Remember the can of Nutella that flashed

on camera during the Aremis 2 mission? A

$5 500 g jar of chocolate spread? To get

that thing into lunar orbit cost roughly

$75,926.

Thank you very much the author of this

post for the calculation. And I agree

with this comment. I really hope it was

worth every penny for the crew. But if a

single jar of Nutella costs 76 grand

just to hitch a ride around the moon,

how did we get to the point when we're

seriously saying, "Yeah, let's put data

centers in space." data centers with

thousands of chips, power systems,

shielding, radiators, networking gear,

tons of metal, all launched into the

orbit before we figured out how to make

data centers work sustainably on Earth.

All you need to know about the economics

of space data centers, why we're getting

them into space in the first place, and

whether it's an economically viable

business to begin with in the new

episode of the series, The Business of

AI. Let's dive in.

Terrestrial AI data centers need three

things in gigantic quantities. Water,

electricity, and land. All three are now

critically constrained. Starting with

water. A typical large data center needs

about 300,000 gallons of water per day

for cooling. This amount of water, for

context, can suffice for about a

thousand households. The largest data

centers for hyperscalers at Google can

use up to 5 million gallons of water per

day. The ones around Phoenix, Arizona

use 385 million gallons per year for

cooling. And this is not counting the

water required to generate the

electricity. As AI data centers grow,

water used for cooling is projected to

skyrocket by 870%.

The second thing is electricity. AI data

centers are hitting similar hard limits

on the power side. The case of Northern

Virginia is probably the most staggering

case of the system reaching its limit.

Northern Virginia hosts nearly 600 data

centers and handles approximately 70% of

the world's internet traffic. A voltage

fluctuation in 2024 almost triggered a

regional blackout and caused 60 data

centers to disconnect from the power

grid. Local utilities and regulators

warned that without massive new

transmission lines, residential

electricity bills in the region can

climb toward $380 per month because

households will be competing with data

centers for the same power. And number

three, land. This specific bottleneck is

simply about finding more physical space

to put all this hardware in. To house

tons of metal, you need a lot of real

estate, acres per site. Most regions

that are best suited for data centers

with cool climates and stable geology,

proximity to fiber routes and power

plants have already been used. and local

governments that once competed to

attract data centers with tax breaks,

for example, are hitting pause because

farmland is paved over, housing projects

are stopped and residents push back

against sprawling industrial jungles on

their doorstep. In more and more

jurisdictions, the cheapest land that

also has grid, water rights, and

community approval simply no longer

exists at the scale that AI builders

wanted to. And so people started looking

elsewhere and that elsewhere is off the

planet earth. They looked down and saw

the water which yes is a thing as well.

We're going to talk about it a little

bit later but today is about looking up

the space. Why put a data center in

space? Let's come back to the three

things that data centers need in

gigantic quantities. Water. Space is the

only place where water supply is not

part of the equation. In orbit, heat can

be rejected directly into the vacuum

through large radiator panels. This

doesn't make cooling easy because those

radiators must be huge and very

carefully engineered, but instead it

breaks the dependency on household or

farmers fighting for the same rivers and

water sources. electricity. Solar power

in orbit doesn't care about heat waves

or winter storms or regional power

grids. A satellite in orbit sees the sun

for roughly 95% of the time, regardless

of clouds, day or night, or atmospheric

losses. What this means is that you get

far more usable energy out of solar

panels in space rather than on the

ground. And again, this also removes the

competition for power plants with

factories or homes or farms or electric

vehicles for a share of an already

strained grid and land. Yeah, it's kind

of obvious that land is unlimited. But

what's even more important is the

absence of politics around the land. On

Earth, every new data center is a

political fight over zoning, noise,

views, jobs, tax breaks. But in space,

you don't need acres of flat and cool

fiber cooling and power adjacent land.

So let's see what it takes to build an

orbital data center step by step.

You're going to need four primary

components. Component number one, the

satellite. To build a flying data center

with its own power generation, computing

cluster, cooling system, and

communications network, you need a

physical piece of equipment that would

house all of it, a satellite. In each

satellite, just like in a terrestrial

data center, we're going to store

clusters of AI chips. However, there's a

problem. If we send the chips that are

currently being produced for terrestrial

data centers, we will inevitably run

into several problems. Problem number

one, cosmic radiation causes bit flips.

On Earth, the atmosphere and magnetic

field block the vast majority of high

energy cosmic rays and solar particles

in orbit above 500 km, which is the

minimum altitude for Space X satellites.

