On September 26, 2025, Horizon Robotics raised HK$6.339 billion, at the cost of a HK$12.4 billion market capitalization loss that same day. This financing was a "pre-existing shares followed by new shares" placement, where the major shareholder sold existing shares and then subscribed to new shares, allowing the company to quickly obtain new funds—60% for R&D investment, 20% for the industry chain investment, and the remainder for overseas expansion and investment in Robotaxi. As a company included in the Hong Kong Stock Connect and whose market capitalization exceeded HK$100 billion in less than a year after listing, Horizon Robotics chose to raise funds at a discount at this time, and this was its third financing since listing (totaling HK$17.19 billion). Another company, Mobileye, also started with automotive-grade AI vision chips and also faces the reshuffling brought about by the rapid iteration of autonomous driving algorithms. On its first day of listing on Nasdaq in October 2022, its market capitalization surged to US$23 billion. However, Horizon Robotics did not choose to raise funds at a discount, and its market value plummeted. By the time it raised funds for the third time, it only had $13.1 billion left, and today it's only $7.3 billion. Horizon Robotics clearly doesn't want to become the next Mobileye. Rather than waiting for the market to drive down its valuation, it's better to proactively offer a discount and put cash in its pocket first, even if the cost is evaporating twice the market value of the cash it receives. At least it buys the company time to survive tomorrow. I. Every Valuation Requires a New Story Since 2015, Horizon Robotics has consistently made choices when there were no other options, but at least when there were still choices. In 2012, Yu Kai returned from Silicon Valley and joined Baidu, becoming the first director of Baidu IDL (Institute of Deep Learning) the following year. This institution focuses on cutting-edge fields such as deep learning, image recognition, and speech recognition. In 2015, Yu Kai chose to leave and found Horizon Robotics, aiming to become the "Intel of the robotics era." He later said in an interview that he had judged the direction of edge intelligence. The reality was that Baidu was all-in on search advertising and O2O at the time, and the chip business was not a high priority. It could only conduct research in the research institute and couldn't wait for mass production orders. Leaving was Yu Kai's only way out at the time. In 2018, Horizon Robotics released its future city solution, covering ADAS (Advanced Driver Assistance Systems), AIoT (Smart Home), smart retail, smart cities, and other fields, practically cramming chips into every imaginable smart hardware. This year was a banner year for AI companies, with the "Four Little Dragons" raising over $3 billion in total, accounting for about one-fifth of the total funding in the entire AI field. They focused on computer vision, valuing their investments based on software, while Horizon Robotics focused on "chips + algorithms," primarily valuing them based on hardware. The two valuation systems differed by an order of magnitude. If Horizon Robotics wanted a high valuation in its Series B round, it had to prove that it was making more than just automotive-grade chips; it was developing general-purpose AI chips applicable to any scenario, including cities, retail, homes, factories, and transportation. Only then could it weave a narrative of a trillion-dollar "general-purpose AI chip." In February 2019, Horizon Robotics secured $600 million in its Series B round, valuing the company at $3 billion. The general-purpose AI chip narrative materialized, but it couldn't continue. Engineering redundancy, slow payments, and fragmented orders revealed the costs of its multi-pronged approach. Ultimately, it had to lay off a large number of employees and retreat to automotive-grade chips to stem the bleeding, since the combined size of its numerous smaller businesses didn't even compare to the scale of the automotive vertical market. Yu Kai later explained, "Instead of digging a bunch of small holes, it's better to focus all your efforts on digging the deepest one." If Horizon Robotics didn't scale back its operations, its cash reserves would struggle to reach Series C. From 2020 to 2021, Horizon Robotics was deeply integrated with Li Auto ONE and Changan UNI-T, providing on-site development of integrated hardware and software. This year and a half was a rare window of opportunity in the automotive chip field. Nvidia was too expensive, Mobileye operated on a black-box model (chips and algorithms were bundled together, limiting the differentiation space for automakers), and Qualcomm and Huawei hadn't yet achieved mass production. Chinese automakers were almost entirely limited to Horizon Robotics' Journey 2 and Journey 3. In December 2020, Series C1 was launched. By June, Series C7 had concluded, bringing the total Series C funding to $1.5 billion, with a valuation of $5 billion. This time, the valuation was no longer supported by the narrative of "general-purpose AI chips," but by hard evidence of automotive-grade mass production. In 2022, Horizon Robotics brought in Volkswagen to participate in its Series D funding round. A $1 billion investment, plus a €1.3 billion joint venture, Coretronic (60% owned by Volkswagen), brought the global automaker's participation, locking in a Series D valuation of $8.71 billion. In 2024, Horizon Robotics went public in Hong Kong, priced at HK$3.99, with a market capitalization of HK$53.4 billion on its first day of trading, lower than its Series D valuation—a discounted listing. However, Horizon Robotics had no other