On August 5, Oriental Starlink launched the Oriental Wisdom High-Spectral AI Dual Satellites, featuring 22 spectral channels and 5-meter resolution, capable of identifying the "spectral fingerprints" of Earth's materials. The dual satellites are equipped with 400 TOPS onboard AI computing power, enabling on-orbit object recognition and anomaly detection. The company envisions a "Planetary Physical AI" future, where satellites serve as eyes, models as brains, communications as nerves, and robots as limbs—ultimately turning the entire Earth into an AI training ground. Oriental Starlink plans to deploy 2 million AI robots on the Moon to form a modular, mobile lunar data center. The company currently operates in over 40 countries and aims to complete the deployment of 258 satellites by 2032, with a lunar landing targeted around 2030.Article author and source: AI World
Too exciting.
Guys, I just completed a nearly one-hour in-depth interview.
The moment I turned off the microphone, a crazy idea popped into my head:
Did I just see the "Elon Musk of China"? And another possibility for China’s space AI innovation?
Don't say I'm exaggerating yet. Look at what they're actually doing.
The first thing they did was build a planetary-scale Physical AI.
The specific approach is to first launch two "eyes" into space to sense the Earth. These hyperspectral satellites can capture the unique "spectral fingerprint" of every substance, much like human fingerprints. So, instead of simply seeing "this is green," the spectral satellites go further—identifying what type of plant this green is, whether it’s healthy, and whether it’s lacking water.
Two "eyes" continuously read the "fingerprints" of mountains, rivers, soil, forests, oceans, minerals, and crops from space, linking computation, communication, and action—enabling AI to gradually understand Earth’s past and present, and predict its future!
In other words, they turned the entire planet into the largest AI training ground ever.
Incredible. This is the real deal Earth physics AI.
But this is only the first step. The second thing is even crazier.
Send embodied intelligent robots to the Moon. Ultimately, build a cosmic-scale data center using 2 million lunar AI robots!
Not those classic robots that are deployed with a few units, take a couple of steps, snap a few photos, and then wait for remote control from Earth.
Instead, let the robot autonomously perceive the lunar surface environment, autonomously assess risks, and autonomously plan routes—even automatically identifying potential resources around it based on its understanding of the terrain and material environment.
Next, we’ll build a modular, mobile, and interoperable lunar data center.
How about that? It’s incredible. Things that just a few years ago could only be imagined in sci-fi movies are now being steadily advanced.
The company is called Oriental Starlink. I interviewed the company's founder.
On August 5, they launched the two "eyes" mentioned above—Dongfanghuiyan High-Spectral AI Dual Satellites 05 and 06—into space.
AI has a super brain
yet still a "terrestrial blind person"
Today's large models are very intelligent. They have read the largest corpus of text in human history, seen hundreds of millions of images, and can write code, solve math problems, and generate videos.
But its understanding of the real Earth remains extremely limited. It knows the term "mangrove," but it does not know whether a specific mangrove forest somewhere on the planet is currently expanding or dying.
In the interview, he repeatedly emphasized a perspective: the AI industry is currently talking about space computing, but most attention is focused on “how to put computing power into space,” while overlooking something even more scarce—continuous data about the real physical world. He even believes that certain unique data may be more important than simply sending computing power into space.
Indeed, even Demis Hassabis of DeepMind has said that the vast majority of data humans possess remains at the level of the internet and language, with very little data about the complete Earth and universe.
The first pair of eyes of a planetary-scale AI
On the morning of August 5, the Dongfang Xinglian 05/06 satellites (Dongfang Huiyan hyperspectral AI dual satellites) were launched into space aboard the Jielong-3 No. 12 rocket off the coast of Haiyang, Shandong.
Regular remote sensing satellites are like taking photos of Earth from space—capturing color, shape, and area. Hyperspectral remote sensing, however, is like performing a medical exam and lab test on Earth from hundreds of kilometers away.
Spectra are the fingerprints of matter; with 22 precisely calibrated spectral channels covering the visible to short-wave infrared range, combined with a 5-meter spatial resolution and a 300-kilometer ultra-wide swath, these two satellites outperform all similar satellites in Europe and the United States.
Meanwhile, these two satellites also began to develop their own understanding.
The previous process involved satellites taking photos, transmitting them back to Earth, processing them at a center, analyzing them by experts, and then generating a report. These two satellites, however, are equipped with onboard AI computing power of 400 TOPS, enabling them to perform anomaly detection, land feature recognition, intelligent compression, and feature extraction in orbit, directly transmitting the results to the ground.
What ultimately reaches the user may not even be a satellite image, but merely a brief message: a risk of crop disease in a certain field, abnormal composition in a stretch of water, or an immediate inspection needed in a specific location.
Provide two real-world examples.
Cassava: AI in food decision-making. Cassava is the primary crop in Lampung Province, Indonesia, and inaccurate yield forecasts have direct consequences—underestimating leads to unnecessary over-imports and waste; overestimating may leave locals genuinely hungry. Even in today’s era, in a country that isn’t particularly poor, this still happens. What they need isn’t just a satellite image, but insights like: “How much was planted this year, how is it growing, what’s the likely final yield, and how should we plan imports and harvests?”—from spectral data all the way to actionable intelligence and decisions.
Locust swarms: Because free data wasn't good enough, they decided to build their own satellites. In the early stages, the team used international open-source data to track locust migrations across Africa for three consecutive years, aiming to predict which crops would be destroyed and which regions would be affected. They found that the data quality and accuracy were far too low. It was because they couldn't solve real-world problems that they decided to build better satellites themselves.
After entering orbit, the two satellites will work in conjunction with the already operational Dongfang Huiyan Gaofen-01 satellite to form a "general survey + detailed survey" service network. UN-Habitat and the Global Mangrove Center have also become partners.
