Will We Really Put Data Centers in Space?
How soon could AI data centers move to orbit? Forethought analyzes launch costs, cooling physics, and other engineering and governance factors needed to make space compute viable.
How soon could AI data centers move to orbit? Forethought analyzes launch costs, cooling physics, and other engineering and governance factors needed to make space compute viable.

22nd May 2026
[AI Narration] Will We Really Put Data Centers in Space?
Playback speed
Volume
0:00 of 1:14:38
Several major technology companies have announced plans to operate AI data centers in orbit. Elon Musk recently claimed: “the lowest-cost place to put AI will be space […] within two years, maybe three.” If a meaningful fraction of new AI compute really is placed in space within a few years, that would be a fairly big deal for AI governance and strategy. Here we try to disentangle the hype from reality and provide a sober assessment of the technical and economic feasibility of orbital data centers (ODCs).
The main case for ODCs is the cost of energy: space solar panels in the right orbits receive more constant and intense sunlight compared to Earth. Moreover, ODCs don’t currently face the same permitting and regulatory delays as on Earth, cause fewer ongoing environmental harms compared to grid or onsite natural gas-powered data centers, and may be more secure against data exfiltration. We find that the cost-competitiveness case for ODCs depends almost entirely on Starship achieving reusability comparable with what SpaceX achieved with Falcon: space-based solar reaches cost parity with present-day off-grid terrestrial power continuously at roughly $250/kg to orbit, and becomes cheaper than any current terrestrial energy source at around $50/kg, from the present-day launch cost of roughly $1,500/kg. Radiative cooling, often cited as a fatal obstacle, appears surprisingly manageable — potentially even cheaper than on Earth. However, ODCs may require substantial (perhaps ~38%) extra non-compute hardware (like solar, racks, and cooling) over 5 years to compensate for their inability to swap out failed chips, and inter-satellite bandwidth limitations likely confine ODCs to inference workloads, at least early on.
Assuming no transformative AI,1 but continued demand for data center buildout, we estimate that ODCs are unlikely to represent a meaningful share of compute before 2030, but become cost-competitive with present-day terrestrial data centers within 3–5 years if Starship development stays on track.
Some of the world's largest technology companies continue racing for compute. If progress continues, demand for data centers may more than double by 2030.2 Increasingly, though, new data center capacity is bottlenecked by multi-year queues to connect to the power grid.3
The result has been a scramble for workarounds. Leading AI labs have increasingly adopted a “Bring Your Own Generation” model to source power, deploying onsite gas turbines and engines to bypass grid bottlenecks. xAI, for example, reportedly installed hundreds of megawatts of onsite gas generation in Memphis to accelerate deployment, and OpenAI and Oracle have placed large turbine orders for new Texas campuses.
Some argue that energy will become the binding constraint on AI progress, given grid interconnection delays as gas turbines are themselves facing multi-year manufacturing backlogs. But the constraint does not appear fundamentally binding (as Epoch notes): turbine manufacture may expand to meet more demand and companies could go off-grid using combinations of gas, solar, and batteries, scaling power in parallel with compute, albeit at a cost premium. This raises a natural question: if you're going off-grid anyway, then what’s the best way to get power and where is the best place to put your data center?
Some think the answer will be in orbit. In November 2025, Google announced Project Suncatcher, a plan to put TPU-equipped satellites in dawn-dusk sun-synchronous orbit. In early 2026, SpaceX filed with the FCC for authorization to launch and operate a constellation of up to one million data center satellites.4 Other entrants include Blue Origin, Ramon.Space and startups like Starcloud, and Aetherflux while China's Three-Body Computing Constellation has launched 12 operational satellites and run Alibaba's Qwen3 model in orbit. Recently, at GTC in March 2026, NVIDIA announced the Space-1 Vera Rubin Module, meant to be a dedicated space-rated GPU platform.
