Horizon: Eleven Years with No Other Choice

On September 26, 2025, Horizon Robotics raised HKD 6.339 billion, at the cost of a market value evaporation of HKD 12.4 billion on that day.
This financing was a “sell old shares first, then issue new shares” placement—major shareholders first sold their existing shares, and then subscribed for new shares—so the company could quickly secure new funds: 60% for R&D, 20% for the industrial chain, and the remainder for overseas expansion and investments in Robotaxi. As a company included in Hong Kong Stock Connect and valued at more than HKD 100 billion in less than a year after listing, Horizon, however, chose to conduct a discounted financing at this time—and it was its third financing since going public (cumulative HKD 17.19 billion).

Another company, Mobileye, also started with automotive-grade AI vision chips and similarly had to deal with the reshuffling caused by how fast autonomous driving algorithms iterate along their roadmap. When it first listed on NASDAQ in October 2022, its market value surged to USD 23 billion. But later it did not choose discounted financing. Its market value kept falling—so that by the time Horizon was doing its third fundraise, Mobileye’s market cap had already fallen to USD 13.1 billion, and today it is USD 7.3 billion.

Horizon clearly does not want to become the next Mobileye. Instead of waiting for the market to drive valuations down, it decided to discount itself proactively, put cash into its pocket first—even if that means the market value evaporates to an amount twice the cash it takes in—so long as it buys time for the company to survive tomorrow.

1. Every valuation needs a new story

Tracing back to 2015, Horizon has been making the choice when it had no alternative—at least when there was still a choice to make.

In 2012, Yu Kai returned from Silicon Valley, joined Baidu, and the following year became the first director of Baidu IDL (Institute of Deep Learning, Institute of Deep Learning), an organization focused on cutting-edge fields such as deep learning, image recognition, and speech recognition.
In 2015, Yu Kai chose to leave and founded Horizon, with the goal of becoming the “Intel of the robot era.” Later, in interviews, he said he had judged the direction of edge intelligence.

The real situation was that at the time, Baidu was all-in on search advertising and O2O, so the chip business was not a priority. The company could only do research within the institute and could not wait for mass production orders. Leaving was Yu Kai’s only way out at that moment.
In 2018, Horizon released future city solutions covering ADAS (Advanced Driver Assistance Systems), AIoT (smart home), smart retail, and smart cities—almost stuffing chips into every smart hardware category you could think of.
That year was a high point for AI companies: the “Four Little Dragons” collectively raised more than USD 3 billion, accounting for about one-fifth of total AI funding. What they were telling was a computer vision narrative, valued mainly on software, while Horizon’s story was “chips + algorithms,” valued primarily as hardware—so the two valuation systems differed by an order of magnitude.
For Horizon to have a high valuation at Series B, it had to prove that it was not only making automotive-grade chips, but also making general-purpose AI chips applicable to any scenario such as cities, retail, homes, factories, and transportation. Only then could it weave the narrative of “general-purpose AI chips” at a trillion-scale.
In February 2019, Horizon raised USD 600 million at Series B with a valuation of USD 3 billion. The general-purpose AI chip narrative was realized, but it could not be sustained. Engineering redundancy, slow collections, and dispersed orders—these costs of spreading across multiple lines began to show. In the end, Horizon could only dramatically cut headcount, retreat to automotive-grade chips to stop the bleeding, because the sum of a pile of small businesses still could not match the scale of a single vertical in automotive.

