Data Centers In Spaaaace!

Pigs in Spaaaace!

The AI bubble has been kept inflated by journalists uncritically reporting whatever CEOs say as they frantically pump the stock. Right now, you can observe a wonderful example of this by searching the Web for "orbital data centers". My recent search turned up pages of links, including SpaceX’s Lofty IPO Valuation Hinges on Big Bet on Outsize Growth from Bloomberg's Bailey Lipschultz, Sana Pashankar, and Loren Grush:

To buy into SpaceX’s audacious $1.5 trillion valuation in a listing next year, investors will need to have faith in Elon Musk’s equally galactic vision for his rocket and satellite maker, from orbital data centers to lunar factories to human settlements on Mars.

I chose one that ought to be more credible than Musk from Scientific American. Jeremy Hsu's Data Centers in Space Aren’t as Wild as They Sound reports that:

In early November Google announced Project Suncatcher, which aims to launch solar-powered satellite constellations carrying its specialty AI chips, with a demonstration mission planned for 2027. Around the same time, the start-up Starcloud celebrated the launch of a 60-kilogram satellite with an NVIDIA H100 GPU as a prelude to an orbital data center that is expected to require five gigawatts of electric power by 2035.

To do Hsu justice, he did point out a few of the problems. But follow me below the fold for more.

A couple of weeks ago an anonymous engineer posted Datacenters in space are a terrible, horrible, no good idea:

In the interests of clarity, I am a former NASA engineer/scientist with a PhD in space electronics. I also worked at Google for 10 years, in various parts of the company including YouTube and the bit of Cloud responsible for deploying AI capacity, so I'm quite well placed to have an opinion here.

The short version: this is an absolutely terrible idea, and really makes zero sense whatsoever. There are multiple reasons for this, but they all amount to saying that the kind of electronics needed to make a datacenter work, particularly a datacenter deploying AI capacity in the form of GPUs and TPUs, is exactly the opposite of what works in space. If you've not worked specifically in this area before, I'll caution against making gut assumptions, because the reality of making space hardware actually function in space is not necessarily intuitively obvious.

You should absolutely read the whole thing, it is magnificent. I don't have this background, but I don't need it to know that they are a terrible idea. All I need is arithmetic.

Steven Clark reports that Investors commit quarter-billion dollars to startup designing “Giga” satellites:

K2 is designing two classes of satellites—Mega and Giga .... The company’s first “Mega Class” satellite is named Gravitas. Gravitas will also deploy twin solar arrays capable of generating 20 kilowatts of power.
...
K2 says Gravitas is “on par with the largest satellites that have ever been produced.” But K2 won’t stop there. The firm’s next satellite iteration, known as Giga, .... Underpinning Giga is its ability to generate up to 100 kilowatts of power per satellite ... Examples of missions Giga can support include AI computing

Gravitas can launch on a Falcon 9, but Giga-class satellites need Starship. In Power: The answer to and source of all your AI datacenter problems Tobias Mann reports that:

Today, Nvidia's rack systems are hovering around 140kW in compute capacity. But we've yet to reach a limit. By 2027, Nvidia plans to launch 600kW racks which pack 576 GPU dies into the space once occupied by just 32.

So the largest satellites ever can power 14% of a current Nvidia rack. The next generation that requires Starship to launch can power 70% of a current Nvidia rack. By the time they can launch it, it will power 17% of a contemporary Nvidia rack. See, progress!

Starcloud's "5GW" satellite could power over 8000 2027 Nvidia racks in 2035 but it would be 50,000 times larger than a satellite that is much larger than one "on par with the largest satellites that have ever been produced".

The next question is "How many racks does one data center hold?" Yasir Shinaishin of ABB wrote in New Strategies in Design to Meet the Demands of AI Data Centers:

It is not uncommon to see data centers of 1 GW with liquid-cooled racks with densities over 100 kW.

Say 7,000 current Nvidia racks. So they need 10,000 Giga-class satellites to match one 2025 data center. Lets say they start launching in 2027 and launch 100 satellites/year. They can match one 2025 data center in 2127. Maybe the AI bubble will have burst by then...

Source

But the useful life of the racks in earth-bound data centers is less that 5 years. In space the chips are subject to high levels of radiation, which as the anonymous engineer explains, is bad news:

For longer duration missions, which would be the case with space based datacenters because they would be so expensive that they would have to fly for a long time in order to be economically viable, it's also necessary to consider total dose effects. Over time, the performance of chips in space degrades, because repeated particle impacts make the tiny field-effect transistors switch more slowly and turn on and off less completely. In practice, this causes maximum viable clock rates to decay over time, and for power consumption to increase.

But lets ignore that and asume a useful life of 5 years. Thus to keep one 2025 data center operating in space they need to launch 2000 enormous satellites per year, or over 5 per day. It is true that solar power in space is free, but launches aren't. Adding the cost of 10,000 launches to the cost of replacing the 60% of the cost of the data center that is the racks is going to make the already impossible economics of AI orders of magnitude worse.

That is not to mention that they will be burning vast amounts of fuel and burning up more than 5 huge satellites a day in the upper atmosphere, which would cause vastly more pollution and damage to the ozone layer than described in Global 3D rocket launch and re-entry air pollutant and CO2 emissions at the onset of the megaconstellation era by Barker et al or Near-future rocket launches could slow ozone recovery by Revell et al.

Thiele et al Fig. 2

But even that isn't the biggest problem. In An Orbital House of Cards: Frequent Megaconstellation Close Conjunctions Thiele et al show that we are close to the Kessler syndrome:

Here we propose a new metric, the CRASH Clock, that measures such stress in terms of the time it takes for a catastrophic collision to occur if there are no collision avoidance manoeuvres or there is a severe loss in situational awareness. Our calculations show the CRASH Clock is currently 2.8 days, which suggests there is now little time to recover from a wide-spread disruptive event, such as a solar storm. This is in stark contrast to the pre-megaconstellation era: in 2018, the CRASH Clock was 121 days.

Co-author Prof. Lawler wrote:

One of the scariest parts of this project was learning more about Starlink's orbital operations. I had always assumed they had some kind of clever configuration of the satellites in the orbital shell that minimized conjunctions, and we would see the number of conjunctions grow over time in our simulations. But no! It's just random! There's no magic here, it's just avoiding collisions by moving a Starlink satellite every 2 minutes. This is bad.

Even if we haven't rendered low earth orbit unusable in the next few years, a constellation of 10,000 huge satellites in low earth orbit would rapidly self-destruct and guarantee humanity lost access to space.

What are these investors smoking?