Putting large models into satellites: Is it a real boom for commercial aerospace, or just another capital story?

Recently, a highly fascinating story has become popular in the commercial aerospace and AI venture capital circles:

Ground AI data centers are almost draining the power grid, while space not only has inexhaustible 24-hour solar energy, but also a natural extremely cold environment of minus 270 degrees Celsius. Moving servers into space can even save electricity and air conditioning costs.

Catalyzed by this narrative, computing power satellites at home and abroad are being launched into orbit in clusters. From the "Pazhou Space Intelligent Computing Satellite Constellation" to the "Star Computing" program of Guoxing Aerospace, and then to the orbital computing power centers of SpaceX and Starcloud, "seeking electricity from the sky and processing space data in space" has become the latest consensus of the capital market.

But if you ask frontline satellite chief designers or thermal control engineers, they will tell you that the judgment on power in the first half is correct, while the common sense about heat dissipation in the second half is completely wrong.

When the global commercial aerospace industry is stuffing GPUs into space, we need to first tear off the packaging of concepts, calculate clearly the physical limit account of space computing power, the completely different commercial routes of China and the United States, and how the delicate AI chips can survive in space.

01

Dividends of the Dawn-Dusk Orbit and the Lie of "Natural Heat Dissipation in Space"

To send high-power consumption AI chips into space, the first step is to solve power supply and heat dissipation, which is exactly the most obvious watershed in the commercial aerospace field.

A close look at the "Star Computing" 02 group of satellites of Guoxing Aerospace, which recently completed the first networking, will reveal a key orbital parameter: this batch of satellites operates in the Dawn-Dusk Sun-Synchronous Orbit (Dawn-Dusk SSO).

An ordinary low-orbit satellite will fly into the Earth's shadow area for about one third of the time every time it orbits the Earth (about 90 minutes). Once entering the shadow, the solar panels will lose power immediately, and the satellite can only rely on heavy lithium batteries to support operation, which cannot drive high-power consumption computing cards at all.

The dawn-dusk orbit is a special orbit flying along the dividing line between day and night on Earth. The solar panels of the satellite can almost face the sun directly 24 hours a day, 365 days a year. Outside the atmosphere, the solar constant is as high as 1361 watts per square meter, without cloud attenuation, and the power generation efficiency is several times that of the ground. Seizing the dawn-dusk orbit does allow satellites to obtain nearly unlimited free electric power.

But with electricity coming in, how can heat be discharged?

The logic that some people like to talk about most is that the background temperature of space is close to minus 270 degrees Celsius (about 3 Kelvin), which is naturally a huge refrigerator. This actually makes the most basic physics mistake. Space is not a refrigerator, but a huge vacuum thermos cup.

In ground data centers, the computer room uses fans to blow (thermal convection) or cold water pipes to flow past the chips (thermal conduction) to quickly take away heat. However, in the high vacuum of space, there is no air medium, so both thermal convection and thermal conduction fail. Satellites have only one way to dissipate heat: thermal radiation.

According to the Stefan-Boltzmann law, the power of heat dissipated by radiation of an object is proportional to the fourth power of its absolute temperature. Under the upper limit of the allowable temperature for normal operation of chips (60°C to 80°C), the efficiency of space radiation cooling panels is actually very low, and only a few hundred watts of heat can be dissipated per square meter of the shaded heat dissipation surface.

We assume that a standard AI server equipped with 8 H100 GPUs in the ground computer room has a full load power consumption of about 10 kilowatts.

If this machine is moved into space as it is, even without considering the external heat flow brought by direct sunlight and Earth's infrared albedo, in order to dissipate 10 kilowatts of waste heat through radiation alone, it is necessary to deploy dozens of square meters of giant radiation cooling wings outside the satellite, and cover them with complex two-phase fluid loops and phase change heat pipes.

NVIDIA CEO Jensen Huang pointed out sharply when evaluating space data centers that the biggest engineering bottleneck of space computing power lies in cooling. The huge heat dissipation mechanism will cause the weight, volume and attitude control cost of satellites to rise exponentially. Most of the in-orbit computing satellites currently in operation have a single-satellite computing power consumption of only a few hundred watts to one or two kilowatts, which is less than half the power of a cabinet in the ground computer room.

Before the heat dissipation problem is completely solved, the "large-scale space data center" only stays on paper.

