Recorded 10 September 2026 · 50 min
Xpeng does not think of itself as a car company, and president Brian Gu says the reason is mechanical, not marketing. A modern car already forces a maker to build its own chips, VLA models, data pipelines and training infrastructure. The same stack carries over to flying cars and humanoid robots. Cars, he says, are becoming "more and more like robots."
Nicolai Tangen, CEO of Norway's sovereign wealth fund, presses Gu on China's EV market, where new energy vehicles now exceed 60% of new car sales, and on whether Chinese speed can travel abroad. The tension is between a real lead in electrification and development cycles, and unfinished work on brand trust outside China. Gu also concedes a generation gap in cloud GPUs. He sidesteps a direct answer on whether Chinese language models will overtake American ones.
Intelligence over electrification — ▶ 1:57
Nine years ago, when Xpeng was a tiny startup and every EV pitch was about the energy transition, Gu told investors and partners that long-term differentiation lay in intelligence, not electrification. The company has since built a full stack across hardware, software and now AI. His test of an intelligent car has layers. It understands your behaviour and preferences, then handles tricky roads at level 2-plus smoothly, "without causing any concerns or anxiety."
Merging cabin and driving — ▶ 3:49
The next step is fusing autonomous driving intelligence with the smart cabin. Xpeng's VLA 2.0 edition, launching the day of the interview, lets a passenger command driving by voice, such as pulling over or making a quick stop. Gu calls level 4, fully autonomous with no steering wheel, "a matter of time" that will happen in our lifetimes. Once it is everywhere, driving time becomes work or entertainment. He expects it to reshape real estate, communications and how people allocate time.
Charging beats range — ▶ 5:56
Current battery chemistry will last, and solid state is more than 5 years from commercial feasibility. Range stays at 500-1000 km, and Gu calls 1000 km the ceiling: "Anything more than that is waste." The real change is charging, down from 30-60 minutes to 5-10, close to a gas station stop. Xpeng buys cells from Chinese suppliers other than CATL, but designs its own battery management system, packing technology and electric platform in-house.
EVs are now mainstream — ▶ 8:12
In the past month China's new energy penetration exceeded 60%, possibly close to 65%. Battery EVs alone outsold combustion cars for the first time. With plug-in hybrids, new energy is over 60% of new car sales. Subsidies are being withdrawn because a mainstream product needs none, and gas stations are being replaced by charging stations. Gu adds that hybrids carry a battery and computing power, so intelligence spreads with them. The trend is now driven by intelligence as much as energy.
The fittest gym — ▶ 9:57
Gu calls China's new energy vehicle market the most competitive in the world, "the fittest gym in the world" for carmakers. Surviving it takes more than before. Sourcing technology from tier-one and tier-two suppliers no longer works because cars are becoming robots. Winners need automotive engineering, AI and software, electric powertrain and service, plus scale and global reach. He gives Xpeng "a decent chance" but calls it a very uphill battle for everybody.
Leads at home, learning abroad — ▶ 11:50
China leads in electrification, with battery technology, capacity and cost, and a local champion is a matter of time. But Chinese players are not used to leading across continents. They must build brand trust, service and local relationships. Xpeng opened its first Norway store at the end of 2020 and now sells in over 70 countries. It must shed old quality perceptions and stand apart from a dozen other Chinese brands going global. Volkswagen's stake reflects a thesis: Xpeng brings speed and consumer insight, partners bring scale.
Where Chinese speed comes from — ▶ 18:02
Gu credits four forces. China's software, internet and consumer electronics expertise comes first, then competitive pressure, then fast consumer adoption as with mobile payment, and finally supply chain advantages. Vehicle development has shrunk from 3-5 years to about 2, and features arrive with every OTA update. He adds an organisational point. A 12-year-old company has no department walls or decades-old processes, while global players decide at headquarters and the decision may reach China a year later.
Flying car in a trunk — ▶ 21:41
Air Ridge, which Xpeng incubated and holds a minority stake in, has a "land aircraft carrier". It is a car that carries a two-person eVTOL, releases it for about 30 minutes of flight and reloads it automatically. It has more than 7,000 indicative orders worldwide and awaits Chinese approval. Gu still expects mass adoption to take 3-5 times longer than in automotive and autonomous driving, because of heavy regulation and a long educational process for buyers and operators.
Humanoids as general-purpose data — ▶ 24:48
Robots reuse Xpeng's edge computing, models, data, training and electric motors. Gu chose a humanoid because the world is built for human bodies, so it is the most versatile form. He sorts Chinese rivals into body specialists like Unitree, brain specialists and component builders such as tactile hands. Xpeng says it has developed all three pieces and aims to unify them. He dismisses fixed-task factory robots as less exciting, since robotic arms already exist. The best data, he argues, comes from deploying a general-purpose body at volume.
Own chips, overlapping supply chain — ▶ 34:08
More than 50% of Xpeng's EV supply chain overlaps with what robots need, but most parts are not off the shelf and must be co-designed. Hands, joints, motors and actuators are designed in-house. The Turing chip, at 700 TOPS, replaced the Nvidia chips used before and, Gu says, rivals the latest Western on-board silicon. Memory cost and supply are a challenge but "within control". He admits a generation gap remains in GPUs and cloud computing.
Cities in three dimensions — ▶ 38:36
In 20 years Gu expects cities to be dispersed, with no central business district, because level 4 and robotaxis remove the need for proximity. Last-mile deliveries arrive by low-altitude craft or autonomous vehicles at the door or roof. Travel moves into designated low-altitude corridors, making cities three-dimensional. Humanoids handle routine and repetitive tasks, and he floats a scenario with more humanoids than humans on the street. Whether people end up happier, he says, depends on spending time on judgment rather than the 90% now taken by routine.