At first glance, it seems very unlikely that any meaningful fraction (say, >10%) of additional data center capacity will be placed in space in the next few years. But if the companies betting on space are right, that would be a fairly big deal, and it could change the landscape of AI governance. For example, terrestrial data centers are subject to national and regional regulations, whereas AI developers could potentially exploit jurisdictional ambiguities around compute in space. Also, the path to low-cost orbital compute likely routes through a single launch company, SpaceX, which also now operates a frontier AI lab since its acquisition of xAI. And that might raise concerns around concentration of power.
We've been looking into the technical and economic viability of orbital data centers (ODCs). Our core model gives estimates for the total cost of Earth and space-based data centers across several scenarios.

Cost breakdown for three Earth-based and three space-based scenarios building out 1 GW of compute. As best we can determine, orbital data centers could become cost competitive with a bullish terrestrial buildout if launch cost reaches around $100/kg given modest reductions to server and cooling system mass, while a bullish case for orbital data centers with substantial mass reductions and launch at $50/kg may offer cost savings.
The report focuses on three questions. First, what is the basic economic case for a meaningful fraction of AI compute being placed in space? Second, the most obvious physical blocker: can you cheaply cool a data center in orbit? Third: how fast could the shift to space data centers happen, how soon, and what would have to go right?
Here is our provisional assessment:
The core of the economic case is energy. A solar panel in a dawn-dusk sun-synchronous orbit9 receives near-continuous direct sunlight, producing roughly 3 to 5 times the energy of the same panel at a good terrestrial site.10 Space-based panels can also be lighter and cheaper than their ground counterparts, since they need no weatherproofing, grid inverters, or heavy mounting hardware.11 But launching solar panels to orbit adds a fixed overhead, and at current prices, space solar doesn't pencil out. The question is: at what launch cost does space solar become cost-competitive with Earth-based power, and are those launch costs feasible within a few years?
To get a first sense of the answer, we can look at the levelized cost of energy (LCOE): the total cost of building and operating a power source, divided by its total lifetime energy output, expressed in dollars per megawatt-hour. This lets us compare across very different power generation technologies on equal footing. On this basis, space solar reaches cost parity with the cheapest present-day terrestrial energy sources at around $50/kg.12

Space solar LCOE compared to terrestrial generation technologies. Dashed curves show space solar LCOE as launch cost falls from $550 to $10/kg. Horizontal lines show median LCOE for terrestrial technologies today (unsubsidized). We model two space solar configurations, current starlink-like efficiency as an upper bound on cost based on current technology (blue line), and an optimized lightweight solar array pushing the limits of packing efficiency and construction as a lower bound (purple line).13
These LCOE figures, however, reflect the cost of energy at the point of generation; they don't capture what it actually costs to supply continuous, reliable power to a large data center operating off-grid. This distinction matters increasingly for hyperscale data centers, which are facing grid interconnection queues that are pushing operators toward building their own independent microgrids. A terrestrial solar facility supplying power 24/7 requires substantial battery storage and backup generation capacity to cover nights and weather interruptions. These are costs that aren’t accounted for in the LCOE figures shown above but would be a core need for off-grid data centers. When we include those system-level costs for the off-grid data center model, the picture for space solar actually looks better.14
We estimate that in-space solar (like that used by Starlink) reaches cost parity with terrestrial solar at around $250/kg in launch costs, which seems achievable in 1–2 years from the time of writing (early 2026).

