Flashy gains in large language models promise smarter chatbots and code. But China’s self-driving truck frontrunners shrug them off. Executives at Inceptio and Pony.ai insist those breakthroughs carry zero weight for hauling freight unsupervised. Timelines hold firm. Mid-2028 remains the target for commercial driverless operations on key routes.
Pony.ai CEO James Peng cuts straight to it. “The world’s best linguistics [expert] doesn’t mean he’s a good driver,” he told CNBC. “AI is a very broad term. They’re completely different things. Absolutely … zero relevance.” Peng draws the line between parsing sentences and dodging obstacles at 80 kilometers per hour. Language skills won’t brake for a pedestrian. Or merge through construction.
Inceptio CEO Julian Ma echoes that divide. His firm sticks to mid-2028 for heavy-duty trucks running empty cabs in select zones. “Achieving that goal in about two years is already quite fast,” Ma said in the same CNBC interview. The company logged 700 million kilometers driven by late April. One billion by year-end. That’s real-world miles, mostly manned for safety, feeding algorithms that build so-called world models—digital twins of highways crammed with variables like weather, potholes, erratic drivers.
Data hunger drives the pace. Inceptio eyes 5 billion kilometers by late 2028. AI extrapolates that to 50 billion in simulated experience. Enough, they claim, for hands-off runs in bounded areas. But collecting it demands trucks on roads today. Sensors. Chips. Not just better prompts for DeepSeek or Claude.
Pony.ai pushes forward too. At Auto China 2026 last week, it unveiled an L4 autonomous light-duty truck with battery giant CATL. Fully redundant systems for braking, steering, power. Cargo space triples low-speed rivals. Range hits 450 kilometers. Aimed at urban deliveries—parcels, groceries, cold chain. “Expected to reduce freight cost per kilometer by 40% to 50%,” Pony.ai stated, per coverage from YouTube reports on the event. Fourth-gen heavy-duty trucks enter mass production this year for cargo hauls.
Regulators loom large. China paused new autonomous licenses after Baidu Apollo Go robotaxis crashed in Wuhan—collisions, mid-road stops— as noted in the CNBC piece citing Bloomberg. Safety first. Trucks carry tons; one glitch spells disaster. Ma calls autos AI’s toughest test. “It involves safety,” he said. “Exceeds the difficulty of embodied AI to some extent.” Partnerships with truck makers. Policy nods. Those bottlenecks persist, hype or no.
Inceptio leads in miles logged. ARK Invest’s Big Ideas 2026 report pegged them at 250 million commercial autonomous miles by January, outpacing Pony.ai’s 4.2 million and U.S. trio Aurora, Kodiak, Gatik’s combined 8.9 million. That edge comes from China’s freight grind—endless hauls between factories, ports, warehouses. AI helps pick test scenarios. But miles don’t lie.
And yet. Broader buzz swirls. Caixin Global detailed auto giants like Geely showing Eva Cab robotaxis with 3,000 TOPS compute at Beijing’s show. Huawei eyes Level 3 commercial by 2027, Level 4 urban by 2028. Bosch tests Level 3 on public roads up to 120 km/h. But trucks? Different beast. Predictable highways beat city chaos.
Pony.ai expands beyond. Joint venture with Sany Trucks and Sinotrans launched unmanned fleets, per Reddit discussions linking to reports. Mining ops too—self-loading in harsh conditions. Inceptio runs in Yimin mine, fleets outperforming diesel, as Instagram freight analysts note. Scale builds quietly.
Challenges stack up. Data from manned runs only. World models demand physics grasp—tires on wet asphalt, wind gusts on loads. LLMs excel at patterns in text. Vehicles need prediction in chaos. Peng’s point: skills don’t transfer. Process language? Fine. Play sports? Separate. Drive? Its own domain.
U.S. lags in truck miles but advances. Aurora runs driverless Class 8 on Texas routes. Still, China’s volume crushes. IMARC forecasts China’s AV market hitting $256 billion by 2034, CAGR 30%. Baidu, Pony.ai, WeRide pilot robotaxis at scale—Pony.ai over 1,000 vehicles, aiming 3,000 by year-end per official reports.
So timelines endure. Ma plans to “make it happen” before regulators buy in. Peng upgrades PonyWorld 2.0 for smarter data hunts. AI aids. But doesn’t dictate. Boom. No speedup.
Freight waits for proof, not promises. Two years. Eyes on 2028.
WebProNews is an iEntry Publication