Foreword:

Life originated in the ocean. As AI extends its reach from land to the deep sea, those "dirty, difficult, and dangerous" tasks that humans cannot, do not do well, or cannot afford to do are brewing an efficiency revolution.

The large-scale commercial deployment of general-purpose embodied intelligence will also find its answer first in the deep blue waters.

Author | Fang Wensan

Image Source | Internet

Behind the $1 Billion A-Round, Capital is Betting on the Certainty of Underwater Operations

Embodied intelligence is extending from land to sea. Recently, Shihang Intelligence completed an A-round financing of over $1 billion in the marine robot sector, setting the highest single-round financing record globally in this field.

New investors include Shanghe Momentum Fund, Vertex Growth, the Agricultural Industry Fund under CITIC Group, Yuzun Capital, and Dayang Motor, among others. Existing investors such as GSR Ventures, Vertex Ventures China, Capital Media, and Changshi Capital continue to follow on.

This certainty comes from the superposition of three layers of demand.

① Existing Operations and Maintenance: There is a vast number of global merchant ships, ports, and marine engineering facilities, with long-standing needs for hull cleaning, inspection, maintenance, and security.

The International Maritime Organization clearly points out that hull biofouling increases ship resistance, affecting fuel costs and air pollutant and greenhouse gas emissions. Biofouling management helps improve energy efficiency.

② Incremental Construction: Offshore wind power, offshore photovoltaics, marine ranches, and deep-sea exploration are all expanding the boundaries of marine engineering.

Data from the Ministry of Natural Resources shows that by 2025, China's gross ocean product will reach RMB 11.018 trillion, a year-on-year increase of 5.5%, accounting for 7.9% of GDP.

Among them, the added value of the marine engineering equipment manufacturing industry increased by 10.2% year-on-year, and the newly added grid-connected capacity of offshore wind power increased by over 60% year-on-year.

The larger the scale of the marine economy, the stronger the rigid demand for underwater inspection, construction, operations and maintenance, and emergency response.

③ Data Gap: NOAA public data shows that as of April 2026, only 28.7% of the global seabed has been mapped using modern high-resolution technology, with less than 0.001% of the deep-sea floor "seen" by humans.

This means that the ocean is not a space that has been fully digitized; it remains the largest under-sensed domain on Earth.

Whoever can continuously operate underwater has the opportunity to accumulate data that others cannot obtain.

The superposition of these three layers of demand explains why capital is willing to give marine robots a higher valuation: they possess both the scarcity of hard technology and chargeable scenarios, as well as the potential to build long-term data barriers.

The Significance of the "Cangqiong" Large Model Lies in Putting AI into the Action Chain

The "Cangqiong CEORION" released by Shihang Intelligence differs in that it is designed for marine operation robots, directly integrating model capabilities into real action scenarios such as inspection, detection, cleaning, grasping, cutting, welding, exploration, search and rescue, and emergency response.

While many previous marine intelligent systems focused on observing sea conditions, routes, and targets, calculating paths, risks, and energy consumption, Cangqiong aims to solve the problem of "how to act after understanding," moving from marine information intelligence to marine embodied intelligence.

Cangqiong CEORION adopts a unified end-to-end architecture, integrating environmental perception, task understanding, and action generation into a single model. It is trained using a combination of real operational data and simulated data and constructs a marine world model based on million-hour-scale commercial operational data.

It uses a unified end-to-end architecture, integrating environmental perception, task understanding, and action generation into the same model, rather than the traditional approach of stitching together multiple subsystems.

This architectural design allows the robot to directly output control commands from raw sensor data, significantly reducing intermediate latency.

In simulation tests, the task success rate exceeds 90%, the success rate of precise control and positioning for grasping exceeds 90%, and the zero-shot adaptation ability exceeds 70% when facing unseen sea areas, water quality, lighting conditions, and different robot platforms.

The built-in physical reasoning module can predict risks before action execution, reducing collision accident rates by 80%.

The barrier for marine robots lies not in the moment they enter the water but in whether they can continue to become smarter after each real operation.

Marine robot data is entirely different from internet data; first-person underwater operational data is extremely scarce.

The million-hour-scale commercial operational data disclosed by Shihang Intelligence provides it with the foundation to train a marine world model.

The term "commercial operations" here is crucial; it means the data comes from real customers, real sea conditions, and real tasks, rather than from closed pools or demonstration scenarios.

Light, turbidity, ocean currents, the surface state of target objects, the thickness of attachments, equipment wear, and operational errors in real tasks continuously constitute learning samples for the model.

The technical system publicly disclosed by Shihang Intelligence is precisely developed around this direction.

The six core systems—propulsion, control, sensing, navigation, sealing, and deployment and recovery—are all independently developed in-house. The robots possess full-ocean-depth (0m to 10,000m) and full-degree-of-freedom operational capabilities, supporting functions such as autonomous navigation and multi-robot collaboration. They have already been applied in scenarios such as ship cleaning, underwater security, offshore wind power, marine ranches, and seabed inspections.

Whoever possesses operational data under real sea conditions has the opportunity to train models into productive forces.

The Deep-Sea New Economy is Taking Shape, from Single-Point Operations to Full-Scenario Intelligence

The market's imaginative space continues to expand. According to Fortune Business Insights, the global underwater robot market size was $5.82 billion in 2025 and is expected to reach $19.66 billion by 2034, with a compound annual growth rate of 14.49%.

The Chinese market is growing even more rapidly. Data from Zhiyan Consulting shows that the domestic underwater robot market size reached RMB 14.6 billion in 2025, with a compound annual growth rate of 24.14% from 2020 to 2024.

