Space Is a Terrible Place to Cool a Data Center

0:00 Google's first AI satellite can only run its chips for about 15 minutes at a time. It's scheduled to launch on the 1st of October. The limit is heat in space, which is about as cold as

0:10 anywhere gets. So, why does a chip overheat in the coldest place there is? Answer that and you can size a space radiator with one formula and a calculator. Then you can use it to judge

0:20 the next orbital AI pitch that lands in your feed. The satellite's called MVP. According to RS Technica, it's about the size of a fridge built with Planet Labs, and it carries four of Google's own AI

0:32 chips called TPUs. Its solar panels make about 1 kilowatt. That's roughly a haird dryer. So, Google's first data center in orbit runs on a haird dryer's worth of power, and it still can't keep cool.

0:44 StarCloud already put an NVIDIA H100 up in November 2025, and SpaceX has filed for up to a million data center satellites. All of them have to get their heat out the same way. Start with

0:55 a thermos. Your coffee stays hot because there's a vacuum between the two walls. Heat can't cross that gap by touching something or riding on air because there's nothing in it. Space is that gap

1:05 with the outer wall taken away. So, a satellite up there is less parked in a fridge than sealed in a thermos with one exit left, glowing its heat away. On Earth, a server sheds heat in two steps.

1:17 The chip warms a metal heat sink by contact. Then air or water carries that heat off somewhere else. In orbit, step two just doesn't exist. Google says it plainly. In space, there's no air flow.

1:29 And in a vacuum, you can only get rid of heat through radiators. A radiator up there is just a panel that glows in infrared. You can't see it, but it carries real power away. That glow is

1:39 the only way out. Every warm surface gives off infrared. And one formula says how much. Take the temperature in Kelvin and raise it to the fourth power. Multiply by a tiny fixed number called

1:50 sigma about 5.67 x 10 the minus 8. Then multiply by emissivity. How good the surface is at glowing on a scale from 0 to 1. Out come watts per square meter. Try it on a radiator at 27°

2:04 about a warm room. That's roughly 300 kel. A perfect surface would shed about 460 watts per square meter from each side. A good real world coding with an emissivity of 0.9 manages about 413.

2:18 These are simplified numbers. We'll add the messy parts back in a minute. Flip that into a size. At 413 watts per square meter, every kilowatt of chip heat needs about 2.4 m of glowing

2:28 surface. A flat panel glows from both faces. So call it 1.2 m of panel per kowatt. Scale that to a single megawatt and you need about 12,200 m of two-sided panel. That's a square roughly 35 m on

2:42 each side for 1 megawatt at room temperature. Even a perfect surface only trims it to about 1,090. Back to MVP. For its 1 kowatt, the sum says about 1.2 m of panel. On a fridge size satellite,

2:56 that sounds doable. MVP moves its heat with heat pipes. Those are sealed tubes where a fluid boils at the hot end and condenses at the cool end, feeding radiators. Yet, it still runs in bursts

3:06 of about 15 minutes. Google hasn't published its radiator size, so those 15 minutes can't be pinned on one cause. But at 27°, that sum is the best case. Sunlight, Earth's glow, and the losses

3:19 between chip and panel only take away from it. Take those one at a time. Sunlight lands on the panel. The warm earth below shines infrared back at it and reflects sunlight, too. And heat

3:30 loses a few degrees at every hop from chip to coolant to panel. So, the panel runs cooler than the chip. Each of those makes a real radiator do worse than this estimate. The lever is in the formula

3:41 itself. It uses the fourth power of temperature. So, a small rise pays off big. Warm the radiator from 27 to 77°. In Kelvin, that's 300 to 350, only about 17% hotter. The heat it sheds jumps to

3:56 about 767 watts per square meter, 1.86 times as much from the same panel. SpaceX is betting on exactly that. In June, it unveiled a satellite called AI1,

4:07 150 kW of compute at peak, 120 on average, a 70 meter wingspan, and a 110 m deployable liquid radiator. Divide 150 kW by 110 m, and you get about 1,360 W per square meter of panel. If both faces