choice at the time. The A-share STAR Market was restrictive for unprofitable chip companies, the US stock market was closed to Chinese concept stocks, and RMB funds were scarce. Hong Kong was the only market that could accommodate this scale, even at a discount. In 2025, Horizon Robotics raised a total of HK$17.19 billion through three rounds of placements, and simultaneously began telling its story about robotics and embodied intelligence. Because the valuation ceiling for Horizon Robotics' intelligent driving business is set at HK$50-80 billion, while its market capitalization has already exceeded HK$100 billion, the gap must be filled by a larger narrative, just as in 2018, when "general-purpose AI chips" were used to justify a valuation exceeding fundamentals. This new narrative was launched at the product launch in April 2026. Yu Kai packaged "automotive intelligent agents" as the next fulcrum and proposed a new concept of "hardware depreciating over time, while agents appreciating over time," attempting to shift Horizon Robotics' valuation anchor from chip prices to the long-term subscription value of software intelligent agents. It can be said that every expansion and contraction of Horizon Robotics was not a choice between optimal and suboptimal solutions, but rather a choice with no other option. All the forced choices over eleven years have combined to create the Horizon Robotics we see today. Attributing it solely to foresight obscures the complex evolution between business, technology, and capital. II. From Solution Provider to Chip Provider Horizon Robotics' gross profit margin of 64.5%, far exceeding that of pure chip manufacturers, is because it doesn't simply sell chips, but rather collaborates with core customers on development. Horizon Robotics sends engineers to the site to package chips, toolchains (Tiangong Kaiwu), and some perception and control algorithm IPs into a unified hardware and software solution, upon which automakers then develop application-level solutions. However, this presents a paradox: the more important the intelligent driving business, the more motivated automakers are to develop their own solutions, and the more they want to break free from their dependence on joint development with Horizon Robotics. When the honeymoon period of joint development between Horizon Robotics and automakers ends largely depends on the maturity speed of the automakers' in-house R&D teams. The first thing automakers will relinquish is the control algorithm. This part relates to differentiated user experience, and the marginal cost of self-developed chips is rapidly decreasing with a sufficient talent market and the open-source nature of toolchain ecosystems (PyTorch, CUDA). Secondly, the perception algorithm is being withdrawn. Weixiaoli's intelligent driving team has thousands of members, and even when using Horizon Robotics chips, they only treat them as a computing power base; the algorithm layer is entirely self-developed. This window of opportunity may only last three to five years. When the control and perception algorithms are withdrawn, even if chips are still being purchased, it will be downgraded from a "hardware-software integrated solution" to a "computing power base," like a car-grade version of MediaTek, focusing solely on chip development. However, mid-to-low-end models will still use externally purchased chips in the medium to long term because, for automakers, the real hurdle to self-developed chips is not the design itself, but rather the cost of tape-out, yield ramp-up, automotive-grade verification, TSMC capacity allocation, and most importantly, the maturity cycle of the software stack. This is precisely the moat that Horizon Robotics' J6E and J6M currently find difficult to shake (Nvidia Thor's cost makes it unfriendly to models priced below 150,000 RMB). This also means that if Horizon Robotics were to become a pure chip company, it could maintain its revenue, but the low gross margin from high-volume sales would not sustain its valuation. Referring to the current PS ratio range of pure chip companies (such as MCU and SoC manufacturers) in the A-share and Hong Kong stock markets, Horizon Robotics' market capitalization may fall back to the 5-7 times PS level. Horizon Robotics can only prepare for this in two ways. For customers with a strong desire for self-developed technology, it can only retreat to "selling chips + toolchains". For customers with a weak desire for self-developed technology, lack of capability, or low production volume, it will provide a full-stack solution from chips all the way to the intelligent driving system. In 2024, it launched the advanced driver assistance system HSD (Horizon SuperDrive), directly competing with Huawei ADS and XPeng XNGP. While Horizon Robotics' current adoption rate and scenario coverage still lag significantly behind Huawei's ADS, the intelligent driving version option has already reached 77% in models like the Exeed ET5 and iCAR V27 equipped with HSD. Horizon Robotics' official statement is that it aims to achieve mass production of tens of millions of units within the next three to five years. Horizon Robotics is also adopting a tiered approach to its Journey 6 products—the high-end J6P is reserved for joint development, while the mid-range J6E and entry-level J6M are supplied in a standardized manner. Furthermore, its toolchain, Tiangong Kaiwu, actively supports CUDA-style development interfaces to reduce migration friction for automakers switching from NVIDIA to Horizon Robotics. Horizon Robotics' current various responses are merely delaying the inevitable, slowing the collapse into a landslide, in exchange for a window of opportunity for chip procurement.
