On November 17, Li Auto CEO Li Xiang said on Weibo that the company would gradually make its M100 chip available to external customers. The move was not driven by Li Auto's sales team. Customers approached the company themselves, with a number of robotics companies asking whether they could buy the chip.

The same day, Li Auto CTO Xie Yan delivered a keynote titled “Design Considerations and Practices for Efficient AI Inference Systems” at TSMC's 2026 China Open Innovation Platform Ecosystem Forum. Li Xiang said previous speakers at the forum had all come from the world's leading chip companies, and Li Auto was the first mainland Chinese company invited to speak.

A Weibo post and a keynote pointed to the same development: an automaker wants to sell automotive-grade chips to robot manufacturers.

Four Weibo Paragraphs, but No Customer List

Li Xiang's Weibo post summarized the M100's progress in four points. Four years ago, Li Auto concluded that conventional GPU architectures faced a data-movement bottleneck in edge computing, so it pursued a data-driven approach and developed what it calls the world's first mass-produced AI chip based on a dynamic dataflow architecture. A paper on the architecture was accepted by the ISCA Industry Track, making Li Auto the first automaker selected; the previous Chinese company accepted by ISCA was DeepSeek in 2025. Several new Li Auto models now come with the chip as standard, enabling parking systems to detect objects 10 meters earlier, forward perception at 300 meters, and a 50% improvement in dynamic-object recognition. The final point was that robotics companies had approached Li Auto about supply.

There was no fifth point. As CnEVPost noted, the post did not identify any prospective customers, provide a supply timetable, or disclose any signed purchase agreements.

That is not nitpicking. The hardest questions in selling chips externally are never simply whether a company is willing to sell, but who will buy, when deliveries will begin, and who bears responsibility if something goes wrong. Li Auto addressed none of the three.

It Eliminates the “Head Chef”

To understand why robotics companies are interested, it helps to see what makes this chip different.

At a launch event in June, Xie Yan used a kitchen analogy that he later repeated in a group interview. A conventional CPU architecture is like a kitchen with a head chef who directs the staff and assigns tasks. It works well with a small team, but as the kitchen grows, the head chef becomes a bottleneck: some workers sit idle awaiting instructions, while others stay busy but cannot get the resources they need. A GPU improves on this by replacing one head chef with many smaller kitchen teams, each led by its own chef. That increases parallelism, but the system remains instruction-driven at its core.

A dataflow architecture eliminates the head chef entirely. Workers do not wait for instructions. As soon as the “ingredients,” or data, reach their stations, they start working and then pass the result directly to the next person on the production line.

Behind the analogy is a technical judgment: AI workloads are inherently parallel, with deterministic data, clear relationships, and fixed paths of movement. If a model's design already specifies what must be computed next, there is no need for a central scheduler to direct operations in real time. That frees up transistors that would otherwise be used to manage the task queue.

The trade-off is embedded in another comment from Xie: dataflow requires compiling across “time and space.” In other words, a model must first be compiled into a graph that maps operations to specific compute units at specific times. The more regular the graph, the greater the efficiency; the more dynamic the workload, the harder it is to compile. Dataflow architectures inherently favor tasks with fixed graphs and are less suited to workloads with complex control flows or constantly changing shapes.

Robot perception and decision-making, however, happen to follow a relatively fixed graph.

82% Measures Computing Power Utilization, Not Usability

Li Auto's figures are a 5-nanometer automotive-grade process, 1,280 TOPS per chip, and computing power utilization above 82%.

That 82% figure stands out in the industry. For conventional GPUs running AI inference, computing power utilization of 30% is considered good, with much of the time spent waiting for data to move from video memory. A dataflow architecture removes much of that movement and scheduling overhead, so the higher utilization is logically consistent.

But it matters what the figure measures. Computing power utilization shows how much time the compute units spend working. It does not necessarily mean faster end-to-end performance, much less ease of use. The user experience is often determined by a different set of metrics: memory bandwidth, latency, and whether the model fits in memory.

