AI Moon. Is Mr. Musk's lunar project pure lunacy?

Last week The New York Times published the story “Elon Musk Wants to Build an A.I. Satellite Factory on the Moon.”

The reporters shared with us the speech by Mr. Musk given to employees at his xAI company. In that speech, he shared his vision for the future where the recently acquired xAI by his SpaceX company will reach out to the stars.

Actually, before the stars, it will be the Moon. One can argue that it is a scaled back ambition because originally SpaceX was meant to ferry people to Mars. Somehow along the way this project got a little off the rails. Aside from all the technical things to be resolved, it would be a one way trip for the foreseeable future. But most importantly, SpaceX still doesn't have the spaceship ready to get us there despite the ambitious plan to build one spaceship a day.

Now the highlights from this new announcement.

Here are the plans:

• There will be AI data centers in outer space. That's the rationale for why SpaceX bought xAI.

• There will be a self-sustaining (e.g. able to continue in a healthy state without outside assistance) city on the Moon.

• There will be a factory on the moon to build AI satellites.

• These satellites will be launched into space using a catapult.

Rather than arguing whether any of it makes any sense, or if it is a vision or delusion, I prefer to use the What Would Have to Be True (WWHTBT) question, a question which is part of the Strategy framework by Prof. Roger Martin.

The terms 'AI data center', 'self-sustaining city' and 'factory building AI satellites' are a little bit ambiguous, but let's start with 'AI satellite'. Why? Because in the recent filing with FCC, SpaceX applied for permission - and it was granted - to launch 1 million satellites as orbital data centers. Fortunately for SpaceX there is no time limit to do that.

We don't know how big and heavy these AI data centers will be but we can use the current Starlink satellite as a reference point. Originally SpaceX launched 60 satellites per launch but as the new generation of Starlink satellites got heavier, the number of satellites per launch decreased to 28. The new Starship, once operational, should launch 100 satellites per flight. That will take us, if there is one rocket launch per day, about 27 years. Currently SpaceX has permission for 25 launches a year and according to the COO of SpaceX, they would like to get to over 4 years to the total of 400.

Now that we have some idea about the scope of what Mr. Musk is proposing, we can now check the WWHTBT to make this happen on and from the moon.

You might remember my post 'Is this SciFi enough?' from last year, where I discussed in length the economics of launching data centers to space and the perils of operating them there. Let's just say it is a badly constructed disguise to move things outside of any government control.

Now, let's focus on the moon.

In order to do anything on the moon, we need energy. Here we are lucky, since Blue Origin (a company founded by Jeff Bezos) has figured out how to manufacture solar cells on the Moon with its Blue Alchemist project. NASA did a similar study with a conclusion that the production of solar cells will become economically beneficial once we are in the MW range.

So this part is plausible.

Once we have energy we can start mining minerals, start construction and get the AI satellite factory going.

Here is a document from Intel 'What does it take to build a fab?' - 'fab' as in a semiconductor factory. This factory includes 1,200 multimillion dollar tools, 1,500 pieces of utility equipment and it takes $10 billion and 7,000 workers to complete. And that's just for the chips.

BTW: One of the devices which you have to have for any chip production is the EUV lithography system made by ASML. To install one of these, it is a months-long, carefully choreographed assembly process of 100,000+ parts, including a 180-metric-ton machine that ships in 40 freight containers.

In order to have an AI chip factory on the moon, this one is non-negotiable.

Also to build an AI chip, you need the following:

• Basic semiconductor materials: Silicon, which forms the foundation of most chips.

• Conductor and connection materials: Copper, silver, gold.

• Critical minerals for advanced functionality: Tantalum, tin, cobalt, gallium, germanium.

• Rare Earth Elements (REEs): Including neodymium, dysprosium, and terbium for high-performance magnets used in AI hardware.

• Other essential minerals: Bismuth, antimony, nickel, palladium, and boron.

And then you need Process Gases - Neon, Helium, Arsine, Krypton and Xenon, Tungsten, Hexafluoride and other chemicals. You can learn about the making of AI Chips here.

We know how to make it on Earth, but none of this was done on the Moon.

Now that we know how difficult it is, the WWHTBT for the AI chips, we still have to build the satellites and launch them.

What is with the catapult launcher?

One option is the mass driver which would use a linear motor to accelerate and catapult payloads to outer space. Let's say that we have all the required electricity available and we can deliver it during the launch and we know how to cool it, what would be the design?

Here we have competing factors between the length of the rail and the G-force which the equipment can sustain. When we send people to space, the G-force is between 3 - 4 Gs to protect the humans. Equipment - satellites, scientific payload - are tested to withstand 10 Gs.

For a payload of 1,000 kg (800 kg satellite + 200kg magnetic sabot), to achieve a moon escape velocity, the track would have to be almost 30 km in length. And then you have to build all the infrastructure and cooling. Just the cooling will double the mass of the system. The 800kg is the mass of a v2 mini Starlink satellite. For comparison a fully loaded Nvidia GB200 NVL72 rack (compute and switch shelves) is 1,360 kg. That would increase the peak power requirement from 291 MW to 495 MW.

Another form of catapult is using the centrifugal force to launch to space is in development on Earth, but it is still in very early stages.

To recap the vision for SpaceX:

• There will be AI data centers in outer space - using cheap electricity as a reason doesn't add up. It is to do things away from any oversight.

• There will be a self-sustaining city on the Moon - with enough energy, there is a lot you can build, but you will still rely on regular shipments from Earth.

• There will be a factory on the moon to build AI satellites - an exciting statement with constraints beyond SpaceX capabilities.

• These satellites will be launched into space using a catapult - once you solve a 'few technical details.’

Is this a vision of what we could achieve in the distant future? Then it sounds good to me. If this is a pitch to potential investors as a near term project, then there will be lots of disappointment.

To follow the What Would Have To Be True question, let's first start with a spaceship which can reliably take off and deliver the maximum payload to the moon. 14 years of development and we are still waiting. Having a truly self-driving car roaming the city streets would add legitimacy to all the other ideas.

The recurrent pattern? When helping companies with aligning the technology with the strategy, I have found the What Would Have To Be True question really grounding with a good dose of reality. It helps to create a roadmap between now and the future. It helps you to differentiate between vision and delusion.