The AI Grid Is Leaving Earth

In 1900, if you wanted to power a factory, you built your own generator.

You hired engineers to design it and bought fuel to feed it. You built your entire operation around the assumption that power was something you produced yourself, on-site, because there was no other option.

Then the grid arrived. The constraint – not the technology, the capital, or the ambition – disappeared overnight. What followed was an entirely new industrial era.

AI is living in its ‘generator moment’ right now. And the grid that’s coming isn’t on the ground.

It’s in orbit.

Orbital compute – AI data centers launched into low Earth orbit, powered by limitless solar energy, cooled by the void of space, beaming processed results back to Earth via laser link – is already being built. And the investment opportunity taking shape around it may be the largest we’ve seen in a generation.

What’s driving it comes down to a constraint the market hasn’t fully priced in yet.

The Real Bottleneck: Land, Power, and Water

The numbers are staggering; and they’re not the problem.

Gartner put 2025 AI infrastructure spending at $965 billion and sees it growing to $1.37 trillion in 2026. McKinsey estimates $7 trillion in total data center investment will be needed by 2030 just to keep pace with projected demand.

Contrary to what those astronomical numbers might have some thinking, the bottleneck here isn’t capital, or even hardware supply. These companies have the money, and securing the chips isn’t a problem. 

The actual binding constraint is something far more intractable: Land. Water. Power.

Prime real estate for data centers – places with cheap power, cool climates, fast fiber – is being consumed.

Water rights for cooling are drawing regulatory scrutiny and community opposition across the American West.

And grid utilities can’t connect new data centers fast enough. Interconnection queues now stretch three to five years in the U.S. Bloomberg estimates that nearly half of all AI data center projects in the U.S. will be delayed this year due to power constraints.

Microsoft (MSFT), Alphabet (GOOGL), and Amazon (AMZN) have all found themselves in the absurd position of having the chips and the capital, yet not being able to get a power connection in time.

These are resource problems. 

The solution has to come from outside Earth’s resource constraints entirely – which is precisely why the tech world is now looking to space for answers.

Why Orbital Data Centers Solve the Problem

Space has three things Earth’s data centers are running out of: energy, cooling capacity, and proximity to where the data originates.

Solar panels in low Earth orbit receive 1,400 watts per square meter of raw energy when in direct sunlight. At Earth’s surface – with ideal conditions – peak sunlight reaches roughly 1,000 W/m², but atmospheric losses, day-night cycles, and weather reduce that average greatly.

Solar panels in space also turn that stronger sunlight into a much steadier stream of power. In low Earth orbit, high-efficiency arrays can produce around 400 watts per square meter in direct sunlight and still average north of 200 watts over time thanks to near-constant exposure. Back on Earth, even top-tier solar farms usually average just 20 to 60 watts per square meter once nightfall, clouds, and shifting sun angles take their toll. That combination of higher intensity and near-continuous exposure gives space-based solar a structural advantage no terrestrial installation can match.

The energy advantage is only half the story above the atmosphere.

Space is a near-perfect vacuum, sitting just a few degrees above absolute zero. That means heat can be dumped straight off of GPUs and into the void through large radiator panels; no fans, water, or intensive cooling infrastructure required. Meanwhile, terrestrial data centers burn through billions of gallons of water each year just to keep systems from overheating. It’s a fundamentally different setup in orbit – one that becomes more important as water constraints tighten and cooling turns into a real bottleneck on Earth.

Then there’s an underappreciated third advantage – one that may become commercially significant faster than anyone expects: data proximity.

A staggering volume of AI workloads analyze satellite imagery, defense surveillance data, weather telemetry, and maritime tracking information – all of which originates in space. Today that data is sent down to Earth and processed in ground-based data centers before results are distributed. The downlink is a bandwidth bottleneck. Move compute closer to the source, and you cut that constraint dramatically. For an Earth observation satellite, it’s the difference between transmitting 100 gigabytes of raw imagery and just 1 megabyte of actionable intelligence. That advantage is available today, with current technology, at current launch costs.

Orbital Compute Is Already Here

This might sound fascinating but distant – years away from mattering to a portfolio.

Consider what has happened in just the last six months.

