Just past the halfway mark of 2026, a stark contrast in the industry can no longer be concealed: while various launch events tout "AI large models in vehicles," "autonomous driving computing power exceeding 1,000 TOPS," and "proactive cabin Agent services," what truly determines whether users dare to drive pure EVs north in winter, whether autonomous driving chips can operate without throttling, and the waiting time at supercharging piles, is a system long regarded as a "behind-the-scenes supporting role"—thermal management.
This is no alarmism. The thermal flux density of NVIDIA's Thor in-vehicle chip has surged to 250W/cm², with local hotspots exceeding 500W/cm². On July 1, 2029, new passenger cars equipped with air conditioning systems using refrigerants with a GWP > 150 (i.e., R134a) will be banned from sale, leaving the industry with a window of less than three years; winter heating energy consumption accounts for the "lion's share" of pure EV power consumption, and traditional rule-based control can no longer handle the complex operating conditions of multi-heat-source coupling.
"AI-defined vehicles" has been chanted for over a year, but thermal management is the first "non-autonomous driving" core system being redefined by AI—it is transforming from a "nanny" of passive heat dissipation into a "butler" that actively allocates the vehicle's total energy.
AI-Driven Predictive Intelligent Control: From "Rule Base" to "Neural Networks"
The logic of traditional thermal management control involves engineers presetting thousands of "if-then" rules—turn on the PTC when the temperature drops below X degrees, and turn on the compressor when it exceeds Y degrees. However, when there are five or six heat sources on the vehicle to coordinate, including the battery, electric drive, cabin, autonomous driving chip, and in-vehicle refrigerator, the rule base collapses. "The smarter it is, the more sensitive it becomes, and the more prone to errors. With a higher switching frequency of components, the challenges to reliability and stability become immense." Recently, at the 4th Automotive Thermal Management Full-Scenario Innovation and Development Forum, Zhou Damei, New Energy Thermodynamics Technical Manager at PATAC (Pan Asia Technical Automotive Center), pinpointed this pain point.
The solution is to rewrite the control logic using AI. Zeng Like, Chief Product and Technology Officer for Platforms at Valeo, provided a highly practical path on-site: • Multi-zone AI Airflow Distribution: The air ducts of the HVAC box involve multiple variable couplings such as internal/external circulation, temperature blend doors, and defrosting. Traditional forward algorithms cannot achieve precise and constant control. Valeo uses neural networks to train bench data and then transfers it to facelifted models, reducing airflow fluctuations to within 1%, making it imperceptible to users. • Sensor Virtualization: Using neural network models to predict outlet/cabin temperatures, replacing some physical sensors to reduce costs. • Automatic Calibration: Feeding engineers' calibration logic into the AI, allowing the controller to adjust parameter maps itself, reducing months of work to weeks. • Fault Prediction: Collecting system characteristics under different refrigerant leak volumes (20g/50g/80g) to train the model, which can output leak values from 0 to 200g—much more precise than the traditional threshold alarm that only detects a leak when "half is lost."
Liu Jun's team at the Dongfeng Motor R&D Institute has gone even further, focusing on Vehicle-Cloud Collaboration + MPC (Model Predictive Control): Based on trip prediction, road conditions, and user habits, they pre-store heat/pre-cool. For short trips, it avoids blindly heating the battery; for long trips, a one-click vehicle preparation pulls the electric drive/battery water temperature to the appropriate range in advance, and predicts/pre-heats the water pipes before charging to reduce fast-charging time. "The high and low-temperature range and fuel consumption on low battery for both hybrid and pure EVs can be optimized from this," said Liu.
Industry data provides a reference: based on an AI predictive thermal control architecture, the winter range of the whole vehicle can be improved by 20%. A leading EV startup improved the winter range of its older models by 8% in February this year through a single OTA update without changing any hardware. The AI thermal management adoption rate in new cars is expected to exceed 90% by 2028—this is no longer a question of "whether to adopt it," but a matter of "being benchmarked and outpaced if you are late."
However, Zeng Like from Valeo also offered a sober judgment: AI is essentially probabilistic generation, inherently prone to hallucinations. "At this stage, AI-generated outputs cannot be run directly; manual rules and reviews must be added." Explainability, ISO 26262 functional safety, and SOTIF (Safety of the Intended Functionality) are the three insurmountable mountains that AI thermal management cannot bypass.
High System Integration: From "Dispersed Piping" to "Dual-Indirect Architecture"
The industry has been chanting the slogan of integration for five years, but the new consensus in 2026 is—integration is not about cramming parts together; it is an architecture-level reconstruction.