Electrification has stopped being a differentiator in China. It is the default. In the past month new energy vehicles passed 60% of new car sales, possibly close to 65%, and battery EVs alone outsold combustion cars for the first time. Subsidies are being withdrawn, because nobody subsidises a mainstream product. Gas stations are giving way to charging stations. Hybrids carry a battery and computing power, so intelligence spreads with them. Brian Gu says Xpeng read this early. Nine years ago, when every EV pitch was about the energy transition, he told investors the lasting edge would be intelligence.
The hardware questions are closing fast. Current battery chemistry will last, and solid state is more than 5 years from commercial feasibility. A range of 500-1000 km is enough, and anything beyond 1000 km is "waste". Charging has fallen from 30-60 minutes to 5-10, close to a gas station stop. Xpeng buys its cells from Chinese suppliers other than CATL and designs the battery management, packing and electric platform itself. What stays open is what the car can do on its own. Gu's test is whether the car understands you, then whether it handles tricky roads at level 2-plus without causing anxiety. Next comes merging driving and cabin intelligence, so a spoken command changes how the car drives. Level 4 is, for him, a matter of time. When it arrives, driving hours become work or entertainment, and industries such as real estate and communications change with them.
That stack explains why Xpeng reads itself as a robot company. A car already forces it to build edge chips, VLA models, data pipelines and training infrastructure. A humanoid needs the same chips, models, data and training, plus efficient electric powertrains and motors. The in-house Turing chip, at 700 TOPS, replaced the Nvidia chips Xpeng used before. Gu says it rivals the latest Western on-board silicon. More than 50% of the EV supply chain overlaps with what a robot needs. But most of those parts are not off the shelf. Xpeng designs the hands, joints, motors and actuators itself and co-designs the rest with suppliers, because no ready supplier base exists for advanced robots.
The humanoid form is a deliberate bet, not a stunt. Offices, shops and restaurants are built for human bodies, so a human-shaped robot is the most widely deployable one. It is also the best data collector. Gu accepts that simulation and egocentric video help. He is cool on the Chinese training centres that buy robots to gather data, which he says serve specific tasks. The most accurate data, he argues, comes from a general-purpose body in regular use at large volume. He sorts Chinese rivals into three groups: body specialists such as Unitree, brain specialists, and builders of components such as tactile hands. Xpeng says it has completed development across all three and aims to unify them. Fixed-task factory robots interest him less, since robotic arms already run car plants.
The same logic makes the car business brutal. China is "the fittest gym in the world" for carmakers, and survival now takes more than it did. Sourcing technology from tier-one and tier-two suppliers no longer works. A winner needs automotive engineering, AI and software, powertrain and service, plus scale and global reach. Gu gives Xpeng a decent chance and calls it a very uphill battle. Carmakers rooted in traditional automotive face a culture problem. Moving to software-defined vehicles was hard. Jumping to robotics is harder.
Speed is the mechanism that lets Xpeng try. Gu credits China's software and consumer electronics expertise, market pressure, consumers who adopt new technology fast, and supply chain efficiency. The measurable effect is vehicle development shortened from 3-5 years to about 2, with features arriving in every OTA update. He adds an organisational point. A 12-year-old company has no department walls and no decades-old processes. Global players decide at headquarters, and by the time a decision reaches China it can be a year old.
Speed does not travel automatically. Gu says Chinese players are not used to leading in a multinational world. They must learn brand trust, service and local relationships. Xpeng opened its first Norway store at the end of 2020 and now sells in over 70 countries. It must shed old quality perceptions and stand apart from a dozen other Chinese brands going global. Volkswagen's stake is his model of the trade. Xpeng brings faster innovation and consumer insight, and the partner brings scale to monetise it. The lead is in speed. Scale is what partners supply.
The stack reaches beyond roads. Air Ridge, an incubated company in which Xpeng holds a minority stake, has built a "land aircraft carrier", a car that carries a two-person eVTOL, releases it for about 30 minutes of flight and reloads it automatically. It has more than 7,000 indicative orders and awaits Chinese approval. Gu still expects mass adoption to take 3-5 times longer than in automotive, because of regulation and buyer education. His city of 20 years out is dispersed, three-dimensional and busy with humanoids, perhaps more than humans on the street.
The weak joint is where the stack meets the frontier. Asked when Chinese language models will pass American ones, Gu declined the framing. It is a different race, he said. US models serve Western governments and companies. Chinese models serve small firms and emerging markets, with open source and low cost. The ecosystems will separate. That is a description, not a timeline. He also concedes a generation gap in GPUs and cloud computing, even as on-board chips reach parity. Yet his own list of what a robot company needs includes AI infrastructure for data and training. The deployment-data thesis has a similar gap. It assumes general-purpose robots are already out in volume. Gu says only that routine home and business tasks arrive very soon, depending on how quickly humanoids acquire more of them. The whole case rests on that rate of acquisition, and he puts no number on it.
The open question. Gu says Chinese and US language models are "a different race" and that the ecosystems will separate. That reframes the question rather than answering it. He also admits a generation gap in GPUs and cloud computing, even as Xpeng's on-board chips reach parity. Yet his own list of what a robot company needs includes AI infrastructure for data and training. The humanoid thesis, that deploying general-purpose robots at volume gives the best training data, assumes the robots are already out in volume. Gu offers no timeline beyond "very soon" for routine tasks. Watch how fast humanoids actually acquire tasks.
Watch the full video at https://www.youtube.com/watch?v=DH8IMwDN0LQ.
Comments ()