On 17 September D-Robotics, known as Dagu Robot, closed a $400 million Series C. Mirae Asset led the round, with Meituan Strategic Investment and institutions including Hefei Guotou, Nanshan Venture Capital, Jingquan Capital and Cathay Capital following on. More than twenty existing shareholders, among them Hillhouse, 5Y Capital, Temasek’s Vertex Growth and Aramco’s Prosperity7, added further capital.
It is one of the largest single rounds in the robotics sector in four years. The twist is that Dagu builds no robot bodies. It positions itself as the picks-and-shovels supplier, selling only chips, an operating system and a developer toolchain, and aims to become the Wintel of the robotics era, even the Nvidia of embodied AI.
Three layers, one job: make building robots easy
The stack has three layers. The chip layer, the Sunrise family, spans 5 to 560 TOPS and fits everything from floor cleaners to humanoids. The software layer, the RDK developer kit, supports many algorithms and open repositories so builders do not start from scratch. The ecosystem layer, the Gravity accelerator, has served more than 500 small teams, 100,000 developers and 500 universities.
The flagship Sunrise S600, released in November 2025, carries 560 TOPS and the selling point of one chip for brain and cerebellum: an 18-core CPU for perception and decisions plus a 6-core microcontroller for real-time motor control. Older robot designs pair an AI chip with an external microcontroller, which adds cost and latency across the chip boundary. S600 does the lot on one die.
The bet is paying off in orders. Sunrise chips have shipped more than 8 million units in total, and first-half 2026 revenue grew several-fold year on year. Within half a year of launch, more than twenty leading customers, including UBTECH and PaXini, adopted S600, and Dagu says its share of embodied-AI customers passed 50 per cent, with most projects now at production scale.
Two years, four rounds, about $770 million
Dagu’s funding pace is unusual. It took a $100 million Series A in May 2025, a $120 million Series B1 in March 2026, a $150 million Series B2 in April, and now the $400 million Series C, roughly $770 million in two years and $670 million within the year, with only twenty days between B1 and B2.
The pattern signals a shift. As whole-robot funding cools, capital is moving upstream to chips and core parts that every builder needs. Dagu’s logic: do not bet on which robot wins, bet on what every robot must buy.
The company spun out of Horizon Robotics in early 2024. Chief executive Wang Cong led Horizon’s edge-AI and robotics unit from 2018. As Dagu kept raising, Horizon’s stake fell from near 70 per cent to about 40 per cent, and from 31 March 2026 Dagu left Horizon’s consolidated accounts, though Horizon remains the largest single shareholder and technology partner.
Taking on Nvidia where it hurts
Horizon founder and chief executive Yu Kai put it plainly at the 2026 Yabuli Forum: the rival in this field is still Nvidia. Nvidia’s Jetson line has been the default robot compute standard for a decade, and Yole puts its share of the global robot system-on-chip market near 69 per cent, and its CUDA ecosystem locks in millions of developers. Agility’s Digit and Boston Dynamics’ Atlas run on Jetson Thor.
Dagu wins no spec war. S600’s 560 TOPS sits between Nvidia’s older Orin at 275 TOPS and the latest Thor at 2070 FP4 sparse TFLOPS, and the two firms quote different precision, so the numbers do not line up. Dagu’s real edges are three. First, cost: Jetson prices keep rising, and on a thousand-dollar consumer robot the chip’s share of the bill of materials gets painful, and Dagu’s single-chip design drops the external microcontroller. Second, power: the Sunrise 5 family draws about 3 watts against Thor’s 40 to 130 watts, fitting home robots with tight power budgets. Third, supply security: with export controls uncertain, Chinese makers treat a domestic chip as both a cost and a risk choice.
The gap is ecosystem. Nvidia’s JetPack, Isaac, Omniverse and GR00T span simulation to deployment, and the CUDA lock-in is hard to break with specs or price. Wang Cong has said training and simulation may stay on Nvidia while deployment shifts to Dagu, so the near-term play is to hold the inference-and-deploy slot, not replace Nvidia. It is a smart and risky line: if Nvidia cheapens its own deploy end, Dagu’s middle ground shrinks.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-09/ART-898890-12003-30704773.html.