Solar Cost Crossover. Earth-to-space cost ratio for powering 1 kW continuously via solar panel as a function of launch cost. This compares space-based solar against two terrestrial baselines: a bullish solar+gas microgrid (solid) and a 100% renewable off-grid facility (dashed) with crossover points marking where space becomes cost-competitive. The lines give ratios of the cost of Earth over the cost of space solar. The blue lines compare the currently flown starlink-like array as a conservative estimate while the purple shows more lightweight, high performance panels for an upper bound.15
Beyond energy costs, we'd like to know how ODCs compare to terrestrial data centers on cost all-told, line-item for line-item. On the build side, both architectures share compute hardware as their dominant cost, but diverge elsewhere: terrestrial facilities require power, cooling systems, physical buildings, land, and permitting, while ODCs replace these with space-specific hardware — solar panels, thermal radiators, satellite bus, and launch. On the operational side, terrestrial centers face ongoing energy bills, staffing, maintenance, and taxes, whereas ODCs largely eliminate these recurring costs by generating solar power freely in orbit. The operational liability for ODCs, however, is the inability to perform maintenance: hardware failures (routine in today’s data centers) would then result in permanent infrastructure loss. We think the business model for ODCs would be more fire-and-forget, accepting losses as they come and launching replacement satellites to compensate for in-orbit hardware failures. We model these failures as costing ~9% compute loss per year, based on estimates for chip death, radiation induced latchup, and mechanical failures detailed in subsequent sections.
| Earth | Space | |
|---|---|---|
| Build costs (CapEx) | Compute, Power (Solar+Batteries+Gas), Cooling Units, Building/Facility, Land, Permitting | Compute, Solar panels, Radiators, Satellite bus & structure, Launch |
| Operational costs (OpEx) | Energy, Staffing/Maintenance, Taxes | Replacement satellites (to compensate for chip bleed due to hardware failures) |
We can try to forecast total costs by modeling six scenarios (three for Earth and three for space). The three Earth scenarios span grid-connected construction with conventional cooling, off-grid microgrids with full battery autonomy, and a hybrid case using on-site solar with gas backup to bypass grid interconnection queues entirely. The three space scenarios are anchored to plausible Starship launch cost trajectories — conservative ($250/kg), median ($100/kg), and bullish ($50/kg) — with corresponding assumptions on thermal management, and server mass reduction.16 Across all scenarios, IT hardware is held constant at $25.6B per GW, isolating infrastructure and operational costs as the key variables. For Earth, non-compute build costs are amortized over 20 years while orbital infrastructure is phased out after a 5-year lifecycle (matching current Starlink lifecycles).17
| Scenario | Per 5 yr Cycle ($B) | Notes |
|---|---|---|
| Earth Grid | 37.6 | Grid-connected, conventional cooling |
| Earth Microgrid (Conservative) | 42.5 | Off-grid microgrid, full battery autonomy |
| Earth Microgrid (Bullish) | 36.8 | On-site solar + gas backup, bypasses grid queue |
| ODCs (Conservative) | 48.2 | $250/kg launch |
| ODCs (Median) | 35.0 | $100/kg launch, 17% mass reduction over terrestrial |
| ODCs (Bullish) | 30.3 | $50/kg launch, 54% mass reduction over terrestrial |
Our estimates suggest ODCs could approach cost parity with terrestrial builds at around $100/kg — a 15× reduction from today's ~$1,500/kg. Whether that's achievable, and on what timeline, depends largely on Starship's reusability trajectory, which we examine next.
The scalability of orbital data centers, and indeed a wide range of economic activity in space, depends on one vehicle. SpaceX's Starship seems to be the only launch system that could deliver the combination of low cost, high cadence, and large payload capacity that orbital data centers would require if they are to make up a meaningful fraction of additional data center capacity anytime soon.
As of early 2026, Starship has flown 11 times. It is a two-stage rocket: the bottom half (a booster called Super Heavy) launches the top half (the Ship) to orbit, then flies back to the launch site where a pair of mechanical arms on the tower catch it mid-air. This catching manoeuvre is now semi-routine, and SpaceX has reflown one booster, demonstrating that the hardware can survive reentry.