02

90% of Satellites Are Idling in Orbit, China and the United States Are Not on the Same Track at All

Since space heat dissipation is so expensive, why do Chinese and American commercial aerospace enterprises rush to send computing power into space?

Many people confuse Musk's "orbital data center" with the "processing space data in space" of Chinese enterprises. In fact, the two sides of the Pacific are calculating two completely different economic accounts.

The route of the United States is "seeking electricity from the sky", which is essentially energy arbitrage.

The expansion speed of the ground power grid in North America can no longer keep up with the throughput appetite of AI data centers.

After the merger of SpaceX and xAI, it directly submitted an application to the Federal Communications Commission (FCC) for up to 1 million orbital data center satellites, and started construction of the Gigasat factory in Texas.

Starcloud, a space computing unicorn incubated by YC, is more radical. It not only sent the first NVIDIA H100 chip into space for model training, but also purchased more than 50 Starlink Mini laser communication terminals from SpaceX.

Their plan is that since transformers and power grid quotas cannot be approved on the ground, they will simply send the computing power cluster to the dawn-dusk orbit to earn the all-weather solar energy dividend, and engage in pure "space cloud computing power leasing".

However, the current wave of sending computing power into space in China's commercial aerospace industry is to develop "space-based edge computing" to solve the bandwidth shortage problem of remote sensing satellites.

Lu Jianhua, academician of the Chinese Academy of Sciences and professor of Tsinghua University, previously pointed out a harsh industry truth that 90% of China's remote sensing satellites are idling in space.

Another data given by Wang Shangguang, dean of the School of Computer Science of Beijing University of Posts and Telecommunications and chief scientist of the Tianconstellation, is equally severe, which shows that less than 10% of the effective data of a remote sensing satellite operating at full load can be successfully transmitted back to the ground.

In the past ten years, China's commercial aerospace industry has launched a large number of remote sensing satellites, but most of them are still at the stage of "digital cameras hanging in the sky". The traditional operation process is "taking photos in space, transmitting data when passing over the ground station, and splicing and analyzing images on the ground". There are two deadlocks here:

The first is the extremely high scrap rate. More than 50% of the Earth's surface is covered by clouds all year round, and more than half of the massive original images taken by optical satellites are worthless white clouds.

However, traditional satellites have no "intelligence" to distinguish which image is cloud and which is the target. They can only store hundreds of GB of scrap images on the on-board hard disk, and wait for the short window of a few minutes when flying over the ground station, occupying the extremely expensive space-ground radio frequency bandwidth to transmit data forcibly.

The second is zero timeliness. In scenarios such as forest fire prevention, flood breach, and offshore oil spill, if the response is delayed by one hour, the number of people to be evacuated and the area of burned forest will be completely different.

After the satellite takes photos, orbits half the Earth to find the ground station for data transmission, and then the ground data center completes processing and sends the results to the commanders, the time often counts in "days", and the golden rescue window has long been closed.

China's "processing space data in space" aims to move data cleaning directly to space.

After the satellite takes 1TB of original remote sensing images, the on-board large model immediately removes the scrap images covered by clouds, identifies the coordinates of mountain fires or ship trajectories, and finally transmits only 1KB of "result instructions" back to the ground through narrowband. Changing from "transmitting original data" to "directly outputting answers", the data volume is compressed by hundreds of millions of times, and the response time is reduced from days to seconds.

03

Two Divergent Routes in China: Developing On-board Intelligence and Building Constellations

Limited by the heat dissipation and volume of a single satellite, it is impossible to install too much computing power in space. Facing this physical constraint, the two leading players in China at present have come up with two completely different engineering solutions.

The first route is the algorithm school represented by the "Pazhou Space Intelligent Computing Satellite Constellation", with the core of "developing on-board intelligence".

The driving force behind this constellation is Galactic Computing, founded by Wang Kai, a Tsinghua graduate and California Institute of Technology doctor.

This company originally focused on AI algorithms and intelligent transportation. Their entry point is that since the computing power of a single satellite is limited, they will not blindly stack hardware, but develop a software and hardware integrated "space-based world model". Through extreme model quantization and pruning, the multi-modal large model is embedded in the power-constrained on-board chips.

In terms of business model, they skip the traditional mode of "selling satellites" or "selling original remote sensing images", and directly sell industry analysis reports processed in orbit to customers in emergency management, agricultural yield estimation, financial insurance and other fields. Even before the satellites are launched, many ground industry customers have prepaid for computing power packages.

The second route is the infrastructure school represented by Guoxing Aerospace, with the core of "building constellations".

