The world's largest AI labs are building data centers the size of small cities. They are draining local power grids, consuming billions of gallons of cooling water, and hunting for cheap land in places that can handle the electrical load. Microsoft, Google, Meta, and Amazon are collectively committing hundreds of billions of dollars to the problem — and the demand is still outpacing supply.
SpaceX has a different answer: leave the atmosphere entirely.
On January 30, 2026, SpaceX filed a request with the Federal Communications Commission to launch up to 1 million solar-powered satellite data centers into low Earth orbit — a constellation it calls the most efficient way to meet the accelerating demand for AI computing. By June 8, the company had unveiled its first piece of hardware for the plan: the AI1, an orbital supercomputer with a 70-meter wingspan, 150 kilowatts of peak AI compute, and cooling straight into the vacuum of space.
This is either one of the most audacious infrastructure bets in the history of computing or an enormously expensive distraction. Probably it is a bit of both. Here is the full picture.
The Problem SpaceX Is Trying to Solve
To understand why SpaceX is doing this, start with the constraint every major AI lab faces right now.
Training and running large AI models requires enormous amounts of power. A single large-scale training run can consume as much electricity as a small city for weeks. The inference load — serving those models to millions of users around the clock — adds a persistent, growing baseline on top of that. The International Energy Agency projects that AI data centers will consume roughly 1,000 terawatt-hours per year by 2026, representing 3–4% of total US electricity consumption and rising sharply.
That creates three interlocking problems:
Power. Grid operators in the US, UK, and Europe are increasingly struggling to accommodate the sudden, massive power demands of new hyperscale data centers. Permitting timelines for new transmission lines run 7–10 years in the US. Several planned data center sites have been rejected outright because local grids cannot supply the load.
Cooling. Large data centers require enormous quantities of water or refrigerant-based cooling. Microsoft's Azure data centers consumed 6.4 million cubic meters of water in a single year. In drought-prone regions, data center water use is drawing regulatory scrutiny and community opposition.
Land and permitting. A hyperscale facility requires not just land, but land near sufficient power, near fiber connectivity, not in a flood zone, not in a flight path, and increasingly not in a jurisdiction where local residents will fight it in court for years.
SpaceX's claim is that a solar-powered orbital data center solves all three in a single move: unlimited solar power with no grid connection, the vacuum of space as an infinite heat sink, and no land use at all.
Whether that claim holds up under engineering and economic scrutiny is the more complicated part of the story.
The FCC Filing: 1 Million Satellites
On January 30, 2026, SpaceX submitted a filing to the FCC that landed like a thunderclap in the data center industry.
The filing requested authorization to operate a constellation of up to 1,000,000 solar-powered satellite data centers in low Earth orbit. SpaceX described the system in blunt terms: "the most efficient way to meet the accelerating demand for AI computing power."
Key technical parameters from the filing:
- The system is expected to remain solar-powered for more than 99% of its operations depending on the orbital plane — minimizing battery requirements
- The satellites are designed as compute nodes, not communications relays, meaning their primary function is running AI workloads rather than routing internet traffic
- The constellation is distinct from Starlink, though it would operate in low Earth orbit at similar altitudes
The scale of the proposal is unlike anything previously filed for orbital infrastructure. Starlink, which is already the world's largest satellite constellation, operates approximately 6,000–7,000 active satellites. A million compute satellites would be roughly 150 times larger.
The FCC filing is not a launch commitment — it is a spectrum and orbital-slot reservation. But it establishes the legal and regulatory footprint for a constellation of a scale no one had publicly planned before.
The xAI Merger: Context That Matters
The orbital data center plan did not emerge from SpaceX in isolation. It is the long-term infrastructure arm of a much bigger corporate reorganization.
On February 4, 2026, Elon Musk announced that SpaceX had acquired xAI, merging his AI company into the rocket company in a single entity. The logic was vertical integration at an extreme scale: xAI's models (Grok), xAI's compute operations (including the Colossus 1 supercomputer in Memphis with 220,000 NVIDIA GPUs), and SpaceX's launch, satellite, and orbital expertise would now operate under one roof.