The next challenge is reusing the Ship, the upper stage which actually reaches orbit. After three consecutive failures in early 2025, the last two flights achieved controlled reentry and ocean splashdown, but the Ship has not yet been caught or flown twice. A recurring obstacle is the heat shield: reentering the atmosphere subjects the underside to temperatures exceeding 1,400°C, and the vehicle needs to survive this without needing expensive refurbishment. Elon Musk has called designing a reusable heat shield the "biggest remaining problem" for Starship.18
China's Long March 9, designed for 150 tonnes to LEO, is not expected to fly until 2030 to 2033. ESA, JAXA, and ISRO have no comparable programmes. New Glenn, Blue Origin's heavy-lift vehicle, landed its first booster in November 2025 but carries just 45 tonnes to LEO versus Starships 150-200 tonnes, with a super-heavy 9×4 variant (70+ tonnes) announced for perhaps 2027. Notably, Blue Origin investigated upper-stage reuse from 2021–2025 but shelved the project, meaning without full reusability, there is no path for New Glenn to approach the launch costs that even partial Starship reusability enables. As best we can tell, these are the closest competitors to Starship and they appear to be 3-5 years on key milestones.
We estimate launch cost across several reusability scenarios, we assume the build cost remains constant at $90M build cost per Starship, and $5M in per-flight operations.19
| Scenario | Booster life | Ship life | Amortized build | Ops / refurb | Per-flight cost | Payload (t) | $/kg |
|---|---|---|---|---|---|---|---|
| Expendable | 1 | 1 | $90M | $5M | $95M | 200* | ~$475 |
| Booster reuse | 10 | 1 | $6M + $30M | $5M | $41M | 180* | ~$228 |
| Full reuse (early) | 10 | 10 | $6M + $3M | $5M | $14M | 150 | ~$93 |
| Full reuse (like Falcon) | 25 | 25 | $2.4M + $1.2M | $4M | ~$7.6M | 150 | ~$51 |
| Full reuse (better than Falcon) | 50 | 50 | $1.2M + $0.6M | $3M | $4.8M | 150 | ~$32 |
| Full reuse (bullish) | 100 | 100 | $0.6M + $0.3M | $2M | $2.9M | 150 | ~$19 |
*Expendable ships offer greater payload to orbit since they don’t need fuel for landing, while grid fins and whole systems can also be removed.
The number of reuses a given Starship will be rated for is a variable on cost. We don’t know how many reuses to expect from Starship but a good point of comparison would be with SpaceX’s previous rockets. SpaceX’s best with the Falcon rocket is 33 reuses while typical numbers seem to be around 15 reuses. We set our median launch cost at $100/kg, corresponding to ~10 reuses and our bullish scenario launch cost at $50/kg corresponding to ~25 reuses. Notably, both of these launch costs are comfortably below the ~$250/kg threshold when space solar wins out over terrestrial power generation and based on our model it is at around $100/kg that ODCs could reach overall cost parity with terrestrial data centers. The (practically irreducible) floor for launch cost is set by propellant (roughly $900K per launch for 4,500 tonnes of liquid oxygen and methane) plus ground operations (perhaps $1.1M), giving about $13/kg at 150 tonnes to orbit. Though recent updates from SpaceX claim the eventual Starship (Block 4) will be capable of delivering 200 tonnes to orbit which would hypothetically set the mature floor for Starship around $10/kg.
Given the above, it looks plausible that space-based data centers could become cost-competitive with terrestrial data centers beyond ~10 reuses per Starship. If this is right, a further question is how long it would take for ODCs to actually become a meaningful fraction of total compute. We’ll look at the main inputs to launch cadence and outline what a slow and fast scenario could look like for scaling orbital compute.
The turnaround time for each Starship. This is the time to catch, restack, load cargo and refuel the Starship before launching again, where SpaceX targets 3x/day. SpaceX’s current record turnaround is 9 days, but this being with Falcon rather than with a fully reusable vehicle. As for a rough lower bound: commercial airlines typically run 3-7 flights a day, with turnaround times of perhaps 1–2 hours. Still the limits of reusable heat shields are not known and the floor for turnaround might be much higher.
The time to construct new rocket factories and ramp their production. SpaceX has two Gigabay production facilities under construction (Boca Chica and Roberts Road, FL), both targeting completion by end 2026, but it's too early to tell what growth rates will be based on this. An apt comparison could be to Tesla, which opened 5 Gigafactories between 2016 and 2022, a pace of roughly one new major factory every 14 months. Construction times also fell sharply over this period: Giga Shanghai was completed in just 168 working days, roughly a third of the typical automotive factory build time.