III. From Dedicated Chips to General-Purpose Chips
The time gained through business tactics is being eaten up by the generational debt at the technological level.
Horizon Robotics' BPU chip follows the DSA (Domain-Specific Architecture) route, optimizing the dataflow layer for tensor computation, convolution, and memory access patterns of neural networks in exchange for several times higher energy efficiency than general-purpose GPUs. This path was correct in the ADAS era of 2017-2020—when convolutional neural networks were dominant, and the algorithms running on the vehicle were relatively stable.
However, since 2020, the evolution of vehicle-side algorithms has taken two intertwined paths. One is the network structure, shifting from CNN-dominated to a CNN+Transformer hybrid, and then to BEV perception dominated by Transformer.
The other path is the system paradigm, shifting from modular perception-decision-control pipelines to end-to-end models, and then evolving towards VLA (Vision-Language-Action). These two paths overlap, creating architectural pressure every three to four years. The BPU architecture faces engineering trade-offs with each paradigm shift: should it harden another layer of operators to maintain energy efficiency, or reserve more general-purpose computing units within the chip to handle unknown algorithms? The former offers higher energy efficiency in the short term, but requires starting over with each next migration. The latter yields to GPUs; every unit of area freed up sacrifices the energy efficiency advantage that originally belonged to dedicated chips. Horizon Robotics chose the latter, and continues to pursue it. The latest version of the chip, the Starry 6P, utilizes a 5nm automotive-grade process with expanded bandwidth and on-chip storage. It boasts the ability to simultaneously support large-scale intelligent driving models and large-scale cockpit-side models, with specifications comparable to mid-to-low-end NVIDIA Thor chips. While competitive in mainstream models priced between 150,000 and 250,000 yuan, its energy efficiency is narrowing compared to the previous generation, forcing it to compromise on general-purpose computing to accommodate large models. It's a cost victory, not an architectural one, and merely delays the pressure of algorithmic paradigm shifts. Additionally, Horizon Robotics has introduced a "magazine system," essentially making the intelligent driving computing unit a relatively independent, replaceable module, allowing automakers to upgrade computing power without recertifying the entire vehicle's EE architecture. This reduces hardware depreciation pressure, but only extends the service life of a single-generation chip, rather than addressing generational debt at the architectural level. Currently, the computing power requirements for mass-produced vehicles have increased from tens of TOPS in the early days to the 500 TOPS level. In the future, with the arrival of VLA (Virtual Reality), perception, decision-making, and control will be merged into a single large model. The industry's optimistic estimate for computing power requirements is around 500 TOPS with sparsity and low bit quantization, while a pessimistic estimate is over 1000 TOPS. Regardless of the figure, the demands on dynamic computation graphs, long context inference, and KV cache management far exceed current ADAS workloads—precisely NVIDIA GPUs' traditional strengths. NVIDIA's Orin to Thor architecture also spans two generations from Ampere to Blackwell, but its advantage lies in the continuity of the CUDA ecosystem—upper-level algorithm code can run on new hardware, performance tuning experience can be reused, and developer migration costs are low. Horizon Robotics' BPU architecture adjustments with each generation require simultaneous rewriting of the toolchain and re-porting of customer algorithms; this is an intergenerational debt at the ecosystem level, which cannot be compensated for by the performance of a single chip. This is why Horizon Robotics maintains increasingly high investment in R&D, far exceeding the estimates made by Yu Kai when he invested in 2018. Even so, Horizon Robotics still faces a seemingly distant "gray rhino." In the end-to-end and VLA paradigms, data, models, and computing power form a closed loop of training and deployment. Automakers, having control of real-world road data, will have bargaining power that extends to chip selection. Leading automakers have already begun negotiating customized solutions with TSMC and Nvidia, bypassing general-purpose automotive chip suppliers. Even if Horizon Robotics keeps pace with chip generations, the industry's focus continues to shift towards automakers who control the data closed loop. Moreover, as VLA and subsequent large-scale model paradigms continue to iterate, driving up demands for computing power and flexibility, the BPU architecture patches will eventually reach a critical point of diminishing returns. At that point, Horizon Robotics may have to redesign its architecture, meaning the technological accumulation of the past eleven years will be rendered meaningless.