The Colossus 1 data center, which SpaceX/xAI built in Memphis at extraordinary speed — the facility went from empty industrial space to operational in roughly 122 days — became the near-term compute asset. The orbital constellation became the long-term plan.
The scale of demand the merged company is already serving is staggering. Anthropic is reportedly paying approximately $1.25 billion per month for access to xAI data-center compute capacity. Google has agreed to pay roughly $920 million per month for AI infrastructure access. These are not future projections — these are reported current contracts, from companies that cannot build compute fast enough to meet their own model training and inference needs.
That revenue base is what makes the orbital bet credible as a business. SpaceX is not pitching a speculative future product. It is building additional capacity for customers who are already paying, who need more, and who face the same land-power-cooling constraints that make orbital compute potentially attractive.
AI1: The Hardware, Explained
On June 8, 2026 — timed deliberately to the week of SpaceX's anticipated IPO — the company unveiled AI1, the first physical satellite in its orbital data center constellation.
The specs are genuinely striking.
Dimensions and Structure
- Wingspan: 70 meters tip-to-tip — wider than a Boeing 747-8 (68.4 meters)
- Height when deployed: 20 meters
- Form factor: A large solar-array platform with a central compute module
AI1 is, in Elon Musk's words, "much simpler than a Starlink satellite." Where Starlink satellites are sophisticated phased-array communications devices, AI1 is essentially a large solar panel attached to a computing rack and a radiator. Musk's framing is that the hardware simplicity works in the system's favor — fewer moving parts, more reliable manufacturing at scale.
Power
- Solar array assumption: 250W per square meter
- Sustained compute power: 120 kW
- Peak compute power: 150 kW
The 70-meter wingspan translates to a very large effective solar collection area. At 250W per square meter, a 70x20 meter array (roughly) would generate approximately 350 kW of raw solar power — leaving margin after accounting for conversion losses and compute overhead.
The 99%+ solar-powered operating time cited in the FCC filing comes from orbital mechanics. In low Earth orbit, satellites spend the majority of each 90-minute orbital period in sunlight. Depending on the inclination and altitude, eclipse periods are short and predictable. For a compute workload that can tolerate brief pauses or route work to other nodes during eclipse, batteries become a secondary concern rather than a primary design driver.
Cooling
- Radiator type: Deployable liquid radiator
- Radiator area: 110 square meters
- Radiator efficiency: Approximately 1,400W per square meter
This is one of the least-discussed but most technically interesting aspects of AI1. In a terrestrial data center, cooling is typically the second-largest cost after power — requiring chillers, cooling towers, water, and associated infrastructure. In orbit, the heat rejection mechanism is thermal radiation into the approximately 2.7 Kelvin background of space.
A deployable liquid radiator circulates coolant through a large flat panel, radiating waste heat directly into space. The 110 square-meter radiator at 1,400W/m² provides approximately 154 kW of heat rejection capacity — slightly above the 150 kW peak compute figure, which is exactly the right margin.
This is free cooling at a planetary scale. There is no water, no refrigerant chiller, no cooling tower, and no energy cost beyond pumping the coolant through the radiator loop. If this approach works at scale, it eliminates one of the two largest cost drivers of terrestrial data center operations.
Compute
- Payload: AI compute hardware (chipmaker not specified)
- Architecture: Interchangeable — SpaceX has deliberately not locked the payload to a single vendor
- Efficiency target: 70 kW per ton
The interchangeable chip payload is a significant design choice. SpaceX is not building AI1 around a specific NVIDIA, AMD, or custom-silicon product. The compute module is designed to be swapped — meaning the satellite bus, solar array, and cooling architecture can remain stable while the compute payload evolves with chip generations.
