Urban assisted driving is expanding into the 100,000 yuan vehicle segment, yet smart-driving vendors are still competing for ever-higher computing power ceilings. One side is expanding coverage at lower cost, while the other is reserving capacity for larger on-vehicle models. These two demands running in parallel make it difficult for this competition to be decided by a single computing power figure.
On October 29, XHEART launched the X7 Ultra, which by the company's own account achieves an equivalent large-model computing power of 700 TOPS for smart-driving scenarios; it also disclosed the development goal for its next-generation X9: 3,000 TOPS of single-chip large-model computing power and 700GB/s of memory bandwidth, planned for release in 2027.
The change in computing power demand first shows up in the tasks that models take on. When XPeng introduced its new second-generation VLA in August this year, it said the on-device model's parameter count had expanded 3.5 times and introduced memory of the past 30 seconds of information plus scene prediction for the next 6 seconds. The model needs to call on longer historical information and simulate possible future changes.
XHEART CEO Li Zonglin laid out the company's core technology路线 on October 29: chips defined by large models. As intelligent driving moves from multiple small models to end-to-end large models, the way chips are designed also needs to change accordingly.
Momenta CEO Cao Xudong said at the launch event that the next stage of advanced autonomous driving will be chips defined by large models. To achieve scaled L4 autonomous driving, the vehicle side will need to carry large models with tens of billions or even hundreds of billions of parameters, which will place higher demands on chip computing power, energy efficiency, bandwidth, and cost.
These tasks must be completed promptly while driving. Larger parameter counts, longer context, and more simulation all increase computing and storage pressure. Beyond computing power, bandwidth has also become critical. If model weights and intermediate data cannot be delivered to the compute units in time, adding computing power may not proportionally shorten inference time. The X9's simultaneous increases in computing power and bandwidth targets respond to exactly this kind of demand.
For suppliers that need to lock in architecture in advance, this is also a bet on the next generation of vehicles. Insufficient reservation may limit model upgrades; excessive reservation adds hardware, power supply, and heat dissipation burdens. As a result, model companies are participating earlier in hardware design so that architecture can be optimized around actual tasks.
XHEART has chosen to collaborate with Momenta. Under Momenta's plan, the R7 Ultra solution based on the X7 Ultra will expand model parameter scale to five times its original size. The two sides are putting models and hardware into the same upgrade plan in an attempt to reduce adaptation losses. But this route already has competitors laying groundwork.
NVIDIA offers tools such as DriveOS, CUDA, and TensorRT around DRIVE AGX Thor. When automakers choose a platform, they are also choosing the toolchain on which subsequent model development and deployment will depend. New entrants seeking customers need to take on far more work than simply delivering hardware.
Domestic suppliers already have mass-production scale. Horizon's planned Journey 7 targets L3 and L4, with native optimization for larger-parameter HSD smart-driving models and cabin models, and is expected to complete tape-out in early Q2 2027.
This means XHEART must confront established platforms that already hold share while also competing with rivals' next-generation products still under development. Software-hardware coordination is becoming a common choice among multiple suppliers.
Automakers' in-house development adds yet another competitive route. XPeng disclosed in October that the Youyou Robotaxi, which obtained Guangzhou's L4 demonstration operation qualification, carries four self-developed Turing AI chips, reaches 3,000 TOPS of effective vehicle computing power, and runs the second-generation VLA. Its XPENG AI ENGINEERING cooperation system has also brought smart driving, chips, and toolchains into the scope of external cooperation. Some potential buyers are becoming solution providers.
It is necessary to distinguish here that XPeng's 3,000 TOPS is the combined total of multiple chips across the vehicle, while XHEART's X9 is a single-chip target. Different vendors also use different computing precisions and equivalent computing power definitions, so these figures cannot be used directly to rank actual driving capability.
But competition at the high end of platforms cannot replace the immediate need for popularization. Horizon disclosed in September that the 128 TOPS Journey 6M is expanding toward the mainstream 100,000 yuan market. As of the end of August, seven automaker brands had already achieved platform-based urban assisted driving rollout based on the product; a single Journey 6M solution can also use air cooling, adapting to the power supply and heat dissipation constraints of fuel vehicles.
For these models, procurement must calculate the cost of the entire system: how much experience improvement the extra computing power can bring, whether more complex cooling is needed, and whether the corresponding development investment can be sustained. The 3,000 TOPS aimed at next-generation models and the current popularization of urban assisted driving correspond to different product budgets and delivery requirements.
XHEART has already built a certain mass-production base. According to company disclosures, the X7 series shipped more than 200,000 units cumulatively in eight months of mass production, is installed in 10 mass-produced models, and has won orders for more than 40 models.
3,000 TOPS shows the direction in which smart-driving vendors are investing for next-generation models, but the market will not pay just for a computing power ceiling. What XHEART needs to win is automakers' belief that, among existing suppliers, self-developing automakers, and low-cost solutions, choosing it with the same hardware and development budget will deliver more deployable smart-driving capability.
Comments