IV. Cashing Out Valuation for Time
For Horizon Robotics, although the various long-term and short-term pressures mentioned above are resisted layer by layer in different directions of business and technology, the capital for resistance comes from capital. It is necessary to cash out unsustainable valuations in a timely manner to gain a buffer time in terms of business and technology.
In the Hong Kong stock market, Horizon Robotics' current valuation of hundreds of billions is based on the narrative of "a company with both chip capabilities and software algorithm capabilities." Mobileye, whose fundamentals are very similar, has a market capitalization of only $7.3 billion today, with a PS ratio of 3.5. The difference between the two is essentially the premium the market paid for "Horizon Robotics will be more than just Mobileye in the future." Once expectations reverse, the decline will be precipitous. Mobileye's 70% drop in three years means the loss of all narrative premiums.
Therefore, Horizon Robotics faces an extremely complex actuarial problem: it must consider the decline in valuations in the secondary market while maintaining its share placement financing, but the share placement financing will inevitably lead to a further decline in valuation. Horizon Robotics must time its moves precisely to successfully convert its valuation into time. Moreover, successful cash-out is only temporary; ultimately, it depends on whether it can be transformed into a sustainable cash flow business. The closest testing milestone as of press time is the Q3 mass production of the Starry Sky 6P. If it successfully implements the "cabin-driver integration + vehicle intelligent body" narrative in its debut with Chery iCAR, Horizon Robotics can secure a new valuation anchor for the next round of share placement. If Q3 mass production is delayed or the post-launch reputation falls short of expectations, the realization of the narrative will be postponed. Overseas markets present a different scenario. For domestic brands, Horizon Robotics is increasingly inclined to directly enter the supply chains of automakers. However, for joint venture brands and overseas markets, Tier 1 suppliers like Continental, Bosch, and ZF are unavoidable intermediaries. Horizon Robotics can only enter overseas markets through Tier 1 suppliers or joint ventures like Coretronic. Coretronic is a joint venture where Volkswagen holds a 60% stake and Horizon Robotics provides the technology, developing intelligent driving solutions for Volkswagen's next-generation electronic and electrical architecture. If successful, Horizon Robotics will have a replicable overseas model. If unsuccessful, its overseas story will be limited to exporting its Journey series chips. Currently, its mass production timeline has been delayed from 2026 to the end of 2027 or even 2028. Zooming out, we find that this timeline is not unique to Horizon Robotics. The A-share market's refusal to accept unprofitable red-chip companies, the closure of US stock exchanges to Chinese companies, the shortage of RMB funds, the blockade of WVR for strategic acquisitions, the shift in algorithm paradigms every three to four years, and automakers' self-developed technologies forcing suppliers to give way—these constraints are not unique to Horizon Robotics, but rather a common predicament for this generation of Chinese hard-tech companies. Horizon Robotics was only exposed to the public market first because of its sheer size. Behind it is a long list of companies that haven't reached that stage yet, but will eventually face the same timeline.
Anais
Alex
Miyuki
Catherine
Weiliang
Kikyo