Oriental Star Chain Satellite 01
Academician Li Deren once said: The East Wisdom High-Spectral AI Twin Satellites represent not only a major breakthrough in high-spectral remote sensing technology, but also a technological model embodying the concept that "lucid waters and lush mountains are invaluable assets." They will transform from merely "seeing" the Earth to truly "understanding" it.
Academician Li Deren and Professor Wang Mi inspect the dual satellites.
Build a "Brain for Earth"
Two satellites, just an entry point.
What truly sets this company apart is their comprehensive vision for "planetary-scale physical AI"—continuously observing Earth, performing intelligent computations in space and on the ground, transmitting results in real time via communication networks, and delivering those results to humans, agents, and robots for execution.
Thus, the satellite becomes the eye, the model becomes the brain, communication becomes the nerve, and the robot becomes the limbs. The entire system operates like a living entity.
They call this the "distilled Earth Large Model." It doesn't just answer questions like "What's here?" or "What's there?"—it gradually addresses deeper questions: Why did this region form such landforms? Why is this body of water changing? Why is this ecosystem degrading? What might happen next? And what actions should humans take?
This is the true core of Physical AI.
Earth is precisely the best place to train this type of model—satellite observations can be continuously validated against ground truth, and humans can provide feedback on whether crops are yielding less, water bodies are being polluted, or forests are degrading.
The internet has trained AI that can speak; the entire planet may train AI that truly understands the physical world.
According to the strategic plan, the "Oriental Eye" constellation will deploy 258 satellites by 2032, with a total computing power of 900,000 POPS across 1,000 "Space Strings" computing satellites. The company has already conducted business in over 40 countries, with its in-orbit AI platform operating securely for more than 2,600 days. In 2026, 23 space computing platforms will be manufactured, ranking first in China for both computing power and shipment volume.
The moon is the first real test.
This part was the most exciting part of the entire interview.
The previous logic has been established—satellites provide the AI with eyes, Earth data trains its brain, and the communication system forms its nerves. Next, it needs a real body.
This body can be a lunar robot.
Dewu II, in collaboration with overseas universities, has developed a project called CHERI—short for "Challenging Environment Intelligent Exploration Robot"—to provide AI capabilities for the dual-star system. It is the first private company to participate in China's lunar exploration program, with plans to land on the Moon around 2030.
The core of the project consists of two miniature lunar robots working in coordination. Though small in size, they are designed to perform remarkable tasks: autonomously navigate the lunar surface, perform real-time mapping (SLAM), intelligently avoid obstacles, communicate with each other in real time via a MESH network, and collaboratively carry out exploration missions.
Two robots simultaneously perceive the environment, share data, and coordinate tasks. One leads the way to map the area, while the other uses the shared map to perform different tasks.
Moon AI Robot concept art
In the interview, he revealed a number that left me stunned: ultimately, 2 million lunar AI robots will be needed.
These two million robots are not simply walking around on the moon. Each one is an intelligent node—capable of sensing, computing, communicating, and collaborating.
When two million such nodes are deployed across the lunar surface, interconnected via a MESH network to share data, allocate computing power, and collaboratively execute tasks, they form a distributed computing network.
In other words, two million AI robots connected together are the Moon data center itself.
Highly modular, mobile, and capable of being deployed distributively or aggregated into ultra-large-scale computing clusters, flexibly adjusted according to lunar illumination and mission requirements. The goal is to reach a significant scale around 2035.
The two robots of CHERI validated the smallest prototype of this architecture: whether two nodes can collaboratively perceive, collaboratively map, and collaboratively decide. Once validated, the next step is engineering scaling from 2 to 2 million.
In the interview, he also described in great detail a training regimen: selecting several beaches in China—because lunar regolith has a texture similar to sand—to collect data under different lighting and temperature conditions; conducting similar experiments in deserts in the Middle East; and then fusing all this data to simulate the lunar surface environment and continuously train the robot’s autonomous navigation model.
More importantly, the robot took to the Moon all the knowledge AI has accumulated on Earth about matter, environment, and planetary evolution.
When the robot lands on the moon, it will use the knowledge it learned on Earth to determine what is beneath its feet, what might be nearby, and where it should go next.
Two hyperspectral satellites are the first eyes through which Planetary AI sees Earth; two lunar robots may become its first steps toward another planet.
He even discussed a more ultimate vision: continuously recording the evolution of matter on Earth using spectral data, training a model that can explain "why life exists on Earth," and then using that model to look outward in search of places beyond Earth where life might exist.
This truly opens a new window in the search for extraterrestrial life.
Is it China's SpaceX?
After this interview, I still can't say for certain whether this company will ultimately become China's SpaceX. After all, SpaceX possesses a complete end-to-end capability with rockets, spacecraft, and Starlink, while Dongfang Starlink is still at the beginning of a long journey.
But by the end of the interview, I realized the question “Is it China’s SpaceX?” was fundamentally misguided. Musk solved the problem of “how to get there”—using reusable rockets to drastically lower the cost of accessing space, and deploying Starlink to expand communications. This is the work of building roads and laying cables—the first layer of infrastructure for the space age.
What happens after the roads are built? When launch costs drop low enough, what becomes truly scarce is no longer capacity, but what you can do once you’re up there.
The answer from Eastern Star Chain is: Don't build more roads; let space grow intelligence on its own.
It’s too early to draw conclusions about whether this answer can be fulfilled. The 258 satellites are scheduled for 2032, lunar robotic landings are expected around 2030, and the lunar data center plan is set for 2035.
But none of these numbers belong to a distant future. Right now, AI has opened eyes that can read the Earth.
While humans are preparing to send AI beyond Earth, it is learning to truly understand its home.