Data Centers in Orbit: The Next Frontier for AI Infrastructure
- Gregory Chassapis
- May 20
- 4 min read
Here’s what we all seem to take for granted: the entire AI industry runs on the assumption that Earth will keep providing enough power, land, water, and silicon to support every chatbot conversation, every generated image, and every frontier model in training.
That assumption is not as safe as it was before mass adoption.
Despite projected hyperscaler spend of approximately $400-$700 billion in 2026 on terrestrial data centers, the bottleneck remains electricity, the demand for which is estimated to roughly double from 485 TWh in 2025 to 950 TWh by 2030, with AI workloads tripling in that span.
That is the context in which an idea once dismissed as science fiction has become a credible line item in venture funds, sovereign space programs, and the prospectus of what may become the largest IPO in history (SpaceX). While our grid remains an issue, we are now faced with the possibility of having to launch capacity into space to ensure it’s all sustainable.
Why This Matters Now
Three forces are converging. First, AI's energy appetite is colliding with grid reality. In the United States, the IEA projects data centers will consume more electricity by 2030 than the production of aluminum, steel, cement, and chemicals combined. Permitting timelines for new transmission and generation now run five to ten years, whereas AI capex cycles run much quicker than that.
Second, the cost of getting to orbit has collapsed. SpaceX's Falcon 9 brought launch costs from roughly $11,500 per kilogram a decade ago to anywhere between $700- $1,000 per kilogram today. Anything below roughly $500 per kilogram, and the case for orbital compute stops being a thought experiment and instead becomes a spreadsheet exercise. Starship, if it hits its operating cadence, is targeted at $100 to $200 per kilogram.
Third, some of the physics actually favor space. Solar irradiance in orbit is roughly 36% higher than at the surface, with no clouds and no night in the right orbital plane. Best of all, there are no neighbors to object, no aquifers to drain, and no zoning boards. For an industry whose biggest constraint is suddenly the substation, those properties are not minor.
From Slide Deck to Funded Buildout
The capital flow has shifted decisively in the last twelve months. Alphabet's Project Suncatcher aims to launch its first prototype orbital AI satellites by 2027, using Trillium TPUs that have already passed some form of radiation testing (with some caveats). Starcloud launched the most powerful GPU ever flown in orbit in late 2025 and raised a $170 million Series A in March 2026, becoming the fastest unicorn in YC history. Cowboy Space raised $275 million in May 2026 at a $2 billion valuation to build rockets whose upper stages double as orbital data centers. And SpaceX, gearing up for a reported IPO at a $1.5 to $2.0 trillion valuation, is in active talks with Google to build orbital infrastructure together, with Blue Origin pursuing a similar plan.
This is no longer a story about white papers. It is a story about supply chains, offtake agreements, and an emerging stack of investable categories: launch, space-grade power and thermal systems, radiation-hardened electronics, optical inter-satellite links, and in-orbit servicing.
The Sustainability Angle
The cleanest gigawatt is the one you do not have to build on Earth. The IEA expects natural gas and coal together to meet over 40% of incremental data center demand through 2030, with data center-related CO2 emissions peaking near 320 megatons annually before slowly declining. Orbital compute, powered by 24/7 solar in a sun-synchronous orbit, sidesteps that emissions curve entirely for the workloads it captures. It also frees terrestrial renewables to decarbonize buildings, vehicles, and industry rather than chasing AI's runaway demand.
Critics will be quick to point out that launch itself has a carbon and atmospheric footprint and that the lifecycle math at scale is still being worked out. But the directional case (replacing fossil-backed baseload with continuous solar above the clouds) is genuinely compelling, and increasingly so as compute intensity climbs.
The Real Risks
Launch economics remain a crucial variable: if Starship's operational cost-per-kilogram does not fall below roughly $500, the entire thesis likely stays niche.
Then there is orbital debris. Plans for constellations of hundreds of thousands (or millions) of compute satellites raise legitimate concerns, where a cascade of collisions could render entire orbital bands unusable. The FCC is beginning to treat this as a policy matter rather than a theoretical one. That's part of the reason SpaceX's own IPO disclosures warn that orbital AI infrastructure may never become commercially viable.
Finally (and perhaps most importantly), there’s the issue of radiation. Charged particles in orbit can flip bits or permanently damage chips, and AI training tightly couples thousands of GPUs in a way that lets a single failure cascade across the system. Radiation-hardened silicon exists but lags multiple generations behind frontier chips. Working around the problem typically demands triple-redundant hardware, which will inevitably add to hardware costs. In the event degradation is more severe than anticipated, swapping a GPU in space is rather difficult, and in an industry where chips age in three years, that is a structural drag.
What Comes Next
The next twenty-four months will likely produce some of the data points that matter. Including Starship's all-in cost-per-kilogram at operational cadence, Project Suncatcher's 2027 prototypes, and Starcloud's second-generation satellite economics. These are catalysts, not forecasts. Either the curve bends and orbital compute becomes a real wedge of the AI infrastructure stack, or it remains a premium niche for specialized workloads.
Just as the build-out of terrestrial hyperscale created durable franchises in power, cooling, and interconnect, the orbital build-out will produce the similar franchises in the areas that matter.
The question is not whether AI will need more compute than Earth can comfortably supply. It already does. The question is where the marginal gigawatt gets built, and the answer might just be directly above us.
Sources
Disclaimer: The content contained herein is provided for general informational purposes and does not constitute a recommendation, offer, or solicitation to buy or sell any securities. The content reflects the writer’s views and analysis as of the time of writing and are intended to support investment decision-making by providing an analytical perspective and context. The content does not address every factor relevant to any particular investor’s circumstances, and investors should evaluate their own facts and circumstances before making any investment decision. Past performance is not indicative of future results.
