Let's look at three sets of numbers first.
First set: In the first half of 2026, Hesai Technology's core LiDAR business generated an operating profit of approximately CNY 108 million, while its new strategic businesses (spatial intelligence and robot actuators) lost about CNY 115 million—the profits earned from the core business were almost entirely consumed by the new businesses.
Second set: Hesai holds a 44% market share in China's long-range ADAS LiDAR market, which is twice that of the second-place player, ranking first for 17 consecutive months.
Third set: The Kosmo One, released on October 14, takes about 15 minutes to scan a 100-square-meter indoor space; whereas traditional manual modeling of the same area by a professional team takes about 300 hours.
On one side, there is the overwhelming advantage of the core business and the swallowed profits; on the other, an efficiency gap of over a thousand times. Putting these three sets of numbers together, the launch event on October 14 was no longer an ordinary new product release—it was a public rewriting of Hesai's own identity.
01 What Exactly Is Kosmo?
On October 14, Hesai Technology unveiled the 3D world data engine Kosmo in Shanghai, simultaneously launching two products: the world's first 8K ultra-HD AI spatial camera, Kosmo One, and the AI-driven cloud platform, Kosmo World.
Kosmo One is the data collection end: It integrates LiDAR, an 8K imaging system, a 1-inch large-format CMOS, and a high-precision IMU, performing microsecond-level synchronous collection of spatial geometry, color, texture, and pose. Coupled with 3DGS (3D Gaussian Splatting) reconstruction, it achieves millimeter-level geometric accuracy. Its most "anti-engineer" design is its ease of use—ready to scan right out of the box, with smart guidance and one-click continuous scanning, turning spatial collection, which previously required professional teams, into an experience akin to using a "point-and-shoot camera."
Kosmo World is the processing end: Its underlying layer is KSC, an AI-native spatial reconstruction and understanding system, featuring AIGC completion (automatically filling in unscanned areas and identifying materials) and AI 3D semantic segmentation capabilities. A single reconstruction simultaneously outputs three spatially aligned data formats: 3DGS, mesh, and depth—addressing three pain points in robot training: "fake visuals," "fake interactions," and "inaccurate spatial representation."
The ecosystem is the open end: It opens up raw collected data and 3D assets, integrating with mainstream simulation and creation platforms such as NVIDIA Isaac Sim, Isaac Lab, Unreal Engine, RenderMan, and Houdini.
Implementation cases are already up and running: Pixar's industrial-grade renderer RenderMan has integrated Kosmo-generated 3DGS data into its rendering pipeline and released official tutorials; the "Flying Street View" in collaboration with AutoNavi has been launched for the Chongqing Art Museum and Pifang Gallery; digitalization of the Du Fu Thatched Cottage in Chengdu is being carried out with TRIC, a UN partner; and Sharpa robots are using Kosmo-generated 3D environments for training in DQ Ice Cream stores, with official sources stating that failure cases have been reduced to one-tenth of the original, and task success rates have increased to 99%.
02 Why Must Hesai Do This?
First layer of reason: The ceiling of the automotive market is already in sight.
Li Yifan's statement is quite candid: "In the past, Hesai spent a long time reducing costs. The current level has reached a relatively reasonable steady state, and the performance basically meets the requirements. When there is no room for ultra-large-scale cost reduction and more than a 10-fold performance improvement, we indeed see that this capability can be used to make other products."
To translate this: LiDAR in vehicles is already approaching "good enough." When a technology shifts from an "arms race" to a "cost race," its ceiling is no longer determined by technology, but by the customers' procurement budgets. Hesai's gross margin dropped from 42.2% to 39.7% in the first half of the year, while RoboSense dropped from 25.9% to 21.8%—price wars are not news; they are the daily routine.
Shifting from a "must-have" to "better performance and more units" sounds like a good thing, but the essence is: automakers are buying hardware as consumables, and the destiny of consumables is continuous price reduction.
Second layer of reason: Capability reuse is almost a zero-cost migration.
LiDAR is the coupling of three capabilities: high-precision spatial perception hardware, chip-level in-house R&D (Hesai's 6D SPAD-SoC "Picasso" has completed mass production preparation), and large-scale precision manufacturing. Directing this set of capabilities to a new scenario incurs extremely low marginal costs—which is also why Li Yifan emphasized that "Kosmo and the handheld cameras from DJI and Insta360 are completely different categories." Its true technological foundation is still that spatial perception system.
Third layer of reason: New markets are enjoying premiums, while old markets are trapped in fierce competition.
Kosmo's two target markets are embodied intelligence and world models on one hand, and film, cultural tourism, and mapping on the other. Both are in a state of "abundant capital but scarce data": in the world model track, Li Fei-Fei's World Labs raised USD 1.23 billion with a valuation of about USD 5 billion, Yann LeCun's AMI Labs raised USD 1.03 billion in its seed round, and ByteDance Seed's budget for world model data is 3 to 4 times that of other directions. These funds will ultimately be spent on data, and Hesai wants to be the one selling the data.
