Pony.ai и Moore Threads: беспилотники переезжают на китайское железо

While Western giants debate autopilot safety, China has decided it's time to stop relying on external supplies of computing power. Pony.ai, one of the most prominent players in the autonomous taxi market, has announced a strategic partnership with Moore Threads. If you haven't been following the semiconductor market, Moore Threads is the team trying to create a full-fledged Chinese alternative to NVIDIA. Now they will work together to perfect the fourth level of autonomy (L4), where a vehicle handles the road without human intervention in virtually any conditions.

The essence of the deal is not simply purchasing graphics cards. The companies plan to build what they call a closed loop of 'algorithms—data—computation—applications'. Pony.ai is moving the training of its systems to MTT S5000 accelerators and KUAE intelligent computing clusters. This is a serious challenge for software: adapting algorithms that have been refined for years under CUDA architecture to new, albeit ambitious, Chinese hardware. But it's worth it, since it gives Pony.ai direct access to silicon-level optimization.

The announcement paid special attention to so-called 'world models'. In the autonomous driving world, this is currently the main trend. Instead of simply reacting to obstacles, a neural network learns to predict the physics of the surrounding world and the behavior of other traffic participants. Such models require colossal computational resources. Moore Threads claims that their clusters will handle this task no worse than Western analogues, while providing significant cost savings when scaling an autonomous fleet.

Why is this important right now? The autonomous driving industry has hit a cost ceiling. To roll out L4 to the streets en masse, it's necessary to drastically reduce the price of the car's 'brain' and the cost of training it. The shift to local solutions like MTT S5000 is not just a matter of political security and protection from sanctions, but also an attempt to make the technology economically viable. If Pony.ai manages to confirm the claimed specifications, we will see a mass exodus of Chinese tech companies from NVIDIA platforms to in-house solutions.

The Pony.ai and Moore Threads pairing also shows that the era of pure software companies in AI is coming to an end. The future belongs to those who control the entire technology stack. Pony.ai stopped being just a startup with good coders long ago; now it is a structure that influences chip development and server architecture. For Moore Threads, this is also a critical moment: if their chips withstand the load of real autonomous traffic in megacities, questions about the competitiveness of Chinese GPUs will resolve themselves.

The key point: Pony.ai is betting on import substitution at the heart of its technologies. Will Moore Threads be able to provide the same stability as NVIDIA, or will we see a delay in L4 development due to 'teething troubles' with the new hardware?