This matters because chip generations in AI compute have been cycling faster than satellite design and production cycles. By separating the satellite bus from the payload, SpaceX preserves optionality in a market where the best-available AI chip changes every 12–18 months.
The Solar Advantage: Why Space Power Is Different
The solar energy case for orbital data centers deserves more careful treatment than it usually gets in coverage of this announcement.
On the Ground, Solar Is Intermittent
Terrestrial solar data centers face a fundamental mismatch: AI compute demand runs 24/7, but solar panels only generate power during daylight. This forces a choice between expensive battery storage, grid backup (which usually means fossil fuels), or simply accepting that the facility will not run at full capacity at night.
Several AI labs have made green energy commitments — Microsoft, Google, and Meta have all pledged 100% renewable electricity. But "100% renewable" in practice often means purchasing renewable energy certificates (RECs) that do not correspond to real-time renewable generation at the data center location. The actual electrons powering the servers may come from whatever is on the grid at midnight.
In Orbit, Solar Is Near-Constant
In low Earth orbit at roughly 550 km altitude (similar to Starlink), a satellite experiences an orbital period of approximately 90 minutes. Depending on the orbital inclination and season, eclipse periods run 20–35 minutes per orbit — meaning the satellite is in sunlight for 65–75% of each orbit, and over the course of a day, near-continuous solar generation is achievable by routing workloads across nodes in different orbital phases.
SpaceX's 99% solar-powered claim reflects this arithmetic. A well-designed constellation with sufficient battery for eclipse periods — or simply with workloads that can migrate between nodes — achieves something terrestrial solar cannot: solar power that is functionally equivalent to baseload power.
This is genuinely different from ground-based solar. It is not intermittent renewable energy. It is renewable energy that behaves like a dispatchable resource.
Scale
The sun delivers approximately 1,361 watts per square meter in space — the solar constant. On the ground, atmospheric attenuation, weather, and the day/night cycle reduce effective collection to roughly 150–250W/m² on an annualized basis. In orbit, there is no atmosphere, no weather, and orbital mechanics replaces the day/night cycle with a much more favorable duty cycle.
A constellation of even 10,000 AI1-scale satellites — 1% of the filed capacity — with 70-meter wingspans each generating ~300 kW of usable compute power, would represent roughly 3 gigawatts of compute capacity, running on solar energy that never has to fight a power purchase agreement, a grid constraint, or a drought.
Is It Feasible? An Honest Assessment
The SpaceX AI1 announcement is real. The FCC filing is real. The xAI merger is real. Anthropic's and Google's compute contracts are reported to be real. But "announced" and "will work" are different things. Here is where genuine uncertainty lives.
What Works in Principle
Solar power in space works. The International Space Station has operated on solar power for 25 years. The physics of orbital solar generation are not in dispute. AI1's solar architecture is a scaling exercise on proven technology, not a fundamental scientific bet.
Radiative cooling in vacuum works. Spacecraft thermal management via deployable radiators is standard engineering. The James Webb Space Telescope uses a sunshield and passive radiative cooling to reach temperatures near absolute zero. AI1's liquid-cooled deployable radiator is well within the bounds of existing space engineering practice.
Starcloud has demonstrated GPU compute in orbit. Before this became a SpaceX story, a startup called Starcloud successfully operated GPU clusters in low Earth orbit, proving that commercial AI compute workloads can run in the space environment. The fundamental concept — GPUs in orbit, powered by solar, cooled by radiation — has been validated at small scale.
LEO latency is adequate for many AI workloads. Low Earth orbit at 550 km altitude introduces approximately 3–7 milliseconds of one-way propagation delay from the satellite to a ground station — sub-20ms round-trip. For batch AI inference, model training (where data is sent up once and results come down later), and edge preprocessing, this is acceptable. It is not adequate for real-time interactive applications that need sub-10ms round-trip, but that is not the primary use case SpaceX is targeting.
