Orbital AI Will Not Replace Terrestrial Data Centers

Executive Summary

Instead of building increasingly large terrestrial data centers, the orbital compute vision places AI infrastructure directly into orbit, powered by solar energy and cooled through thermal radiation into space.

The concept is serious, but it is often discussed with the wrong comparison.

A 120 kW AI satellite is not equivalent to a modern 2 MW terrestrial AI container. While orbital systems may become exceptionally valuable for processing data already generated in space, global edge inference and strategic resilience, Earth-based data centers remain fundamentally superior for large-scale AI training, hardware upgrades, maintenance and tightly coupled GPU clusters.

The future is therefore unlikely to be Earth versus Space.

It is far more likely to become Earth plus Space: two complementary layers of one global computing infrastructure.

1. The Vision Behind Orbital Compute

The central idea is straightforward.

Traditional AI infrastructure faces three increasingly difficult constraints:

• available land
• electrical power
• cooling

Large AI clusters now require hundreds of megawatts and, in some cases, gigawatts of continuous electrical power. Obtaining permits, grid connections and water rights is becoming slower and more expensive.

Orbital computing attempts to remove these bottlenecks by moving computation into Low Earth Orbit, where solar energy is abundant and heat can be rejected through thermal radiation instead of water-based cooling.

At first glance, the concept appears revolutionary.

In reality, it replaces terrestrial infrastructure challenges with orbital engineering challenges.

2. Electrical Power Does Not Equal Compute

One of the most misunderstood numbers in recent discussions is:

Actual AI throughput depends on:

• accelerator architecture
• memory bandwidth
• interconnect topology
• numerical precision
• workload efficiency
• software stack
• redundancy
• radiation mitigation

A modern AI accelerator typically consumes approximately one kilowatt of total system power.

Therefore, a 120 kW satellite might realistically host on the order of one hundred AI accelerators.

By comparison, a modern 2 MW terrestrial AI container with an efficient PUE provides approximately 1.7 MW of IT power, enough for roughly 1,500 to 2,000 accelerators, or the equivalent IT capacity of around 15 orbital satellites.

Calculation

2 MW terrestrial container versus 120 kW orbital satellite

$$P_{IT}=2MW/1.2=1.67MW$$

Assumptions

• Terrestrial facility power is 2 MW.
• Illustrative PUE is 1.2.
• Orbital satellite electrical input is 120 kW.
• One accelerator-class system is approximated at 1 kW total system power.

A 2 MW terrestrial AI container can carry roughly 14 times the IT power of a 120 kW orbital compute satellite.

Electrical power alone is not an adequate measure of computational capability, but it shows the scale gap clearly.

Electrical power alone is therefore not an adequate measure of computational capability.

3. Space Does Not Provide Free Cooling

One of the most common misconceptions is that vacuum somehow makes cooling easy.

The opposite is true.

On Earth, heat is removed by:

• convection
• forced airflow
• liquid cooling
• cooling towers
• rivers
• seawater
• ambient air

None of these mechanisms exist in space.

The only way to remove heat is through thermal radiation, governed by the Stefan-Boltzmann equation:

Calculation

Radiative heat rejection

$$P=\epsilon\sigma AT^4$$

Assumptions

• P is radiated power.
• epsilon is emissivity.
• sigma is the Stefan-Boltzmann constant.
• A is radiator area.
• T is absolute temperature.

Every watt consumed by AI hardware must ultimately leave the spacecraft through radiator panels.

The challenge is not only cooling chips. It is moving heat to large radiating surfaces while maintaining spacecraft stability, reliability and orientation.

Space replaces cooling towers with radiators spanning hundreds of square meters.

Physics does not disappear.

It simply changes form.

4. Solar Power Is Excellent, But Not Infinite

Outside Earth's atmosphere, solar irradiance reaches approximately 1,360 W/m².

This creates an enormous opportunity.

However, solar panels convert only a fraction of incoming energy into electricity.

Generating approximately 150 kW continuously requires several hundred square meters of photovoltaic area, together with batteries, power electronics and deployment mechanisms.

Every additional kilowatt requires:

• larger solar arrays
• larger radiators
• more structure
• greater launch mass
• increased cost

Energy remains fundamentally linked to mass.

No engineering design escapes this relationship.

5. Latency: Physics Still Wins

The speed of light defines the absolute lower limit for communication.

Signals traveling from Earth to a satellite and back inevitably require milliseconds, even before routing, switching and processing are considered.

Nearby terrestrial data centers often respond faster.