After Tesla introduced the Octovalve in 2020, domestic companies followed quickly. Wang Tianying, Thermal Management Director at IM Motors, explained the route more clearly: the industry is taking another step forward from "integrated heat pumps" to "dual-indirect systems"—the refrigerant is sealed in a small integrated machine within the engine compartment, leaving only three water pipes entering the cabin, with the water side coupling the battery, electric drive, and cabin.
IM Motors' own ITMS 3.0 pure EV platform and ITM 4.0 extended-range platform have already validated this logic: • The cold side flow channel plate weighs 0.82kg, the lightest in the industry; • The refrigerant charge is reduced to the 120g level, significantly lowering the risk of collision leakage; • OEMs do not add refrigerant on their end; suppliers seal the integrated machine and deliver it for installation, "which actually reduces vehicle manufacturing costs" (confirmed by Wang Tianying in his original words). • In terms of energy efficiency, the CAERI (China Automotive Engineering Research Institute) evaluated the air conditioning to reach Level 1 at both +40℃ and -20℃.
Wang Yimin, Asia-Pacific R&D Director at Halla, added a sharp point from the component side: "The five-way valve on the refrigerant side has reached its limit; competition has shifted to the water side." Halla's two-way and three-way valves have been mass-produced for over two years, and the five-way valve B-sample is compatible with R290. The ceramic valve core has an internal leakage of 2mL/min (after 40,000 durability cycles at 1MPa), with a lifespan of 15 years/300,000 km. The five-way valve on the refrigerant side is the ceiling; the nine-way valve on the water side is the next battleground.
However, integration also has "side effects." Li Jingyuan, Senior Supervisor of Thermal System Integration at Changan Automobile, frankly admitted the pitfalls in his sharing: "Indirect heat pumps consume about 100W more electricity than direct ones. In our low-temperature tests, we found that component sharing + supplementary liquid pipe design leads to 'cross-water and heat leakage.' When the battery residual cold is in the low-temperature loop, there can be several liters of flow and tens of watts of heat leakage." —The higher the integration, the more system-level heat leakage, maintenance assembly replacement costs, and fault localization difficulty will jump. This is a problem that OEMs and Tier 1 suppliers need to solve together next.
Tao Hong from Highly New Energy gave a sharp judgment: "The thermal management system for NEVs (New Energy Vehicles) will gradually converge from the current complex and diverse exploration to simplicity, classic design, reliability, and low cost, stabilizing like traditional vehicles. The massive customization work currently done by compressor manufacturers (different mounting points/connectors/communication protocols) does not help reduce costs; standardization is something the industry must do together."
High-Computing Chip Cooling: The "Invisible Bottleneck" Under a Thermal Flux Density of 250W/cm²
This is the chapter most easily overshadowed by the "AI-defined vehicle" narrative in 2026, yet it is the most headache-inducing for the engineering side.
Let's lay out the data first: L4/L5 autonomous driving computing power has broken through 1,000 TOPS, with NVIDIA's Thor reaching 2,000 TOPS. The process nodes are 4nm/5nm, the thermal flux density is 250W/cm², and local hotspots exceed 500W/cm². 3D packaging makes thermal coupling more complex. The in-vehicle environment also stacks high temperature/high altitude/extreme cold/high-frequency vibration—traditional air cooling and passive heat dissipation are out; cold plate liquid cooling is the mainstream direction for autonomous driving domain controllers.
However, putting liquid cooling "on vehicles" is not just about moving data center solutions over. There are three unique pain points for in-vehicle applications:
- Condensation: When the coolant temperature drops below the dew point in the domain, condensation forms on the outer wall of the flow channel and drips onto the circuit, causing a direct short circuit. Huawei's MDC approach is to add holes to the shell + an e-PTFE breathable membrane; a more advanced approach is the "breathable valve + desiccant" CMD structure.
- Sealing and Leakage: O-ring aging + pipeline vibration; a certain automaker has already had a recall case due to O-ring failure. Fluoroelastomer O-rings currently still need to be procured from Japan/Germany, accounting for 15%–20% of the system cost.
- Low-Temperature Fluidity: At -40℃, the viscosity of water-based coolants increases sharply, and the pump may fail to start. Switching to low-viscosity silicone oil requires further cost increases.
The materials side is also keeping pace. Qi Dan, Chief of Material Development at the FAW R&D Institute, pointed out a bottleneck chain in the forum: "For silicon nitride heat dissipation substrates, imported powder costs 380–400 RMB/kg. If domestic production can achieve 120–150 RMB/kg, there is huge room for cost reduction—but for high thermal conductivity (targeting over 110W/m·K) silicon nitride powder, China still lags behind Japan and South Korea." More forward-looking diamond nano-films are "not as expensive as imagined, but there are still technical issues for automotive applications."
