Google wants to move the data center into space, but the first difficulty is exposed: 4 TPUs can only run at full speed for 15 minutes. (Diagram, Reuters)
[Financial Channel/Comprehensive Report] Google is preparing to put AI computing into space, but the first bottleneck that emerges may not be the cost of rocket launch, but how to dissipate the heat generated by the AI chip. According to reports, every watt of power consumed by a space chip is converted into heat. A chip that consumes one kilowatt of power will generate one kilowatt of heat, but this heat must be dissipated, otherwise the chip will overheat and stop working.
The New York Times reported that the MVP satellite operates with a power of about 1 kilowatt and is equipped with four chips. These chips will shut down for cooling after about 15 minutes of operation. Even the smallest test requires a stop. Therefore, objectively speaking, heat determines the development speed of data centers. On the ground, electricity determines the speed of data center development, while in orbit, current evidence suggests that heat will play a decisive role.
Please read on...
The article pointed out that Google's "Sunshine Catcher" project will put four AI chips into orbit on October 1. The test satellite MVP was built with the help of Earth observation satellite company Planet and launched from California on a SpaceX rocket. Google hopes to test the feasibility of deploying AI computing equipment in space and use orbits close to continuous sunlight to obtain solar energy.
According to Google research, in a specific morning and evening orbit, the energy obtained by solar panels in a year can be up to 8 times that of ground-based solar panels in mid-latitude areas such as the United States or Europe. But when the outside world discusses Project Suncatcher today, the focus is mostly on cost.
Estimates used in Google research show that currently, Falcon 9 The cost of sending a kilogram of objects into low Earth orbit is about US$3,600. If it drops to about US$200 per kilogram (approximately NT$6,340) by the mid-2030s, based on a design similar to Starlink satellites, the launch cost will be amortized. The post-advance cost is approximately US$810 per kilowatt per year (approximately NT$25,667), which falls within the range of US$570 to US$3,000 per kilowatt per year (approximately NT$18,069 to NT$95,100) for electricity costs for ground data centers in the United States.
It is worth noting that this comparison is not "the total cost of space data centers and ground data centers." Google's study also made it clear that the calculation did not include costs such as chips and construction and "does not constitute a complete economic analysis." In other words, the comparison is between satellite launch costs and ground data center power costs.
In addition, to bring the launch cost down to about $200 per kilogram, the study estimates that about 1,800 Starship launches will be needed, which is equivalent to about 180 times per year. In addition to cost, another problem exposed by the MVP test is heat dissipation.
The electricity used by AI chips will eventually be converted into heat. In ground-based data centers, fans, air or water can be used to remove heat, but there is no air in space, so heat must mainly be emitted by radiation through radiators. Google engineers pointed out that each square meter of heat sink can only dissipate a few hundred watts of heat, but a small AI chip may generate thousands of watts of heat, so each TPU may require several square meters of heat dissipation area.
The MVP satellite has a power of about 1 kilowatt and is equipped with 4 chips. The most noteworthy thing is that these chips can only operate for approximately 15 minutes at a time, and then must be stopped and cooled down. This "15 minutes" has also become an important observation indicator for the first test of Project Suncatcher.
Experts point out that the proportion of time the machine can continuously operate at full power will directly affect the actual computing efficiency; if the chip can only work intermittently, the effective computing time it can provide will decrease with the same satellite weight and launch cost.
The article points out that this message does have a downside, as the cooling method is not new. According to Google engineers, this thermal control technology has a rich history of application on previous spacecraft, and the team has a sealed room that can simulate the cold and airless environment of space.
Today, they are working on heat pipes and pumped fluid circulation systems to deliver heat to radiators. Over time, more advanced tools may increase thermal limits. Therefore, an objective interpretation is that heat determines the development speed of data centers. On the ground, electricity determines the speed of data center development. And in orbit, current evidence suggests that heat will play a decisive role.
However, this time MVP was not tailor-made for space AI computing from the beginning. The report pointed out that Google hopes to put the chip into orbit this year. Therefore, using satellites that have been designed by Planet, if larger radiators and specially designed heat pipes and fluid circulation systems are used in the future, the operation time may be different.
On the other hand, the satellites planned by Google in the future will carry dozens of chips. The more chips there are, the greater the heat generated, and the required heat dissipation area and weight also increase. The cost of rocket launch is directly related to the weight.
Therefore, one of the questions that this test really wants to answer is how long the AI chip can maintain full power operation after it enters orbit. MVP is expected to operate in orbit for about a year. The actual operation results and the size of the cooling system required for the next generation of satellites will become key data for Google to verify the feasibility of space AI computing.
Google research director James Manyika also said that we should not expect this technology to reach a truly practical scale of operation in the next few years.
Grasp the economic pulse with one hand. Click here to subscribe to Free Finance Youtube Channel
