► Xpeng is about to launch the next generation of self-driving cars
► We’ve been to the company’s home market to see how they work
► They’ll have super-powerful processors and an all-new AI model
AI already has a strong foothold in modern society. Loads of my mates now use ChatGPT as their replacement for Google, while AI content generators are flooding the media market with (largely incorrect) information on everything from politics to celebrity gossip.
It’s already reshaping the way we interact with the internet, but the Chinese technology company, Xpeng, is looking beyond such trivial uses for AI. It sees a future in which AI lives amongst us in physical entities – and it’s starting by giving it control of its cars.
In the first quarter of 2026, Xpeng will launch its new SEPA 3.0 platform. It’ll sit beneath its next generation of electric cars, such as the X9 MPV and P7 sports saloon. Mechanically, the chassis is near identical to the SEPA 2.0 platform beneath the G6 SUV and P7+ saloon – but it gets a huge computing upgrade that Xpeng reckons can provide enough processing power to make these cars ready for fully autonomous driving.
The brains of the operation are Xpeng’s latest Turing AI chips, on which will sit the company’s shiny new AI model. This pairing promises to be faster and more human-like than anything we’ve yet seen from Tesla, or indeed established AI developers such as OpenAI and Google. Scroll down to find out how it all works.
So, what’s so special about the Turing chip?
It has a lot of waft. Xpeng tells me it can perform 750 trillion operations per second (TOPS). To put that performance into perspective, the Xbox Series X can manage around 12 trillion operations per second, while the punchiest version of Apple’s current Macbook Pro can compute 38 trillion operations per second.
What’s more, Xpeng says it can stack its Turing chips to create even faster processing speeds. To prove the concept, the brand has packed three of them into its humanoid AI-powered robot, IRON, which the brand says is capable of 3000 trillion operations per second.
The same tech will eventually worm its way into Xpeng’s cars. Dr. Xiaoming Liu, Xpeng’s head of autonomous driving, tells me his future cars will be capable of 2000 trillion operations per second, as he believes having plenty of processing headroom is the key to delivering a safe fully autonomous driving system.
Liu also mentioned his (unnamed, but I’m pretty sure he was referring to Tesla given the processing power he quoted) competitors are planning to deliver Level 3 autonomous driving systems on chipsets that can only perform a couple hundred trillion operations per second. But what good’s a fancy chip if you don’t have any software for it?
Xpeng VLA 2.0: the new way to do AI
Xpeng has designed a new type of AI which Liu says ‘has dropped the crutch of language.’ That means, rather than translating its inputs into a format humans can understand before acting on them, it takes direct visual cues from the landscape and makes decisions based solely on what it sees. It’s called the VLA 2.0 Large Model.
Here’s how it works. Xpeng’s newest cars will be fitted with an array of seven cameras, which will track everything happening around the car. It feeds that data at a speed of 5000 inputs per second into VLA 2.0, which has been trained on 65,000 years-worth of driving footage and has built up a knowledge base of what constitutes good and bad driving.
The AI model drives the car on two categories of tokens, called trajectory tokens and latent tokens. The former group relates to the forces acting on the car – such as speed, braking and steering angle – while the latter relates to the world. This includes the actions of other cars and pedestrians, as well as hazards such as road closures or blocked pathways.
What’s particularly clever is that VLA 2.0 uses the data it gathered during its training to infer what will happen with the cars, people and objects surrounding it – and it’s prepared to act on their entropy. For example, if a car is trying to sneak out of its lane and cut in front, the AI will recognise the behaviour from its time in school and know how to act to avoid an accident. In short, it’s thinking like a human.
It can also react like a human, which is to say quickly. The reason why Xpeng designed an insanely powerful chip is because it needed something fast enough to match the refresh rate of the cameras feeding the AI with data, therefore allowing it to avoid accidents in real time.
Western manufacturers are taking Xpeng’s progress seriously, too. Volkswagen has already signed a partnership with Xpeng and will be the first external company to use its Turing chip and VLA 2.0 AI in its upcoming production vehicles for China. Whether or not it’ll ever reach Volkswagen’s European models remains to be seen.
Why go through all this effort? What’s the end goal for Xpeng?
Xpeng isn’t shy about the fact it benchmarks its products against Tesla. In fact, Dr. Xiaoming Liu told me he thought Tesla’s Full Self-Driving technology was the best in the segment right now – and he’s quite open about the fact that he’s been using Musk’s models to improve his own.
However, after spending a week with Xpeng in its home market, it became increasingly obvious the brand was doing more than just benchmarking its products against Tesla. From my outside perspective, it seems it’s also benchmarking Musk’s business model.
I’ll explain. Musk’s numerous technology companies – Tesla, Space X, Starlink, The Boring Company and OpenAI to name a few – appear completely unrelated to the casual observer. But every new enterprise Musk explores provides additional support for his ultimate goal of landing on Mars. In essence, he’s spreading the workload (and the financial risk) over several firms. Like him or loathe him, that’s smart.
He Xiaoping, Xpeng’s CEO (pictured above), has adopted a remarkably similar strategy only, instead of attempting extraterrestrial colonisation, he’s trying to create a reliable artificial general intelligence. To get there faster, he’s partitioned his company into three segments to efficiently pull data from as many human touchpoints as possible.
Xpeng’s autonomous driving models are being trained by observing our driving habits. The company has also deployed its IRON humanoid robots into retail centres in China, so its language and reasoning models are being trained by interacting with us. Xpeng’s AI is even learning how to fly thanks to the efforts of Aridge.
Autonomous cars are just the beginning. Liu said: ‘AI cars will be the first production [item] where physical AI will be deployed into the real world’ – and making these AI cars work will be crucial to the success of Xpeng’s wider AI programme. Liu referenced his colleague LC Mi, head of Xpeng’s humanoid robot programme, saying: ‘we always joke that if I fail then the robots fail.’
That’s a lot of weight for two tonnes of steel, rubber and copper to bear. Let’s see how it pans out.