4:25 radiate, that's about 680 per face. run the formula backwards and the radiator has to sit somewhere around 58 to 67° C. That temperature is my inference, not a SpaceX figure. If the 110 m already

4:39 counts both faces, it's closer to 130°. Either way, compare it with room temperature. At 27°, the same 150 kW would need about 180 m of panel. Hotter radiator, smaller radiator. Musk says

4:54 most of AI1, including the solar and the radiators, is technology already built for Starlink V3. There's not some magic that's necessary that doesn't exist, he said. He also called it much simpler

5:06 than a Starlink satellite, which is a comforting thing to hear about a machine with a 70 m wingspan. All of that is company claims, and AI1 hasn't flown. So, why not run everything hot? Because

5:17 heat only flows downhill. The chip has to be hotter than the coolant and the coolant hotter than the radiator. Raise the radiator's temperature and the whole slope lifts. With the chip at the top,

5:27 the fourth power rewards a hot radiator. The chips would rather stay cool. Every orbital design has to pick a spot on that slope. Google doesn't claim to have won it. Its paper calls effective

5:37 thermal management a critical optimization challenge for power dense TPUs in a vacuum. It names heat pipes and radiators, gives no radiator area or temperature, and leaves testing in orbit

5:48 as future work. Its research blog lists heat among the significant remaining challenges. MVP is the first chance to get real numbers from orbit. Heat's the core, but four other problems come with

5:59 the job, and each has a proposed fix. Power first. Google's design uses a dawn dusk sun synchronous orbit, which rides the line between day and night. There, a panel can collect up to eight times more

6:10 solar energy a year than one at mid latitude on Earth, nearly non-stop. So, it needs fewer batteries. Radiation. Next, Google put its trillium TPUs in a proton beam on the ground. The high

6:22 bandwidth memory started showing irregularities at 2 kads, almost three times the 750 rads a shielded chip should absorb over 5 years. Nothing failed outright up to 15 kads, the

6:33 highest dose tested. So, the fix here is mostly shielding plus chips that turned out tougher than the job needs, at least in a lab beam. Then the network. Google's illustrative design flies 81

6:44 satellites in a cluster 1 km in radius, about 650 km up, with neighbors only 100 to 200 m apart. They talk by laser. On a bench, one transceiver pair hit 800 Gbits per second each way over a short

6:58 open air path. Promising, but it's a lab link in a model cluster. Cost comes down to launch because every square meter of that radiator has to be lifted. Google's own analysis projects launched to low

7:09 Earth orbit could fall to about $200 a kilo or less by the mid 2030s if that forecast holds and repairs a failed chip in orbit stays failed. Google's paper calls swapping hardware up there

7:22 obviously impracticable. There's no remote hands ticket that gets a technician to 650 km. So the simplest fix is to launch spare hardware from day one. Quick status check. In December

7:33 2025, StarCloud 1 ran Google's Gemma model and trained NanoGPT by its own account. In February, the FCC accepted SpaceX's application for filing. That starts the review. It isn't an approval.

7:46 And Google's MVP is an extra earlier test. The two satellite laser link mission is still planned for 2027. Pull the numbers together. about 413 watts per square meter per side at 27° around

8:00 1,200 square meters of two-sided panel per megawatt. Going from 27 to 77° buys you 1.86 times the cooling from the same panel. And AI1's radiator only adds up if it runs somewhere near 60° or more.

8:13 Google's own SVP James Mana is blunt about the timeline. We don't expect to be perfectly frank that we'll have anything usefully operational in the next few years. My verdict. If you're a

8:25 developer or a team wondering whether to plan around compute in orbit, I wouldn't. Not for the next few years. Google saying the same thing. When the next pitch shows up, ask two things.

8:35 What temperature does the radiator run at? And how many square meters per megawatt? 1200 at room temperature is your yard stick. A pitch that skips both hasn't answered the hard part. MVP is

8:46 scheduled to fly on the 1st of October, running in 15-minute bursts. Once it's up, the question worth watching is simple. Does its thermal data show a way past the problem Google itself lists as

8:57 a remaining challenge?