As an example, there is no such

shielding. High energy particles fly

straight into the chip circuiting. When

a charged particle hits the memory cell,

it can flip a bit from 0ero to one or

one to zero. Here is an example. Imagine

a chip storing the number 22 in binary.

Then a cosmic ray hits the chip and

flips a single bit from 0 to 1. And

suddenly the number isn't 22 anymore.

It's 150. This is the bit flip. One tiny

change at the hardware level can cause

completely different data. And this

error is going to propagate through the

entire neural network without triggering

any visible flag. For example, Meta's

training of Lama 3 on Nvidia H100's

experienced 419 unexpected interruptions

in 54 days on Earth. And these failures

were happening every few minutes in a

fully protected terrestrial environment.

In orbit, this system would face all

those failures plus radiation induced

bit flips. So, you may ask, why don't we

just use radiation hardened chips?

Radiation hardened chips do exist, and

they're typically used in military

satellites, nuclear power plants, space

probes, but they're typically much

larger and a lot older. They're more

resistant to radiation by nature because

the transistors are physically bigger

and harder for a particle to hit, but

instead, they lag generations behind

leading AI chips in compute power. The

latest, the most advanced Radhar chips

contain between 500 million and two

billion transistors. In contrast,

Nvidia's flagship H100's contain 80

billion transistors and Blackwell chips

around 200 billion. The solution to this

would be a combination of triple modular

redundancy, running at least three

parallel computation pipelines and

comparing outputs to catch errors,

radiation tolerant wrappers around the

chip, error correcting memory at every

level, and new firmware designed

specifically for high radiation

environments. Triple redundancy means

launching three times as many chips,

three times the cooling, and three times

the launch cost. Meaning the economics

of this simply will not work. Google in

fact tested their trillium TBU chips on

the ground and projected only five years

of reliable orbital operation. But the

problem is that the ground radiation

exposes chips to only one particle type

while space has continuous hail of

cosmic rays at a far wider range of

energies. The second big problem is that

chips wear out and cannot be replaced.

On Earth, data centers swap out GPUs

every two to three years as those chips

degrade from use or as newer generations

arrive. In orbit, you can't swap out the

chips. You're not going to hop on a

rocket and magically fly up to the

satellite and take them out. Chips do

not get brought back. It's a one-way

trip. They run there for about 5 years

until solar panels and chips start to

degrade and then burn up. There's no

retrieval of satellites back to Earth.

In contrast, for terrestrial data

centers, old chips get a second and

sometimes third life. They can be

cascaded to lighter loads. They can be

sold to smaller cloud providers. The

secondhand market for chips is actually

quite big. But in space, recovering and

refurbishing those chips would require a

whole additional class of vehicles like

space tugs or cargo return capsules,

which nobody has at scale yet. Moving on

to the second component, power. A solar

plant on Earth receives sunlight

filtered through the atmosphere. Only

when the sun is up and only when there

are no clouds. In space, the exposure to

the sun is approximately 95% compared to

roughly 30% for a solar panel on Earth.

And it all sounds fine and dandy, but

the risk here is that generating power

in space requires very large systems of

solar panels to feed a sizable chip

cluster. We're talking solar structures

on the order of a football field folded

inside a rocket and then somehow

autonomously unfolding into 170 m wide

wing. These structures should also be

able to survive thousands of cycles

between minus 120 and plus 120° Celsius

temperature fluctuation. Component

number three, cooling. On Earth, data

centers keep chips cool by pushing heat

into air or water. In space, there is

neither air nor water. So, that trick

doesn't work. The only option is to let

the hardware dump heat as in frared

light into empty space using big

radiator panels. The problem is the

size. To get rid of heat from a data

center, those panels have to be massive

and work for several years with no

maintenance or somebody being there to

fix them. As you add more compute, those

panels have to get larger and larger.

And component number four, networking.

For a cluster of satellites to function

as a distributed data center, they need

to talk to each other and they need to

send results back down to Earth. This

satellite to satellite problem is solved

with laser links that are quite

literally shooting laser beams between

the satellites. However, for those

lasers to carry enough bandwidth for AI

computation, the satellites must fly in

formation from hundreds of meters to a

few kilometers apart to stay in each

other's area of visibility. That means

that every satellite must hold its exact

position relative to its neighbors or be

able to correct its position

autonomously to maintain connectivity.