Yu Kai later explained, “With that many small pits to dig, it’s better to devote everything to digging one deepest pit.” If Horizon did not shrink its front line again, it would be hard for cash to last until Series C.
From 2020 to 2021, Horizon deeply bound with Li Auto ONE and Changan UNI-T, stationed on site to do integrated hardware-software development. This year and more was a rare window in the in-vehicle chip space. NVIDIA was too expensive. Mobileye operated as a black-box model (selling chips and algorithms bundled together, leaving limited room for differentiation for automakers). Qualcomm and Huawei had not yet moved into mass production either. As a result, Chinese automakers could almost only choose Horizon’s Journey 2 and Journey 3.
In December 2020, the C1 round kicked off. By June, the C7 round was wrapped up. Totaling USD 1.5 billion with a valuation of USD 5 billion. This time, the valuation was no longer supported by the “general-purpose AI chip” narrative, but by hard evidence of mass production in automotive.
In 2022, Horizon introduced Volkswagen to participate in the D round. With USD 1 billion in investment plus EUR 1.3 billion to set up the joint venture CoreRun (Volkswagen holding 60%), the entry of global automakers helped lock the Series D valuation at USD 8.71 billion.
In 2024, Horizon’s Hong Kong stock IPO set the price at HKD 3.99, and its closing market cap on the first day was HKD 53.4 billion—below the D-round valuation, i.e., a discounted listing. At that time, Horizon had no choice: on the A-share STAR Market, rules were tightened for unprofitable chip companies; the U.S. market had shut the door to Chinese concept stocks; and RMB funds were short of supply. Hong Kong was the only market that could accommodate this scale—even if it was already at a discount.
In 2025, Horizon’s three placements totaled HKD 17.19 billion, and it also began telling stories about robots and embodied intelligence. Because the valuation ceiling for the autonomous driving business was capped in the HKD 50-80 billion range, and Horizon’s market cap had already stood above HKD 100 billion, the gap had to be filled by a larger narrative—just like in 2018, when “general-purpose AI chips” carried valuations beyond fundamentals.
This new narrative took shape at a product launch in April 2026. Yu Kai packaged “car intelligence agents” as the next fulcrum and proposed a new concept: “hardware depreciates over time, agents appreciate over time.” He attempted to shift Horizon’s valuation anchor from chip prices to the long-term subscription value of software agents.
It can be said that every expansion and contraction at Horizon is not a choice between an optimal solution and a suboptimal one; rather, there is no alternative. All the forced choices accumulated over eleven years combined to become today’s Horizon. If you reduce it to vision alone, you would obscure the complex evolution among business, technology, and capital.

2. Sliding from a solution provider into a chip supplier

Horizon’s gross margin is able to stay at 64.5%, far higher than pure-play chip manufacturers, because it is not merely selling chips. Instead, it co-develops with its core customers—Horizon sends engineers to work on site, bundling chips, toolchains (Tian Gong Kai Wu), parts of perception and planning/control algorithm IP into a unified hardware-software solution, on top of which automakers develop application-layer software.

But a paradox arises here: the more important the autonomous driving business becomes, the stronger automakers’ incentives to develop in-house, and the more they need to break free from dependence on joint development with Horizon. When does the honeymoon period of joint development between Horizon and automakers end? It mostly depends on how quickly automakers’ in-house teams mature.
The first thing automakers recover are the planning/control algorithms. This portion is tied to differentiated user experience, and the marginal cost of developing these in-house drops quickly as talent markets are abundant and toolchain ecosystems (PyTorch, CUDA) become open source.

Next come perception algorithms. The autonomous driving teams at companies such as WeRide, XPeng, and Li Auto each have teams on the scale of thousands. Even if they use Horizon chips, they treat the chips only as a computing base, while fully self-developing algorithms at the software level. This window may last only three to five years.

When planning/control algorithms and then perception algorithms are both pulled back in-house, even if chips are still being purchased, the unified “hardware-software solution” downgrades to a “computing base”—like a car-grade MediaTek: only making chips.

However, in the medium to long term, mid- to low-end models will still use externally sourced chips. For automakers, the real threshold for self-designed chips is not the design itself, but wafer fabrication cost, yield ramp-up, automotive-grade validation, foundry capacity allocation at TSMC, and—most critically—the software stack maturity cycle. This is the moat that Horizon J6E and J6M currently have that is difficult to dislodge (NVIDIA Thor’s cost structure makes it unfriendly to models below 150,000 RMB).

This also means that if Horizon slides into being a pure chip company, it can protect revenue—but it cannot protect valuation with low gross margins at scale. Referring to the current PS ranges of pure chip companies in A-shares and Hong Kong (such as MCU and SoC vendors), Horizon’s market cap could fall back to a 5-7x PS level.

