HangJiaShuo Robotics
Just now, D-Robotics announced the completion of its USD 400 million Series C funding round, equivalent to approximately CNY 2.7 billion.
South Korea's Mirae Asset led the round, with strategic investments from Meituan, Hefei State-owned Capital Operation Co., Ltd., and Nanshan Strategic Emerging Investment following, and over twenty existing shareholders including Hillhouse, 5Y Capital, Linear Capital, Vertex Growth, and Prosperity7 increasing their stakes. Combined with the previous Series A and Series B rounds, the cumulative financing of this company, spun off from Horizon Robotics, has now approached USD 800 million.
Several figures are worth noting: The cumulative shipments of the Sunrise series chips have exceeded 8 million units, revenue in the first half of 2026 has grown several times year-on-year, and the customer coverage rate in the embodied intelligence sector has surpassed 50%.
However, the Sunrise S600, D-Robotics's high-computing-power chip for embodied intelligence, was only released in November 2025, less than a year ago. The vast majority of the 8 million units could not possibly come from it.
What truly drove the numbers were the earlier orders: robot vacuums, robotic mowers, drones, and AI imaging devices. In 2025, D-Robotics's shipments grew by 180% year-on-year, the number of customers increased by 200%, and more than 100 products were commercialized.
01. Where the 560 TOPS Stands
The S600 delivers 560 TOPS at INT8 precision. On the same scale, NVIDIA's Jetson AGX Orin offers 275 TOPS, while the Thor launched in 2025 achieves 2070 TFLOPS at FP4 precision, with its developer kit priced at USD 3,499, equivalent to approximately CNY 25,000. The S600 sits right in the middle.
D-Robotics has chosen the "sufficient and scalable" segment: enabling 3B parameter-level embodied models to independently complete the closed loop of perception, decision-making, and control at the edge, at the cost of not competing for the highest performance label.
What better illustrates the point is not the TOPS, but the architecture. Inside a single S600, 18 A78AE cores run Linux for perception and decision-making, while 6 Cortex-R52+ cores run a real-time system for motor control; the two domains share data via an on-chip bus. 32 channels of PWM and 12 channels of CAN FD are routed directly from this chip, eliminating the need for an external MCU (Microcontroller Unit). This is a division of labor redefined for robotic applications. The savings it brings are not found in benchmark tables, but in one less component on the BOM, one less layer in the wiring harness, and the reduction of control command latency from milliseconds to microseconds.
02. The Most Difficult Part Is Not on the Chip
Wang Cong, CEO of D-Robotics, discussed a sequence at an industry forum this year: the simulation and training of robots will most likely be completed first on NVIDIA's platforms—Isaac, GR00T, Cosmos—before considering how to migrate the models to the actual deployment chip for mass production.
This statement lays bare the real situation of domestic substitution. What has always been lacking is not just computing power: the code developers have written in CUDA, the tools they learned in school, and the tuned simulation environments constitute the true migration costs. And such costs are not borne by chip companies, but absorbed by each robotics company themselves.
Therefore, the allocation of this round of funding is worth noting: half goes to the "full computing power segment layout of the Sunrise chips," and the other half to the "full-chain software platform integrating data collection, model training, simulation verification, and inference deployment." The latter is not as easy to market as 560 TOPS, but it can bring developers back from competitors' doorsteps.
03. Many Are Already on the Path
According to statistics from research firm Yole, in 2025, NVIDIA held a 69% share of the robotics SoC (System on Chip) market, while Chinese manufacturers combined accounted for only 5%. Leading the charge in domestic substitution is Rockchip—the basic versions of Unitree G1 and Agibot Lingxi X2 both use the RK3588 for their basic computing boards, which represents shipped volumes. Further up the ladder are Axera's 1500 TOPS "embodied brain controller," Moore Threads' "Yangtze River" series, and UBTECH's joint venture with MetaX, XILECT, which plans to tape out in the second half of 2027 and mass-produce in 2028.
There are also those pouring cold water on the trend. Song Jiqiang, President of Intel China Research Institute, judged earlier this year: dedicated robotics chips need to reach million-unit mass production to amortize R&D costs effectively, and the timing is not yet right.
A fork in the road that remains unanswered to this day: spend money to build the toolchain first, or wait for scale to come knocking. This USD 400 million is D-Robotics's answer.
04. After Coverage Surpasses Half
When the clients on the list are all competitors with each other—TARS, Spirit AI, X Square Robot, Zhi Zai Wujie, UBTECH, Paxini Sensing, plus the Beijing Humanoid Robot Innovation Center and Humanoid Robot (Shanghai) Co., Ltd.—the matter takes on another layer of meaning. TARS's wheeled dual-arm robot is fully equipped with the S600 and has entered mass production at the thousand-unit scale; Beijing Humanoid's "Embodied Tiangong 3.0" has completed computing power adaptation and is planned for delivery in the second half of the year.
After the same chip and the same set of toolchains are deployed across more than half of the industry peers, the foundation no longer constitutes differentiation for any single company, pushing the competition upwards: to models, data, and the understanding of specific workstations. The stronger the commonality at the hardware layer, the more valuable the individual differences at the software and data layers become.
The 8 million units represent cash flow already received, the over 20 clients are a work in progress, and that full-chain software platform remains a blueprint.
This article is compiled based on publicly available online information, for reference only, and does not constitute investment advice.