• November 2025: Starcloud launches a satellite containing an Nvidia (NVDA) H100 GPU – the first chip of that class ever sent into orbit.
• December 2025: That satellite trains NanoGPT on the complete works of Shakespeare – the first time a language model has been trained in space.
• January 2026: SpaceX files with the FCC for authorization to launch up to 1 million satellites as orbital AI data centers. Not a press release – a regulatory filing.
• February 2026: SpaceX acquires xAI in the largest private merger to date, valuing the combined entity at $1.25 trillion. Orbital compute is the stated rationale.
• March 2026: Nvidia announces the Vera Rubin Space-1 module at its GPU Technology Conference. The module is a chip platform purpose-built for orbital data centers. During his keynote, Jensen Huang said: “Space computing, the final frontier, has arrived.”
• April 2026: SpaceX confidentially files with the SEC for an IPO targeting a $1.75 trillion valuation – the largest in market history – with orbital compute as the central thesis.

The Economics of Space-Based AI Infrastructure

Now for the numbers… because this is what has folks so skeptical.

Today, in a state-of-the-art terrestrial data center, running one H100-equivalent GPU for one hour costs roughly $1.00. Hardware amortization, electricity, cooling, facility costs, network, labor – all in, about a dollar. That’s what the hyperscalers pay before they mark it up for cloud customers.

In orbit today, that same GPU-hour costs approximately $142 – a 142x premium.

But that $142 breaks down in a very specific way: $85.62 of it is pure launch cost. That’s 60 cents of every orbital dollar going toward getting hardware off the ground. The “free” solar energy of space costs $22.83 per hour to access because you have to deploy the solar panels in orbit first. In other words, the energy itself is free; the infrastructure to harvest it is not.

The entire orbital compute thesis is a bet on one number declining: dollars per kilogram to low Earth orbit.

Starship – SpaceX’s next-generation reusable rocket – is designed to bring that cost from ~$3,000/kg today to somewhere in the $50-100/kg range at scale. Google’s own engineers published a feasibility study concluding that at $200/kg, orbital compute becomes cost-competitive with terrestrial. The trajectory of launch cost reduction is real. Costs have fallen roughly 10x in 15 years. 

The question is whether Starship crosses the final threshold.

Our analysis suggests the economic crossover happens around 2038. At that point, orbital compute flips from being the premium option to being the low-cost option for AI inference. 

How? Orbital compute costs are technology-bound: subject to Wright’s Law and learning curves. They reliably fall. Terrestrial compute costs are increasingly resource-bound: subject to physical scarcity in power, land, and water. They reliably rise. Both forces are already in motion. Neither is reversing. The lines cross around 2038 – and then they keep diverging.

Who Adopts Orbital Data Centers First

If orbital compute is still more expensive than terrestrial through the early 2030s, how does this market scale before the economics fully flip?

The answer comes down to urgency. Several buyer groups already have strong reasons to move early, even without perfect cost alignment.

Defense and sovereign systems. Next-generation missile defense and space-based sensing programs already run into the hundreds of billions. These architectures depend on real-time data processing, and pushing everything through ground infrastructure introduces latency and operational constraints that don’t always work. In certain scenarios, keeping compute in orbit becomes a practical requirement.

Power-constrained regions. Energy costs in Europe and parts of Asia remain structurally higher than in the U.S., putting pressure on large-scale compute economics. That dynamic shifts the timeline, making alternative energy models more attractive sooner for some operators.

Carbon pressure. Hyperscalers have made aggressive decarbonization commitments, while policy frameworks continue tightening. As carbon accounting evolves and clean energy becomes more contested, access to reliable power starts to shape infrastructure decisions in a more direct way.

Strategic positioning. Infrastructure transitions tend to reward early movers. Companies that built cloud capacity ahead of demand ended up owning the stack. A similar dynamic could play out here, where establishing orbital capacity early provides long-term leverage.

Space-native workloads. A growing set of applications – real-time Earth observation, autonomous satellite operations, on-orbit systems – benefits from having compute close to the source. In some cases, performance drops off meaningfully without it.

The economics don’t have to be perfect for a real market to form. Strategic necessity has a way of moving faster than spreadsheets.