Horizon can prepare two options. For customers with strong in-house development intent, it can revert to “selling chips + toolchains.” For customers with weak in-house intent, or insufficient capability, or with low production volumes, it provides a full-stack solution packaged end-to-end, from chips all the way to autonomous driving systems.

In 2024, Horizon launched a complete high-end assisted driving solution, HSD (Horizon SuperDrive), directly targeting Huawei ADS and XPeng XNGP. Compared with Huawei ADS, Horizon’s current takeover rate and scenario coverage still show a clear gap. However, for models equipped with HSD such as Starry ET5 and iCAR V27, the optional upgrade rate for the autonomous driving package has reached 77%. Horizon’s stated target for the outside world is to achieve “ten million-scale” mass production in the next three to five years.
Horizon also layers its product strategy on Journey 6—J6P high-end is reserved for joint development; J6E mid-range and J6M entry-level are standardized supplies. In addition, the toolchain Tian Gong Kai Wu proactively supports CUDA-style development interfaces to reduce migration friction for automakers moving from NVIDIA to Horizon.
At present, Horizon’s various responses only delay collapse—turning a crash into a slow slide—to secure a window for chip procurement.

3. Retreating from dedicated chips back toward general-purpose chips

The time bought through business tactics is being eaten away by generational technical debt.
Horizon’s BPU chips follow a DSA (Domain-Specific Architecture) route. They optimize at the dataflow layer for tensor calculations, convolution, and memory access patterns for neural networks, achieving energy efficiency several times higher than general-purpose GPUs. In the ADAS era from 2017 to 2020, this path was correct—back then convolutional neural networks dominated, and the algorithms running on the vehicle side were relatively stable.

But starting in 2020, the evolution of vehicle-side algorithms has two intertwined lines. One is network structure: shifting from CNN dominance to a CNN+Transformer hybrid, and then to BEV perception dominated by Transformer. The other is system paradigm: moving from modular perception-decision-control pipelines to end-to-end models, and then evolving into VLA (Vision-Language-Action, vision-language-action large models). With these two lines overlapping, architecture pressure arrives about every three to four years.

For the BPU architecture, each paradigm migration requires engineering trade-offs: whether to harden an additional layer of new operators to preserve energy efficiency, or to leave more general compute units in the chip to cover unknown algorithms. The former yields higher energy efficiency in the short term, but the next migration has to be redone again. The latter yields to GPUs: every slice of area set aside for general compute steals away a portion of the energy-efficiency advantage originally unique to dedicated chips.

Horizon chose the latter—and it has gone further and further. The latest chip is Starry 6P, built on a 5nm automotive process. It expands bandwidth and on-chip storage, and it claims that a single chip can simultaneously support both autonomous driving large models and cockpit-side edge large models, with core parameters positioned against the mid-to-lower configurations of NVIDIA Thor. It is competitive in mainstream models priced in the 150,000-250,000 RMB range. But compared with the previous generation, its energy efficiency has narrowed, because to accommodate large models it has to compromise and give way to more general-purpose computing. It is a victory in cost—not a victory in architecture—and it still only delays the pressure of algorithm paradigm migration.
In addition, Horizon has introduced a “cartridge system.” Essentially, it makes the autonomous driving compute unit a relatively independent, swappable module. That allows automakers to upgrade compute without re-certifying the vehicle’s entire EE architecture. This reduces hardware depreciation pressure, but it only extends the service life of a single generation of chips—it does not address generational debt at the architectural level.
At present, compute demand for mass-production vehicles has grown from the early dozens of TOPS to the scale of around 500 TOPS. When the VLA wave arrives—when perception, decision-making, and control are merged into one large model—compute needs are estimated at roughly 500 TOPS (optimistic) on the low end and 1000 TOPS or more (pessimistic) on the high end, and can be made workable by sparsification and low-bit quantization. But regardless of which number, the requirements for dynamic computation graphs, long-context reasoning, and KV cache management are far beyond today’s ADAS workloads—exactly the areas where NVIDIA GPUs traditionally excel.

