traditional air cooling is out. Cold-plate liquid cooling is now the mainstream path for intelligent driving domain controllers.
Simply transplanting data center solutions does not work. Onboard applications face three unique pain points.
1. Condensation: When coolant temperature falls below the dew point, condensation forms on channel walls. It drips onto circuits, causing short circuits. Huawei MDC uses shell perforations with e-PTFE breathable membranes. A more advanced approach is the "breather valve + desiccant" CMD structure.
2. Sealing and Leaks: O-ring aging combined with pipeline vibration is a risk. One automaker issued a recall due to O-ring failure. Fluororubber O-rings still rely on imports from Japan or Germany. They account for 15%–20% of system cost.
3. Low-Temperature Fluidity: At -40°C, the viscosity of water-based coolant spikes. This can prevent the pump from starting. Switching to low-viscosity silicone oil adds further cost.
Materials are catching up too. Qi Dan, chief of material development at FAW's R&D center, highlighted a bottleneck. "Imported silicon nitride powder costs 380–400 yuan per kg," he said. "If domestic production hits 120–150 yuan, there is room for cost cuts." For high thermal conductivity silicon nitride powder, China lags behind Japan and South Korea. Diamond nano films "are not as expensive as imagined, but technical issues remain."
The spillover of thermal management is happening. Gu Jifeng, deputy technical director at Air International, discussed humanoid robot solutions. Servo joints account for 60%–80% of total heat dissipation. Computing chips like Nvidia AGX face heat flux densities of 150–300W/cm². They throttle above 85°C. "Robots are like humans, requiring multi-point heat dissipation," Gu said. "We want to create bionic 'vascular' water jackets." This shares the same technological DNA as intelligent driving domain liquid cooling.
An easily overlooked signal is emerging. Thermal management is spreading from "cars" to "robots and low-altitude flight." The same liquid cooling and phase-change materials can be sold four times. This is the second growth curve for component manufacturers.
Policy-Driven Green Transition: The Hard Deadline of July 1, 2029
The final chapter must land on policy. This is the strongest force forcing change in thermal management over the next three years.
The "National Plan for China's Implementation of the Montreal Protocol (2025–2030)" is strict. By January 1, 2029, controlled uses of HFCs must be cut by 10%. Starting July 1, 2029, the automotive industry is banned from using refrigerants with a GWP over 150 in new M1 class vehicles. R134a is directly out. This leaves a window of less than three years. Prices for third-generation refrigerants have hit a ten-year high. The question is not "whether to switch," but "if you don't decide now, you won't make mass production by 2027."
Currently, three paths run in parallel. Each has its own calculus.
Tao Hong, chief technology expert at Haili New Energy, offered a pragmatic view. "COP and refrigerant charge are eternal challenges," he said. "R1234yf can at least be used until the 2029 deadline." He confirmed that Honeywell is negotiating domestic capacity with the government. This implies supply and costs for 1234yf will fluctuate after 2029.
Liu Jun's team is working on an R290 double-secondary-loop indirect heat pump. It features a "cross-generation compatible design." Changing interfaces allows switching between R1234a, R290, and R744 refrigerants. This is a strategy of not putting all eggs in one basket.
The math for commercial vehicles is compelling. Huang Jiayue, deputy head of thermal management development at Foton, calculated the figures. A 4.5-ton pure electric logistics truck saves about 24,000 yuan annually if efficiency improves by 10%. This equals a 15% boost in net profit. Net margins in freight are under 5%. This math determines a company's survival. Under policy and TCO pressure, the low-GWP switch for commercial vehicles will be decisive.
Conclusion:
Returning to the opening question: How will thermal management develop in the era of AI-defined cars?
The answer is no longer the linear iteration of "bigger radiators, more valves." It is a simultaneous leap across four dimensions.
• Control: From rules to AI prediction, MPC, and vehicle-cloud collaboration. A 20% boost in winter range is the first quantifiable dividend.
• Architecture: From scattered systems to dual indirect, multi-way valves, and standardization. Beware of the heat loss and maintenance costs brought by integration.
• Scope: From protecting the three-electric system to covering AD chips, robots, and low-altitude flight. A heat flux density of 250W/cm² is the new engineering ceiling.
• Compliance: July 1, 2029, is a hard countdown. The R1234yf transition and the R290/R744 dual track are the most certain investment directions for the next three years.
Thermal management is evolving from an auxiliary subsystem to a vehicle energy dispatch center. It is moving from "passively preventing hardware overheating" to "actively defining range, experience, and safety." Two years ago, this sounded like a slogan. After the 2026 forums, it is a project on everyone's PPTs.
IM Motors' Wang Tianying said, "Dual indirect is the 'last mile,' and it's coming soon." Valeo's Zeng Like noted, "If AI generates counter-intuitive strategies that prove valuable, engineers must learn from AI." These capture the mindset of the thermal management community. The architecture's "last mile" is being stepped on. The control paradigm is shifting. The policy countdown is ticking.
There is no slowing down now.