What Remains Unproven
Launch costs at the required scale. The entire economic case for orbital data centers depends on Starship reducing launch costs by an order of magnitude relative to Falcon 9. SpaceX has quoted Starship targets of $100/kg or lower; current Falcon 9 costs are roughly $2,700/kg to LEO. AI1 weighs an estimated 2–3 tonnes based on its dimensions and the 70 kW/tonne efficiency figure. At $2,700/kg, launching one AI1 satellite costs $5–8 million before the satellite hardware cost — making it uneconomical against terrestrial alternatives. At $100/kg, the launch cost per satellite drops to $200,000–300,000, which changes the calculus entirely. Starship has not yet achieved commercial payload operations, let alone $100/kg pricing.
Radiation hardening at scale. Low Earth orbit exposes hardware to significantly elevated radiation levels compared to Earth's surface — particularly when traversing the South Atlantic Anomaly, a region of intensified radiation belt exposure that satellites cross multiple times per day. Consumer-grade AI chips are not radiation-hardened. Radiation-hardened chips are expensive, lag commercial performance by 2–3 chip generations, and are produced in small volumes. SpaceX's interchangeable payload approach suggests the company is not planning full radiation hardening of the compute module — but then the question is how frequently radiation-induced failures will require satellite replacement, and what that replacement cadence costs.
In-orbit servicing. When a terrestrial server fails, a technician replaces it in minutes. When an orbital compute node fails or degrades, the options are: (a) accept the loss and route workloads elsewhere, (b) launch a replacement satellite (expensive), or (c) develop robotic in-orbit servicing capability (expensive and not yet commercially available at scale). SpaceX has not disclosed a servicing architecture for the AI1 constellation. The interchangeable payload module hints at future servicing capability, but the logistics remain unaddressed publicly.
Per-kWh fully-loaded compute cost. SpaceX has published no comprehensive cost model that accounts for satellite manufacturing, launch, operations, radiation-induced failure rate, replacement cadence, and ground-station data throughput costs. The "free solar, free cooling" framing is real but incomplete. Until there is a fully-loaded $/kWh or $/FLOP figure for AI1 compute, comparisons with terrestrial data centers (where the fully-loaded cost is well-understood) are difficult.
Bandwidth to and from orbit. A data center is useless if you cannot get data in and results out fast enough. AI training requires moving enormous datasets — petabytes — into the compute cluster. LEO-to-ground optical laser communications (which SpaceX uses on Starlink) can achieve multi-gigabit throughput, but scaling this to data-center-class I/O bandwidth across a constellation of millions of satellites is a non-trivial systems engineering problem. SpaceX has not detailed the inter-satellite link or ground-station bandwidth architecture for the AI compute constellation.
The Gigasat Factory: Scaling Production
The orbital data center vision requires hardware at a scale that no spacecraft manufacturer has ever attempted. SpaceX has publicly announced plans to build what it calls a Gigasat factory in Bastrop, Texas — the same county where SpaceX's Starbase orbital launch facility operates.
The planned facility covers more than 1,000 acres with over 11 million square feet of potential building space. Production targets include not just the AI1 satellites themselves but vertically integrated solar manufacturing — ingots, wafers, and solar cells — with production targeted to begin by the end of 2027.
The vertical integration of solar manufacturing is notable. SpaceX is not planning to buy solar panels from existing suppliers; it is planning to manufacture them. This mirrors the company's approach to Starlink hardware — building the satellites internally rather than purchasing from traditional aerospace suppliers, enabling cost reductions through volume and vertical control.
For context, Tesla's Gigafactory in Nevada covers 5.3 million square feet and produces batteries at scale for electric vehicles. SpaceX's Gigasat facility, at a stated target of 11 million square feet, would be roughly twice the size — dedicated to satellite AI compute hardware.
Whether the factory reaches that scale, and on what timeline, will be one of the key indicators of whether SpaceX's orbital data center plans are on track.