The heat dissipation and power of a single satellite cannot support a large model with hundreds of billions of parameters, so inter-satellite laser communication is used to connect scattered computing power fragments together.

The "Star Computing" program jointly built by Guoxing Aerospace and more than 100 universities and research institutes in China plans to deploy 2800 computing satellites, of which 2400 are mainly for reasoning and 400 are mainly for training. Through high-bandwidth inter-satellite laser links, the originally isolated satellites are connected into a space-based local area network.

After a satellite captures data, if its own computing card is too busy to process it, it will directly transmit the data to the idle satellite nearby through laser for collaborative computing.

There is no distinction between superior and inferior for these two routes. The former is closer to cash flow, while the latter is betting on the underlying infrastructure ticket for the future space-based distributed computing power network.

04

How Commercial GPUs Resist Space Particle Bombardment

Even if power supply, heat dissipation and inter-satellite networking are solved, there is still a deep-seated engineering deadlock for space computing power, that is, chip service life.

According to the rules of traditional aerospace, the chips sent to space must be "space-grade (Rad-Hard)". These chips are processed through special radiation-hardening processes such as SOI Silicon-On-Insulator, which can resist the bombardment of high-energy protons and heavy ions in space.

But the cost is extremely high, and the process technology is seriously backward. The computing power of the current mainstream space-grade chips is roughly at the level of smart phones more than ten years ago, and even the basic Transformer visual model cannot run.

To run AI in space, industrial-grade or even commercial-grade GPUs and NPUs must be used, such as NVIDIA Jetson series or domestic high-computing AI chips. However, once these chips with advanced process technology lose the strict protection of the Earth's magnetic field, they are extremely vulnerable to Single Event Effect (SEE) in space:

Once high-energy particles hit the transistor, at the slightest level, Single Event Upset (SEU) will occur, turning 0 in the memory into 1, resulting in disordered reasoning weights of the large model, garbled output, or even direct system crash; at the worst, Single Event Latchup (SEL) will be triggered, forming a large current short circuit inside the chip, burning the entire computing board in a few seconds.

How to make the delicate commercial high-computing chips survive in space?

Another core player deeply engaged in space-based computing in China, Zhongke Tiansuan, recently disclosed a very representative underlying patent - "domestic high-computing payload computing platform based on dual-domain architecture". This architecture is a very practical breakthrough idea.

Engineers divide two physical domains on the same computing payload. The high-performance domain directly uses cheap, high-computing domestic commercial AI chips, which are specially responsible for heavy workloads and running complex neural network reasoning; while the high-reliability domain uses space-grade small chips with extremely high safety level and radiation resistance, which do not undertake heavy workloads, but act as "supervisor" and "watchdog".

Once the high-energy particle bombardment causes Single Event Upset or current abnormality in the commercial AI chip, the space-grade chip in the high-reliability domain will detect the problem in milliseconds, immediately cut off the power of the faulty chip to prevent burnout, and then reload the clean system image for it and resume the computing task.

Through this "high-low matching" architectural innovation, the availability of commercial chips in orbit operation has been raised above the engineering passing line.

From the intensive in-orbit networking of domestic computing satellites to the heavy bet of SpaceX and Starcloud on orbital data centers in North America, a key industrial inflection point of commercial aerospace has emerged.

As the launch cost is continuously reduced by the liquid oxygen methane reusable rockets, the manufacturing logic of satellites is undergoing a fundamental reversal.

In the past, rocket transportation was as expensive as gold, so satellites had to be extremely lightweight and extremely conservative, and even backward chips from ten years ago had to be used to ensure absolute safety;

Now, with the decline of launch cost and the maturity of inter-satellite laser communication, the commercial aerospace industry finally has the confidence to send industrial-grade high-computing chips, phase change heat dissipation loops and large models into the dawn-dusk orbit.

But for the investors and industry customers who actually pay, the disenchantment has only just begun. For a very long period of time, space data centers cannot compete with ground thermal power and hydropower clusters in terms of general-purpose computing power cost. Its real golden scenario is still locked in "data generated by space-based system is digested locally in space".

In this hidden war of sending computing power into space, what ultimately determines the survival of a company is how much loss the 1KB of data transmitted back to the ground can help ground customers avoid when the satellite flies over forest fires, ocean routes and farmland.

This article is from the WeChat Official Account "Xingdong Wuji", author: Julian, published with authorization from 36Kr.