03 The True Ambition: From Automotive Tier 1 to Data Infrastructure for Physical AI
Xiang Shaoqing presented an "impossible triangle" on-site: 2D images and videos lack depth and pose; pure simulation data cannot reproduce materials, friction, and deformation; real-world teleoperation is authentic but expensive, slow, and unscalable—high quality, large scale, and low cost cannot be achieved simultaneously in the long term.
This statement reveals the most real bottleneck in current physical AI: it is not a lack of computing power, nor a lack of models, but a lack of "authentic, scalable 3D world data."
And who holds "hardware capable of collecting data + production lines capable of mass-producing the hardware + ten years of accumulated real-world road data" in their hands? The list of answers is short. This is where Hesai's true scarcity lies—it is not crossing over to make cameras, but rather changing its ten years of accumulated collection capabilities from "selling to automakers" to "selling to the AI industry."
From "giving machines eyes" to "enabling robots to understand and change the world"—this is the story Li Yifan tells, but the business logic behind the story is very simple: since data is more expensive than sensors, don't just sell sensors.
04 Three Layers of Impact on the Industry
First layer: Redefining the ceiling of a "LiDAR company." In the past, the imagination of this industry was locked into "how many tens of thousands of units per year and how much value per vehicle." Kosmo elevates the narrative from "automotive parts" to "physical AI infrastructure"—this is a switch in valuation logic. At the same time, it also sets an example: the paths of the top three have diverged very clearly—Hesai uses money to buy time (cash reserves of approximately CNY 7.05 billion), RoboSense uses first-mover advantage to exchange for space (shipping 283,000 robot LiDAR units in the first half of the year, surpassing Hesai), and Tudatong uses focus to exchange for survival (cash of only about CNY 760 million, with an overall gross margin of 11.2% in the first half of the year).
Second layer: Supplying "data fuel" for embodied intelligence. 15 minutes vs. 300 hours is not an efficiency optimization, but a change in cost structure. When the collection threshold is lowered to the level of a "point-and-shoot camera," the supply curve of real-world 3D data may be rewritten. But note: the value of the data business obeys the law of scale—collecting data from 10,000 rooms and 1 million rooms are two completely different businesses.
Third layer: Monetizing 3DGS first in the content industry. Pixar's rendering pipeline, AutoNavi's Flying Street View (an entry point with nearly 1 billion monthly active users), and cultural tourism digitalization—these are cash flows closer than robotics. This also means Hesai's new business will not stand on just one leg.
05 Three Sharp Questions
Question one: What the new business consumes is precisely the profit of the core business. In the first half of the year, the core business generated an operating profit of CNY 108 million, while the new business incurred an operating loss of CNY 115 million. In other words, Hesai's current "profitability" is barely sustained by the core business. If the 2027 target for SGI to achieve about CNY 700 million in revenue and break even is delayed by a year, and the automotive price war continues to escalate—the squeeze from both ends will happen simultaneously. Cash reserves determine how long it can endure, but enduring itself does not produce victory.
Question two: The "consumer camera" might be the most dangerous step. Li Yifan has confirmed that Kosmo will subsequently be sold to C-end consumers. But the C-end is a completely different game: channels, branding, supply chain, after-sales service, and opponents like DJI and Insta360 who have been winning for a decade. Selling data to B-end is monetizing capabilities; selling cameras to C-end is gambling with capabilities. If this step goes awry, it will consume the most precious window of opportunity for the new business.
Question three: Open ecosystems and data assets are inherently contradictory. Hesai is opening up raw data and 3D assets, integrating with various simulation platforms, while at the same time relying on "scalable data production services" to make money. To what extent should it be open? When your clients include robot companies, mapping platforms, and film studios, and you are simultaneously providing integrated hardware and software products—once the boundary blurs, clients will start evaluating "whether they should build their own collection teams." This is a common long-term risk for all "infrastructure stories."
Conclusion: Data Is the Next Bottleneck
Ten years ago, Hesai turned LiDAR from a luxury item costing over a hundred thousand US dollars into a mass-produced automotive-grade component; ten years later, what it wants to do is turn the "real world" into reusable data assets.
What this company is truly betting on has never been a specific product, but a judgment: the next bottleneck for physical AI is data.
If this judgment holds true, Kosmo is a crucial foundation; if falsified, it will be an expensive imagination experiment conducted by a perception leader at the time when its core business profits were at their peak.
The results of both scenarios will be clear within three years.
What do you think? Will the next major client for LiDAR companies be automakers, robot companies, or AI labs? Let's discuss in the comments.