The spillover of thermal management is also occurring. Gu Jifeng, Deputy Director of Thermal Management Technology at Air International, discussed humanoid robot solutions in the forum: the heat exchange of servo joints accounts for 60%–80% of the whole machine, and the thermal flux density of computing chips (NVIDIA AGX class) is 150–300W/cm². The chip throttles when it exceeds 85℃—"Robots are actually somewhat like humans, with multi-point heat dissipation. What we want to do is embed a bionic 'vascular' water jacket into all heat-generating components, with centralized heat dissipation in the chest cavity." This shares the same technological matrix as autonomous driving domain controller liquid cooling, just with tighter space and higher weight sensitivity.
An easily overlooked signal: Thermal management is spreading from "vehicles" to "robots + low-altitude + eVTOL + AI computing centers." The same set of liquid cooling/phase change/high thermal conductivity material technologies 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—because this is the strongest hand forcing thermal management transformation in the next three years.
The "National Implementation Plan of the People's Republic of China on the Montreal Protocol (2025-2030)" issued by the Ministry of Ecology and Environment and other ministries is very strict: by January 1, 2029, the use of HFCs for controlled purposes must be reduced by at least 10%. Starting July 1, 2029, the automotive industry will ban the use of refrigerants with a GWP > 150 in the air conditioning systems of newly announced M1 vehicles. R134a (GWP=1430) is directly eliminated, leaving a window of less than three years. Coupled with the R134a quota being "welded at the current level," the prices of third-generation refrigerants have hit a ten-year high—it is no longer a question of "whether to switch," but "if you don't finalize the route now, it will be too late for mass production in 2027."
Currently, three routes are running in parallel, each with its own calculations:
Tao Hong, Chief Technical Expert of Thermal Management Systems at Highly New Energy, gave a pragmatic judgment in the forum: "COP and charge volume are always topics that must be done well. Environmental protection must be economically sustainable, otherwise it cannot be sustained. R1234yf can at least be used until the 2029 node, no need to do it overnight." But he also confirmed that Honeywell is negotiating with the Chinese government/factories regarding domestic capacity—implying that the supply and cost of 1234yf will still fluctuate after 2029.
Liu Jun's team is already working on an R290 double-secondary-loop indirect heat pump, "cross-generation compatible design, where interface changes can achieve switching between R1234yf/R290/R744 three refrigerants"—this is a "don't put all your eggs in one basket" approach.
The calculations on the commercial vehicle side are more convincing to decision-makers. Huang Jiayue, Deputy Director of the Thermal Management Development Department at BAIC Foton, calculated on-site: "A 4.5-ton pure electric logistics vehicle runs 100,000 km a year. A 10% optimization in thermal management energy efficiency saves about 24,000 RMB annually, equivalent to a +15% increase in net profit—the net profit margin in the freight industry is generally less than 5%, so this calculation directly relates to the survival of the enterprise." Driven by the dual forces of policy and TCO (Total Cost of Ownership), the switch to low-GWP refrigerants in commercial vehicles will actually be more resolute than in passenger cars.
Conclusion:
Returning to the question at the beginning—in the era of AI-defined vehicles, how will thermal management develop?
The answer is no longer linear iteration like "making heat sinks bigger and adding more valves," but a simultaneous leap across four dimensions: • Control side: Rules → AI prediction + MPC + vehicle-cloud collaboration, with a +20% winter range as the first quantifiable dividend; • Architecture side: Dispersed → dual-indirect + multi-way valves + standardization, but vigilance is needed against heat leakage and maintenance costs caused by integration; • Object side: From protecting the three electrics → accommodating autonomous driving chips/robots/low-altitude, with a thermal flux density of 250W/cm² as the new engineering ceiling; • Compliance side: July 1, 2029, is a hard countdown. R1234yf transition + R290/R744 dual-track is the most certain investment direction for the next three years.
Thermal management is transforming from an auxiliary subsystem of "passively ensuring hardware does not overheat" into a "vehicle energy dispatch center that actively defines range, experience, and safety"—this statement sounded like a slogan two years ago, but after listening to the 2026 forum, it is already a project running on everyone's PPTs.
Wang Tianying's quote from IM Motors, "Dual-indirect is the 'last mile,' and it's coming soon," and Zeng Like's quote from Valeo, "If the counter-intuitive strategies generated by AI have good value, engineers need to learn from AI"—reading these two together is probably a microcosm of the mindset of thermal management professionals this year: the "last mile" of architecture is being stepped on, the paradigm of control is changing, and the policy countdown is ticking.
It cannot slow down.