Orbital laser networks become advantageous primarily for:

• intercontinental routing
• remote regions
• space assets
• globally distributed edge applications

For local AI services, terrestrial fiber remains difficult to beat.

6. AI Training Depends on Communication

Modern frontier AI models do not simply require many GPUs.

They require thousands of GPUs communicating continuously.

Gradient synchronization occurs millions of times during training.

Within terrestrial AI clusters, communication latency is measured in microseconds.

Orbital constellations introduce millisecond-scale communication delays.

This difference, roughly three orders of magnitude, fundamentally changes distributed training efficiency.

Orbital compute therefore appears significantly better suited for:

• inference
• independent workloads
• Earth observation
• defense applications
• weather processing
• navigation
• autonomous decision support

rather than tightly synchronized AI training.

7. Hardware Ages Faster Than Satellites

One challenge receives surprisingly little attention.

AI hardware evolves extraordinarily quickly.

A terrestrial data center upgrades:

• GPUs
• networking
• memory
• storage
• cooling

every few years.

Satellites cannot.

Once launched, their hardware remains essentially fixed throughout operational life.

If accelerator efficiency doubles every several years, orbital hardware inevitably becomes economically obsolete long before the spacecraft reaches end-of-life.

Orbital AI therefore requires continuous replacement launches.

The infrastructure itself becomes a permanent manufacturing cycle.

8. Reliability Is More Complex Than On Earth

Terrestrial infrastructure benefits from technicians, spare parts and immediate repairs.

Orbital infrastructure does not.

Space introduces entirely different risks:

• radiation-induced memory errors
• thermal cycling
• micrometeoroids
• orbital debris
• pump failures
• deployment failures
• attitude control limitations

Every additional redundancy increases:

• mass
• cost
• power consumption

Space engineering therefore becomes an exercise in balancing reliability against launch economics.

9. Comparing Orbit with Earth

Suppose we deploy a modern 2 MW AI container in:

• the Sahara with dedicated solar and batteries
• Norway with hydroelectric power
• Iceland with geothermal energy

Each location offers:

• hardware replacement
• scalable fiber connectivity
• modular expansion
• straightforward maintenance
• continuously improving AI hardware

Orbital infrastructure offers:

• global reach
• resilience
• direct access to space-generated data
• independence from local grids

The comparison is therefore not one of superiority.

It is one of specialization.

Different environments optimize different workloads.

10. Bandwidth and Hardware Refresh Decide the Boundary

Raw satellite downlink has improved dramatically. That does not make orbit the natural place for massive Earth-originated training data movement. Frontier training consumes enormous datasets, checkpoint movement, model synchronization and operational telemetry. Moving that workflow into orbit can add a logistics problem unless the data itself is already generated in space.

This is where the economics become clearer. If most data originates on Earth and most users are on Earth, orbital compute must justify the extra communications layer. If the data originates in orbit, such as imagery, signals intelligence, weather sensing or spacecraft telemetry, local processing becomes far more compelling.

Chip obsolescence reinforces the same conclusion. AI accelerators evolve faster than most space platforms. A terrestrial operator can replace racks, add switches, change cooling loops and adopt new accelerator generations. A satellite operator must live with launch schedules, qualification cycles and orbital lifetime. For orbital inference, that may be acceptable. For frontier training, it is a structural disadvantage.

11. The Future Is Hybrid

The discussion should never become:

Space versus Earth.

Instead, it should become:

Which workload belongs where?

The emerging computing architecture may ultimately consist of three layers:

Terrestrial hyperscale

Training frontier AI models in dense, maintainable, upgradeable data centers with high-bandwidth internal networks.

Regional edge infrastructure

Serving users with minimal latency and local operational control.

Orbital compute

Processing space-native data, supporting autonomous systems and extending global computational resilience.

This layered architecture appears significantly more realistic than replacing terrestrial AI infrastructure altogether.

Why this matters

Investors, operators and governments should evaluate orbital compute against the workloads it can structurally improve, not against the fantasy that it eliminates terrestrial AI infrastructure. The commercial opportunity is in integration, not replacement.

My conclusion

Space-based AI represents one of the most exciting infrastructure developments of the coming decade. But engineering is ultimately governed by physics rather than marketing. Electrical power is not compute. Vacuum does not provide free cooling. Solar energy is abundant but requires enormous structures. Latency cannot exceed the speed of light. Bandwidth remains valuable. Hardware becomes obsolete. Large AI models still depend on extremely dense communication. Rather than replacing terrestrial data centers, orbital compute is likely to become a powerful complementary layer within a broader global computing ecosystem. The future belongs neither exclusively to Earth nor to orbit. It belongs to the intelligent integration of both.