The number of rockets/factory/year produced. SpaceX has discussed an eventual goal of producing ~365 Starships per year (~1/day). As a useful anchor, Boeing’s mature 737 program produces on the order of ~38–42 aircraft per month today, though this rate is currently capped by the FAA following quality issues, and historically has reached higher levels. An airliner contains on the order of hundreds of thousands of parts, whereas Starship’s stainless-steel architecture is relatively simpler. However, Boeing benefits from decades of production tooling and supply-chain maturity. SpaceX’s Falcon-9 ramp could also be illustrative. Early production went from about 25 rockets in 2020 to 165 by 2025.
Musk has claimed ODCs will likely be the cheapest way to generate AI compute in 2.5–3 years, though a more calibrated adjustment on SpaceX’s intentionally ambitious timelines could be 3–5 years. Below we give two example scenarios, one conservative and one bullish.21 In this way we try to give an upper and lower forecast for how quickly compute could be launched to orbit. Notably, there are relatively few good anchors for these timelines, so we’re especially uncertain about these estimates. Still we’re showing our best-guesses, for the sake of concreteness.
On Earth, data centers stay cool by blowing air or pumping liquid coolant over hot chips, and then venting that heat into the atmosphere. Google and Meta, which have pushed cooling overhead to best-in-class levels, dedicate roughly 7–10% of their electricity and perhaps 15–20% of facility capital expenditure to thermal management.23
In the vacuum of space, there is no air, so another method is needed. Liquid cooling can be used to pump heat away from chips, but that heat must ultimately be radiated away by emitting infrared light into the void. The physics governing this (the Stefan-Boltzmann law) requires a large surface area to shed a modest amount of heat, so first-pass estimates suggest you'd need prohibitively large radiator panels to cool the heat generated by just a few GPUs.24
On more careful analysis, however, the cooling problem seems both less important and more tractable than it first appears. Configured in the right way, it looks like relatively cheap and light radiators will be sufficient.
The radiators on the International Space Station, designed in the 1990s, manage an efficiency of about ~13 W/kg. But there has not been significant commercial pressure to optimize space radiators, since they’re currently a small fraction of the cost to launch and operate satellites. Thus, we expect some fairly low-hanging fruit for efficiency improvements. Compute modules would also escape needing the sprawling cooling loops and cold plates the ISS uses to sustain humans, and could run at much warmer radiator temperatures (panels at 25–45°C) than ISS radiators (whose ammonia outlet is controlled to about -40°C to stay above the freeze point), exploiting the T⁴ dependence in the Stefan-Boltzmann law.
A flat panel coated with standard off-the-shelf reflective paint radiating from both sides rejects a surprisingly large amount of heat.25 Built from lightweight carbon fibre composites with fluid loops to carry heat from the chips to the radiating surface, these passive radiators achieve roughly 163 W/kg at the system level in our estimates, a 13× improvement over ISS heritage technology.26 NASA Marshall's testing of High-Conductivity (Hi-K) carbon-composite radiators has demonstrated specific powers roughly 7× that of conventional metal fins (implying system-level specific powers of order ~90 W/kg at moderately elevated temperatures, using ISS-style metal-fin baselines), with hardware built and tested on the ground.
Crucially, the power rejected by a radiator scales to the fourth power of its temperature, per the Stefan-Boltzmann law. So moving from the ISS’s human-friendly temperatures, to industrial-grade rejection temperatures, allows ODCs to get notably more efficient at radiating heat.
