Horizon Robotics Inc.
9660 · HKEX · China
Makes AI chips and matching software so Chinese car companies can build self-driving systems without foreign components.
Horizon Robotics designs automotive AI chips, called BPUs, and the Journey compiler that runs on them — the two are built together from the start so that the chip exposes custom instructions only Journey knows how to schedule, meaning an OEM that builds a perception model for one automatically depends on the other. Because those models then enter the 18–24 month AEC-Q100 safety certification process tied to a specific BPU generation, customers like Geely and SAIC are effectively locked to that chip's architecture for the length of a model-year cycle — switching to a rival chip means abandoning Journey workflows, retraining engineers, and restarting the certification clock. The whole system, however, can only be manufactured at the advanced process node TSMC allocated when the BPU was designed, because the power-efficiency targets that justified AEC-Q100 qualification in the first place cannot be met on an older or alternative geometry. If US-China export controls cause TSMC to cut off that node access, the co-design advantage runs in reverse — the Journey compiler is optimized for a chip that can no longer be built, existing qualifications are voided, and the same switching costs that trap OEM customers also trap Horizon on its way to any fallback.
How does this company make money?
The company charges automotive OEMs and Tier 1 suppliers a price per BPU chip sold, with the price depending on processing capability and the volume the customer commits to buying. It also collects software licensing fees for the Journey development tools and charges for ongoing technical support contracts that help customers integrate and optimize the chip in their vehicles.
What makes this company hard to replace?
Automotive engineers at OEMs build their entire model development and deployment workflow inside the Journey software environment, using tools written specifically for the BPU. Switching to a rival chip means abandoning those workflows, not just the hardware. It also means restarting the AEC-Q100 safety certification, which takes 18 to 24 months per chip generation. On top of that, validated reference designs and manufacturing relationships already built into the Chinese automotive supply chain create additional practical barriers to changing course.
What limits this company?
TSMC is the only factory that can make the BPU at the precise manufacturing process the chip was designed for. The power efficiency that qualifies the chip for automotive use cannot be achieved on an older or alternative process. TSMC's automotive-grade capacity runs on multi-year allocation cycles, and no amount of money can instantly expand those slots or redirect them to a different factory.
What does this company depend on?
The company cannot operate without TSMC's advanced process node fabrication capacity to manufacture the BPU. It relies on ARM processor IP licensing for the CPU components inside its chips. It needs automotive-grade memory suppliers that meet AEC-Q100 standards. It requires export approvals from China's Ministry of Industry and Information Technology. And the entire product depends on continued development of the Journey software stack and its compiler optimizations.
Who depends on this company?
Chinese automotive OEMs like Geely and SAIC depend on it as their primary source of domestically developed ADAS processing. If it stopped, Tier 1 automotive suppliers would have to redesign their smart driving systems around entirely different chip architectures. Sensor manufacturers that make cameras and LiDAR units tuned to work with the BPU's processing pipelines would lose the optimization support those integrations require.
How does this company scale?
The Journey development tools and AI model libraries can be copied and deployed to any number of customers at almost no additional cost once they are built. What cannot scale cheaply is the deep engineering knowledge required to optimize neural networks for real-time vehicle safety — that expertise takes years to build and cannot be hired or bought quickly.
What external forces can significantly affect this company?
US-China semiconductor export controls are the most direct threat, potentially cutting off access to the advanced design tools and manufacturing processes needed to build the next generation of BPU chips. On the other side, Chinese government mandates pushing domestic car companies to use locally developed technology actively work in the company's favor. Separately, EU and US automotive safety regulations require different validation standards for AI driving systems, which complicates any attempt to sell into export markets.
Where is this company structurally vulnerable?
If US-China export controls force TSMC to cut off advanced manufacturing access for this supplier, the BPU cannot be built at the process node its power targets require. The Journey compiler's optimizations would no longer match whatever fallback chip could be produced. Every existing AEC-Q100 qualification would be voided, and the same lock-in that prevents OEM customers from leaving would also prevent the company from moving its own product to a different fabrication path.