This is particularly evident with language models. Large-model inference has two stages. Prefill is compute-intensive, while decode generates tokens one by one and is constrained by memory bandwidth rather than computing power. A chip with 82% computing power utilization will still decode slowly if its bandwidth cannot keep up. Xie said the chip “can run large language models and support agents,” but no third-party language-model benchmarks or token-generation speeds have been made public.

Li Auto has offered a different set of comparisons: three times the performance of NVIDIA's Thor-U on large-model workloads and a 40% reduction in end-to-end latency for intelligent driving. Both figures come from Li Auto's own tests and have not been independently reproduced.

A Company Yet to Begin Operations, Valued at 17.1 Billion Yuan

Any discussion of external chip sales must start with Li Auto's October 16 announcement.

That day, Li Auto announced that its chip subsidiary, Shanghai Mahexin Technology, had signed an agreement for its first equity financing round, raising 1.625 billion yuan. Li Auto itself invested 1 billion yuan, while outside investors contributed 625 million yuan, with an option to subscribe for up to another 500 million yuan. If the option is exercised in full, the company's post-money valuation will reach as much as 17.1 billion yuan. Li Auto retained a 74.01% stake after the financing and will continue to control and consolidate the subsidiary.

Mahexin was incorporated only on August 12. Li Auto stated in the announcement that the company had yet to begin substantive commercial operations. For now, its entire asset base consists of a mass-produced 5-nanometer chip already installed in vehicles and a team of about 200 people.

One chip, 200 people, and a valuation of 17.1 billion yuan. There is only one way that valuation works: the chip cannot be sold solely for use in Li Auto's own vehicles.

That makes the claim that “robotics companies are asking” both a signal of demand and part of the valuation narrative. Separating the two would distort the picture. Without external customers, the 17.1 billion yuan valuation is difficult to justify. With them, Li Auto can evolve from an automaker developing its own chips into a chip supplier. Li Xiang said the company would open supply “gradually,” leaving room to maneuver: crack open the door first, then see what orders materialize.

Selling Automotive-Grade Chips to Robot Makers: Performance Is Not the Problem

Assuming robotics companies do place orders, the first hurdle will not be computing power but liability.

Automotive-grade chips are certified under functional-safety standards, and the industry has established frameworks for determining the consequences of a chip failure in a vehicle and assigning liability. Robotics has no equivalent system. If a humanoid robot injures someone in a factory, there is no consensus on whether the robot manufacturer, chip supplier, or algorithm provider is responsible. Whether a chipmaker is willing to put its name into that chain of liability is a business decision, not a technical one.

The second hurdle is the supply commitment. The automotive industry expects long-term supply: a vehicle model can remain in service for more than a decade, and its chips cannot be discontinued midway through that lifecycle. Robotics volumes are still small. A robot manufacturer may produce only a few thousand units a year, translating into perhaps tens of thousands of chips. Committing to ten years of supply for that volume may not make economic sense.

The third hurdle is the compiler. Much of a dataflow architecture's value lies in its compiler: how the model is partitioned, which compute units operators are mapped to, and how operations are scheduled over time. This toolchain is a core asset that Li Auto has spent four years building. Customers cannot use the chip without the compiler, but providing it would mean handing over Li Auto's methodology. Public disclosures offer no answer on where the company will draw that line.

Li Auto has laid some groundwork. In Mahexin's financing announcement, the company said it planned to offer chips, domain controllers, and operating systems to external customers, while continuing to serve Li Auto vehicles and expanding into applications beyond the automotive sector. The Weibo post moved that plan one step closer to reality.

As for why robotics companies are now willing to consider alternatives to NVIDIA, there is a cautionary example close at hand. NVIDIA unveiled Thor in 2022 and initially planned to begin mass production by the end of 2024. After three delays, deliveries did not begin until August 2025, and the advertised computing power fell from 2,000 TOPS to 700 TOPS. Li Auto had planned to launch an updated L-series lineup in March 2025 with Thor as the centerpiece of the upgrade, but was forced to postpone it until May. Based on the difference in sales before and after the refresh, some media outlets estimated that the delay cost Li Auto about 6 billion yuan in revenue.

Waiting on a supplier's roadmap can carry a multibillion-yuan price tag. That lesson applies just as much to robotics companies as it does to automakers.