Flashy AI headlines grab attention. Large language models like Anthropic’s Claude and China’s DeepSeek drop every few weeks. Yet leaders at China’s top self-driving truck firms say these don’t budge their deployment schedules. Pony.ai CEO James Peng put it bluntly last week: “The world’s best linguistics [expert] doesn’t mean he’s a good driver.” AI’s a broad term, he added. Completely different from driving skills. Zero relevance.
Autonomous trucking demands sensors, chips, algorithms mimicking human drivers. Real-world road data fuels the systems. Not text corpora powering ChatGPT. Inceptio CEO Julian Ma sticks to mid-2028 for full commercialization. That’s when his firm expects 5 billion kilometers of driving data—enough for AI world models to simulate 50 billion kilometers of experience. Heavy-duty trucks could then run public roads sans humans in select areas. Fast? Sure. But no shortcuts from chatbot buzz.
Inceptio leads in miles logged. ARK Invest’s Big Ideas 2026 report pegged them at 250 million commercial autonomous miles as of January, dwarfing Pony.ai’s 4.2 million and U.S. trio Aurora, Kodiak, Gatik’s combined 8.9 million. Late April update: 700 million kilometers. Target: 1 billion by year-end. AI helps pinpoint test scenarios now. Still, partnerships with makers. Regulatory nods. Those gatekeep scale.
Peng’s Pony.ai unveiled upgrades at Beijing’s Auto China 2026. PonyWorld 2.0 collects data smarter, trains faster. They debuted a driverless light-duty truck with CATL—fully redundant, L4-ready for urban logistics like parcel runs, cold-chain delivery. Cuts freight costs 40-50% per kilometer versus humans. Cargo space triples low-speed rivals. Range: 320-450 km. All-weather, full-scenario. Impressive. But CEO dismisses LLM hype. “When we process language, play sports, drive—we use different skills.”
Regulators paused new autonomous licenses recently. Baidu Apollo Go robotaxis stalled mid-Wuhan traffic, sparking collisions, Bloomberg reported. Safety first in autos, Ma notes—tougher than embodied AI like humanoid bots. Companies push; Beijing follows once convinced.
And broader hurdles persist. Compute demands skyrocket for L4: 100-150 EFLOPS training, per Caixin Global. Data scarcity across scenarios. Video optimization. Simulation-to-reality gaps nag physical AI, as Gasgoo Auto News highlighted from the show. World models predict physical dynamics. Yet power limits simulations. Virtual trials can’t fully bridge to chaos of highways.
ZYT offers contrast—and acceleration hints. Their mobility foundation model outdrives CEO Shen Shaojie on Shenzhen streets. Trained on eclectic video: roads, drones, vacuums, bikes. Adapted to heavy trucks in six weeks. Partners cover 98% of China’s truck market—XCMG, SHACMAN, SINOTRUK for highway systems this year. FAW took 35.8% stake, easing compliance post-DJI ties. “It actually drives better than me,” Shen told Reuters. Fuel savings: low single digits percent. Compelling for fleets. Still, compressing for cheap chips ongoing. No U.S. plans.
China’s edge shows in scale. Pony.ai’s Gen-4 trucks mass-produce thousands this year with SANY—1 million km service life per unit. Inceptio deploys cabless cargo haulers in Inner Mongolia mining. Yet U.S. lags data: Tesla’s China FSD hobbled by export rules till recent local training nods, Electrek noted. Beijing’s Five-Year Plans prioritize tech self-reliance. Auto China buzzed physical AI—XPeng’s L4 pilots target driverless 2027, Horizon calls it embodied AI’s first big test, per People’s Daily Online.
But timelines hold firm. Data grinds accumulate. Two years to 5 billion km? Aggressive. Regulations tighten post-incidents. Partnerships essential. LLM fireworks? Spectacle. Not fuel for rigs.
Freight’s trillion-dollar prize tempts. China logs billions ton-km autonomously already—Pony.ai over 1 billion. Costs plummet long-term. Safety climbs. Jobs shift. Execution decides winners.
Leaders know. Hype accelerates stock ticks. Miles move trucks.
WebProNews is an iEntry Publication