Finding Space. [3/4]

“An economist is an expert who will know tomorrow why the things he predicted yesterday didn't happen today.” - Laurence J. Peter

If you know this topic well and notice I've made a mistake or overlooked something (which is likely), leave a comment and let me know. We can learn together!

In order to have an apples to apples comparison, we’re using “Levelized Cost”. If you aren’t familiar, it calculates the unit cost over a finite time ($/_MWh) which lets us compare costs while accounting for lifetime and project scale (click the link if you’re curious). These are the primary cost assumptions:

This gives us a total cost to beat of $2,285/_MWh.

While highly simplified, we assume any orbital data center can be split into four primary costs:

We mentioned last week that mega-satellites are likely the best solution since they avoid the “slow” east-west communication involved with splitting compute across many smaller satellites. The easiest way to find our levelized cost is to find the total cost for a certain capacity (I chose 1GW) and divide this by the estimated lifetime (MWh) of the system.

Starlink satellites have an estimated life of 5 years and this is conveniently also around the lifetime of modern GPU’s. Given this, 5 years is good enough for us and we will assume 5 years as the lifetime of our orbital data center.

Our first cost will be GPU’s. An Nvidia H200 (which contains 8 cores) consumes 10.2 kW and costs roughly $250,000. Since this is space and weight is important, it also weighs just under 300lbs. For our 1 GW data center, this gives us a total dollar value of 25B and a total mass of 13,000 tonnes.

Our second cost is the satellite bus which encompasses the main structure and subsystems of the satellite. As a general rule, computer chips will convert all the energy they consume into heat. So for a 1 GW data center, the GPU’s will generate roughly 1 GW of heat.

As we discussed in the engineering post the only way to dissipate heat in space will be specialized radiative coolers which are neither cheap nor light. Functionally what this means is that at the GW scale our bus costs are roughly equivalent to our radiator costs. Lets assume these specialized radiators are around $10/_W of cooling capacity and for every 1 kg of GPU’s we will need 1.2 kg of radiators. Extrapolating this out we find a total cost of 10B and total mass of 15,000 tonnes.

*These are very rough estimates so take the radiator calculations with a grain of salt

Next we need to determine the cost of our solar panels. Redwire makes some that have a specific power of 100 ^W/_kg and we will assume they cost around $50/_W (space panels cost a lot more than terrestrial ones which are only $0.26/_W). Applying this to our 1GW data center we get a total cost of 50B and mass of 10,000 tonnes.

Tallying everything up and dividing by our lifetime of 5 years we get a levelized cost of $3,245/_MWh

So obviously $3,245 is bigger than $2,285. What does this tell us?

Well it says that with the current price to send things into orbit, from an economics perspective, we are better off building data centers on earth.

Is that it? Should we give up?

Nonsense.

The most logical question to address next is, "What would the launch costs need to be?". If we set the levelized cost for an orbital data center equal to that of a terrestrial data center and do a little rearranging of our initial equation, we get this:

Solving this equation we find that space based data centers start to make sense when launch costs get to around $400/_kg (Google researchers said it was closer to $200/_kg but who’s counting).

The current cheapest launch costs are $1500/_kg aboard Space X’s Falcon 9 Heavy. At more than 3x the required cost, we clearly still have a ways to go. In spite of this, I remain optimistic.

The largest cost in a rocket launch is the rocket itself (duh). While this may not be a be a groundbreaking truth it does highlight something interesting. With Space X’s reusable rockets the cost of the rocket is split across every launch it does. As this technology gets better and the number of times a rocket is reused increases, the cost of a launch actually approaches the only truly consumable item…the cost of fuel. Which is considerably less than $400/_kg. With ever increasing investment into data centers and space technologies I just don’t think we are that far off. Time will tell.

All the math from the post can be found here: (Terrestrial) (Space)

Next week I am going to finish up this series with risks and I can say that after only a little bit of research I am genuinely concerned about potential consequences. Stay tuned.