NVIDIA’s transition from Orin to Thor also spans two architectural generations—Ampere to Blackwell. But NVIDIA’s advantage lies in the continuity of the CUDA ecosystem: algorithm code can run on new hardware, performance tuning experience can be reused, and developer migration costs are low. Horizon’s BPU architecture, with each generational adjustment, requires synchronized toolchain rewrites and customer algorithm re-porting. That is a form of generational debt at the ecosystem level; it cannot be made up by the performance of a single chip. This is why Horizon’s R&D investment continues to stay at an increasingly high level—at least far higher than what was estimated when Yu Kai bet on the company in 2018.

Even so, Horizon still faces a “gray rhino” that looks distant. Under the end-to-end and VLA paradigms, data, models, and compute form a closed loop from training to deployment. Since automakers hold the real-world driving data, their bargaining power extends in the opposite direction to chip selection. Some leading automakers have already started negotiating customized solutions with TSMC and NVIDIA, bypassing general-purpose automotive-grade chip suppliers. Even if Horizon keeps up with chip generations, the industry’s center of gravity will continue shifting toward automakers that control the data loop.
Moreover, as VLA and subsequent large-model paradigms keep iterating, and as demand for higher compute and more flexibility rises, the patches to the BPU architecture will eventually reach a point where marginal benefits level off. At that point, Horizon may need to redo its architecture—meaning the technological accumulation from the past eleven years could effectively reset to zero.

4. Cashing valuation into time

For Horizon, these long-term and short-term pressures—though they resist in different directions—ultimately all come from capital. It has to seize the chance to monetize an unsustainable valuation in a timely manner, so as to gain buffers in business and technology.

In the Hong Kong market, Horizon’s current valuation above HKD 100 billion is built on the narrative of a “company that has both chip capabilities and software algorithm capabilities.” A company with fundamentals similar to Horizon, Mobileye, has today’s market value reduced to only USD 7.3 billion—about 3.5x PS. The difference essentially reflects the market premium that “Horizon will not be just Mobileye in the future.” Once this expectation reverses, the downside could be abrupt, like a cliff. Mobileye’s 70% drop over three years is a drop of all its narrative premium.

So Horizon faces an extremely complex actuarial problem: it must worry about valuation declines in the secondary market, while at the same time it must maintain placement financing—but placement financing itself inevitably brings further downward pressure on valuation. Horizon has to hit the right timing to successfully convert valuation into time.

And even if the cash-out succeeds, it is only temporary. Ultimately, it still depends on whether it can be converted into a sustainable cash-flow business.

The closest test node as of the time of publication is the Q3 mass production of Starry 6P. If it can run the “cockpit-integration + vehicle intelligence” narrative in the debut of Chery iCAR, Horizon can secure a new valuation anchor for the next round of placement. If Q3 mass production is delayed, or if post-launch reputation after being installed in vehicles falls short of expectations, then the realization of the narrative will have to be pushed out.

Overseas is another curve. For domestic brands, Horizon is increasingly inclined to enter the supply chain of carmakers directly. But for joint-venture brands and overseas markets, Tier 1 suppliers such as Continental, Bosch, and ZF are the unavoidable middle layer. Horizon’s overseas business must go through Tier 1 partners or through joint ventures such as CoreRun to enter.
CoreRun is a joint venture with Volkswagen holding 60%, with Horizon providing technology. It develops autonomous driving solutions for Volkswagen’s next-generation E/E architecture. If it runs through successfully, Horizon will gain a replicable overseas model. If it does not, the overseas story will shrink to only a narrow path—exporting Journey series chips. As of now, its mass production timeline has already been pushed from 2026 to the end of 2027, and even to 2028.
Zooming out reveals that this timetable is not unique to Horizon. In A-shares, the market does not accept unprofitable red-chip listings; in the U.S., the door for Chinese concept stocks is shut; RMB funds are in short supply; WVR locks out strategic mergers and acquisitions; algorithm paradigms switch every three to four years; and automakers’ in-house development push forces suppliers to give way. These constraints are not Horizon’s alone—they are the shared predicament of this generation of China’s hard-tech companies.

Horizon is only the first to be exposed to the public market because its scale is large enough. Behind it is a long list of companies that have not reached this stage yet—but will inevitably face the same timetable sooner or later.