Behind this growth is the concentrated outbreak (outbreak) of multiple billion-dollar scenarios, including offshore wind power, offshore photovoltaics, deep-sea aquaculture, and seabed pipeline operations and maintenance.

Marine robots are moving from single-task ship cleaning to a broader underwater operations market.

Scenarios such as marine photovoltaics, offshore wind power, marine ranches, seabed pipelines, underwater security, emergency rescue, and deep-sea scientific exploration all share common characteristics of high operational difficulty, strong human replacement potential, and sufficient payment capacity.

Shihang Intelligence's service to the CNNC Tianwan Photovoltaic Station in January and its selection for the Singapore Maritime Authority's underwater hull inspection and cleaning program in April indicate that the boundaries of commercialization have begun to expand.

As large model capabilities improve, robots will also move from cleaning and inspection to complex operations and maintenance tasks such as cutting, welding, and assembly.

Competition on the technological front is shifting toward full-stack deep cultivation. Shihang Intelligence has achieved independent development of the six core systems—propulsion, control, sensing, navigation, sealing, and deployment and recovery—and has entered the pain points of lifespan and reliability with its "Pangu" ten-thousand-hour magnetically coupled thruster.

Full-stack investment involves long cycles and high costs, but once successful, hardware and algorithms can form deep synergy, building a thicker technological moat.

Marine robots are moving from fragmented trial deployments to ecological collaboration, with the "chip-algorithm-body-operations and maintenance-data" chain gradually taking shape.

Moore Threads and Kunlun Core provide the computational power base, Shihang Intelligence delivers the complete machine and large model system, and enterprises like Dayang Motor support key components, while downstream shipping and energy companies release application scenarios.

Cost reductions and accelerated iteration will drive the industry away from its niche, hardcore status.

Continuous policy efforts have accelerated the pace of localization substitution. The "14th Five-Year Plan" explicitly lists deep-sea operational robots and intelligent underwater vehicles as key development directions for high-end marine equipment, setting a target of over 70% localization rate for key components by 2025.

The "Three-Year Action Plan for Building a Maritime Power," released in 2025, allocated RMB 2.86 billion in special funds, with nearly half directed toward subsidies for core component research and demonstration applications in underwater robots.

Coastal provinces such as Shandong, Hainan, and Guangdong have successively established industrial funds to promote industry-university-research collaboration.

The global landscape is also shifting. While European and American companies have long dominated the high-end market, Chinese enterprises are accelerating their catch-up through engineering efficiency, cost advantages, and the embodied large model route.

With the warm up (heating up) of port construction and marine energy development along the "Belt and Road," Chinese marine robots are expected to gain a more proactive position in the international market, technical standards, and industrial discourse.

The Large-Scale Deployment of General-Purpose Embodied Intelligence Will Happen in the Ocean Before Land

Issues such as fragmented scenarios, high manufacturing costs, insufficient general capabilities, and safety and compliance risks keep the large-scale commercialization of humanoid robots at the expected level.

However, progress in the marine sector is refreshing the industry's understanding of the deployment path for embodied intelligence.

Life originated in the ocean, and the large-scale commercial deployment of general-purpose embodied intelligence will also find its answer first in the deep blue waters.

Marine scenarios have stronger rigid demand attributes and shorter paths for commercial validation.

Humanoid robots attempt to replace generalized human labor in life and industry, with fragmented scenarios and highly elastic demand.

In contrast, marine robots directly replace high-risk, high-cost, and high-threshold underwater human operations, with rigid demand and quantifiable value.

The marine environment has stronger regularity, making it more suitable for the iterative evolution of embodied intelligence.

The terrestrial environment is complex and variable, with pedestrians, vehicles, and obstacles appearing unpredictably, imposing extremely high requirements on the robot's general decision-making capabilities.

In contrast, while the underwater environment has poor perception conditions, the operational scenarios are relatively closed, with highly standardized task processes and clearer decision boundaries for the model.

At the same time, the marine operations have relatively controllable error tolerance. Robots performing tasks underwater do not face the severe ethical pressures regarding life safety that terrestrial autonomous driving does, resulting in lower compliance costs for technological deployment.

The "data flywheel" is easier to form a closed loop in marine scenarios. The core competitiveness of embodied intelligence comes from data, which can only be accumulated through real operations.

Due to the fragmented deployment scenarios of humanoid robots, the data volume from a single scenario is limited, making it difficult to achieve large-scale data accumulation.

In contrast, once a marine robot enters a specific vertical scenario, it can continuously generate data through sustained operations, iteratively improving model capabilities. Enhanced model capabilities can then expand into more scenarios, forming a positive cycle.

Conclusion:

The AI industry is no longer solely focused on screens, factories, and urban roads; part of the next round of hard-tech competition is diving into the deep blue.

Whether marine robots can become the earliest branch of embodied intelligence to achieve large-scale profitability remains to be verified by time.

Partial references: 36 krypton : "Marine Embodied Intelligence Company 'Shihang Intelligence' Secures Record-Breaking $1 Billion Financing, with Jonah Zhu Betting Big," Wen Hui Bao: "The Largest Global Financing in the Marine Robot Sector is Born! Shihang Intelligence's Robots Can Retrieve Objects from the 10,000-Meter Deep Sea," China Fund News: "Jonah Zhu Makes Consecutive Moves: Marine Robot Company Shihang Intelligence Completes Over $1 Billion in Financing," NetEase Technology: "Cangqiong CEORION 'Dives' into the Ocean, with Large Models Taking on 'Physical Labor.'"