Everything about data

centers in space lives or dies on how

cheap it is to put each kilogram of

hardware into low Earth orbit. It's the

cost of the launch that determines

whether the effort to launch orbital

data centers can compete with

terrestrial alternatives. Whatever you

do to solve the radiation and cooling

and power, you still end up with a ton

of stuff that needs to be carried to the

satellite. Chips, shielding, radiators,

structures, propellants. The heavier the

load, the more fuel the rocket burns,

the more expensive the flight. To get it

into space, you have to pay for every

kilogram that you put into the orbit,

regardless of whether that kilogram is a

gold chip or a steel bolt. To deliver

the chips to the orbit, you need three

things: a rocket, a satellite, and

payload. Payload is the actual cargo,

the chips, the radiators, the coolers

that go on the satellite. The rocket is

basically a delivery truck. It burns

fuel to accelerate whatever it is

carrying to up to the orbital speed. The

rocket has multiple stages and each

stage fires, runs out of fuel and gets

dropped. Only the top stage plus the

payload make it to the orbit. The

satellite is what will stay in orbit

once the rocket is done. The satellite

carries the structures, propulsion,

power, cooling, communications, and

compute hardware. By compute hardware,

we're talking racks of GPUs or TPUs or

any other kind of chip. When someone

says put H100s in space, they literally

mean bolting H100 boards inside a

satellite and putting that satellite on

a rocket. So the rocket carries the

satellite, the satellite carries the

chips, the chips sit on boards inside

the satellite. Now who supplies the

rocket? The rocket company. SpaceX, Blue

Origin, United Launch Alliance mainly

for US military and government or area

and space in Europe. They sell you

kilograms to orbit. Now, for the

economics of this to work in favor of

data centers, the cost per kilogram has

to go down to about $200.

Why $200? Because this number is roughly

the point where the cost of lodge stops

overpowering the cost of compute. And

the math starts to look similar to a

good terrestrial data center. As of

today, SpaceX is the only one openly

targeting $200 per kilogram goal with

more or less promising plan on how

they're going to get to that number. But

again, it's a future projection. At

today's launch prices, space power costs

five to 20 times more than the ground.

And the cost per kilogram sits around

$3,000 for the cheapest widely used

rocket. Who supplies the satellites?

StarCloud, Axiom, ADA Space. These guys

decide what the chips are going to

physically sit in. They design the

chassis, panels, radiators, thrusters,

and build the actual satellite with

those things. So, how much does a

satellite that would be compatible with

AI hardware cost? There's no publicly

disclosed number, but StarCloud raised

$3 million precede and with that, they

designed, built, and launched StarCloud

1 in 21 months. And the $3 million

figure well covered the engineering and

manufacturing costs. So, speaking of how

much it costs to build a satellite like

that, we're talking in the ballpark of

$2 million for a StarCloud class

satellite. But on top of that, add the

GPU costs. Speaking of chips, it is

increasingly becoming in the interest of

semiconductor manufacturers to start

producing radiation hardened chips

because it's a massive growing market.

Project Suncatcher is Google's attempt

to step into payload territory. They

design the rad hardened chips, their own

TPUs, and the space compute concept. And

then they partner with launch providers

that actually put that hardware into

space. And this is what the cost of a

space data center more or less comes

down to. So to sum it up, are we about

to see AI infra getting into space? Yes.

At first, it's going to be for AI

inference and then probably within the

next 10 years, we will most likely see

AI training as well. But as it stands,

it is not a justifiable business. You

can see the numbers. The math ain't

mathing. So the industry is not going to

blow up overnight. But it is something

that we're actively marching towards.

Another big reason why getting into

space would be a huge step towards

decentralization of AI is because

putting data centers in space would in

theory give countries that are unable to

run AI workloads an opportunity to build

their own AI tech market. AI

infrastructure crisis in the global

north is defined by 7 years of grid

cues. In the global south, the problem

is either unreliable power or no power

at all or power too expensive to run

compute heavy workloads. As a result,

we're dealing with what researchers call

compute deserts, countries where there

is no public cloud AI hardware at all,

compared with compute north, the US,

China, and the EU, where GPU data

centers live, and compute south, which

is most of the developing world, where

you can use but not train AI, and where

you pay US cloud providers the dollar

prices on local wages. Space-based AI

compute would change the situation, but

again, it would do so in theory. A

country with no functional grid, but sun

overhead could theoretically access AI

inference from an orbital data center,

beam results down, and bypass the entire

landbased infrastructure. So, let's see

where this goes. As always, we hope this

was helpful. We'll see you in the next

episode of the business of AI. Bye.