Timeline: What Is Actually Scheduled
| Milestone | Target Date |
|---|---|
| FCC filing for 1M-satellite constellation | January 30, 2026 (done) |
| SpaceX-xAI merger announced | February 4, 2026 (done) |
| AI1 satellite unveiled | June 8, 2026 (done) |
| Technical specs published | June 9, 2026 (done) |
| AI1 prototype satellites launch | Early 2027 |
| Gigasat factory solar production begins | End of 2027 |
| Commercial constellation deployment begins | 2028 |
| 1 GW orbital compute capacity | Late 2027 (SpaceX target) |
The 1 GW target by late 2027 deserves scrutiny. At 150 kW peak per AI1 satellite, reaching 1 GW would require approximately 6,700 AI1 satellites — all manufactured, launched, and operational within roughly 18 months of the prototypes flying. That is an aggressive production ramp by any standard, comparable to the early Starlink constellation deployment pace.
Starlink reached 6,000 satellites in orbit in approximately four years from first launch. SpaceX would need to approach that scale in under two years for AI1. The Gigasat factory timeline — production starting end of 2027 — suggests the 1 GW figure is aspirational rather than committed.
What This Means for the AI Industry
Whether or not AI1 succeeds on its stated timeline, the announcement reshapes several conversations.
The energy constraint is SpaceX's wedge. The company did not position orbital data centers as a curiosity or a moonshot. It positioned them as the solution to a problem that is already limiting AI scaling. When Anthropic and Google are paying billions of dollars per month for compute access and still cannot get enough capacity, the pitch for "unlimited solar-powered compute above the atmosphere" lands very differently than it would have five years ago.
xAI is the near-term revenue engine; orbit is the long-term infrastructure play. The Colossus 1 supercomputer in Memphis is generating real revenue today. The orbital constellation is a multi-year buildout. SpaceX is not betting on orbit to the exclusion of ground-based capacity — it is using ground-based revenue to fund the longer-term build.
The interchangeable chip payload is a supplier-agnostic architecture. By designing AI1 to accept swappable compute modules, SpaceX has positioned itself as an infrastructure provider rather than a compute buyer locked to NVIDIA. If AMD, custom AI silicon, or a future chipmaker produces better AI performance per watt, AI1 can accommodate it. This is the compute equivalent of a cloud platform staying provider-agnostic.
The scale of the FCC filing signals intent, not capability. Filing for 1 million satellites is not the same as deploying 1 million satellites. It is a regulatory reservation — a way of establishing priority for orbital slots and radio spectrum in case the technology and economics develop favorably. SpaceX will almost certainly not launch a million satellites. But having the slot reservations means it can scale as fast as production and launch cadence allow, without hitting regulatory bottlenecks later.
The Bigger Picture: Space as Infrastructure
There is a way to look at this that makes the scale less surprising.
SpaceX is already the world's largest satellite operator by a factor of roughly 5x. It builds its own rockets, its own satellites, and (increasingly) its own solar cells. It launches more mass to orbit annually than all other launch providers combined. The Colossus 1 data center went from empty building to operational AI supercomputer in 122 days — a deployment speed that no traditional data center developer has approached.
The orbital AI data center plan is the same organizational capability applied to a new problem: abundant, cheap, low-carbon compute for AI workloads that are straining terrestrial infrastructure. The question is not whether SpaceX has the technical organization capable of attempting this — it demonstrably does. The question is whether the economics work out at the launch costs and satellite lifetimes that the plan requires.
Early 2027 will be the first real data point. Two AI1 prototype satellites will either demonstrate that solar-powered orbital AI compute is operationally viable, or surface the engineering challenges that have not yet been publicly acknowledged. Until then, the honest answer to "is this possible?" is: the physics work, the engineering is largely precedented, the economics are unproven, and the timeline is aggressive.
That is a narrow aperture of uncertainty for a plan this ambitious. It is also why this announcement deserves to be taken more seriously than most "space future" pitches.