Running a coolant loop at 25–45°C instead increases heat rejection per square metre by 2.7 to 3.5×. The practical ceiling is set by GPU thermal margins: modern chips throttle at around 83–85°C, and direct liquid cooling typically runs 15–25°C below the chip, capping useful coolant temperatures at perhaps ~60°C. At 25–45°C, the physics is sufficient to reduce radiator mass substantially. Lastly, this same T4 scaling means cooling could become even more efficient by actively pumping heat away from the chips and towards the radiators; though on our first-pass analysis, this looks like more hassle than it’s worth.27
If this is correct, thermal management looks like a modest cost. In our conservative scenario, radiator hardware is about 14% of system mass and ~5% of total cost. In our median scenario, given a warmer operating temperature, it drops to 8% of mass and roughly 2% of cost. Strikingly, this is less than what terrestrial data centers spend on cooling over a comparable lifecycle.28
That said, the kind of radiators we’re speculating about haven’t been flight proven, so we should also consider the more conservative scenario. Our estimates extrapolate from this baseline using thinner panels and higher operating temperatures, which is why we reach 163 W/kg and above. The most pessimistic assumptions we’ve seen have radiators accounting for as much as 60% of system mass. When launch costs are high, this is clearly a big deal. For example, if Starship indeed has a $250/kg launch cost because it is only partially reusable and there can be no improvement on the 1990s radiator design, then this would amount to an increase of over $8B per giga-watt. Compare this to ~$3B spent on cooling Earth-side. At our median launch cost $100/kg or bullish case of $50/kg this worst case scenario becomes a $3.5B increase or $1.7B increase respectively. This underscores how the most sensitive parameter by far for orbital data centers is launch cost. Making ODCs economical will require improving on existing radiator designs, but we conclude that cooling is unlikely to be a major bottleneck to ODCs becoming economically feasible, if they ever do.
A data center in orbit also faces a number of other difficulties that a terrestrial facility does not. We work through the most important ones here. Some of these appear manageable; others are unproven and uncertain.
We think chip failure and inter-satellite bandwidth are the hardest hitting challenges for ODCs. If solutions can’t be found, dead infrastructure due to chip failure could result in something like a 38% premium on non-IT infrastructure and launch costs (accounted for as ‘chip bleed’ in Figure 1); while low bandwidth may mean orbital data centers are useful only for inference and not competitive in frontier training runs.
On Earth, GPU failures are routinely handled. Meta's 16,384-GPU H100 cluster experienced 419 hardware failures in 54 days (roughly one every three hours), but a technician simply swaps the dead card while the training run continues via checkpointing.29 In space, this kind of maintenance is not feasible for the foreseeable future. Failed chips stay failed, and you bleed compute from day one.
How bad is this? Based on Meta's observed annualised failure rate of roughly 9%, a space data center would lose about 38% of its initial compute over a 5-year lifecycle.30 We accounted for this in the figures above by applying a 38% cost increase to all non-chip hardware, since a terrestrial operator would be buying the same replacement chips but could install them without losing any infrastructure.31
That’s a significant penalty. It might be partially offset by factors favouring the space environment (i.e. no dust, no humidity, after launch, no temperature cycling from HVAC), but additional space-specific failure modes (radiation-induced latchup, thermal cycling from eclipse periods for non-SSO orbits) could push the rate higher. The net effect is uncertain, and we think this one of the larger unknowns regarding orbital data centers.
How you physically arrange compute in orbit matters a great deal, particularly for training workloads that require fast communication between GPUs.
At the modest end, individual “smart” satellites could carry a few kilowatts of onboard compute, for example enhanced Starlink V3 nodes doing inference, caching, or on-orbit data filtering. Starcloud-1, a 60 kg test satellite launched in 2025 carrying an Nvidia H100, represents early hardware in this category. These are deployable today in principle, but the compute per node is small.

Illustration of two types of orbital data center architecture, the optically linked constellation model (most like today’s space tech) and the modular station model (most like traditional data centers).
At the ambitious end, two broad approaches are being pursued for data-center-scale compute, each with distinct tradeoffs.
The implications of the bandwidth constraint depends on whether ODCs are being used for training, or inference:
Radiation is substantially higher in orbit. Which causes two distinct problems: cumulative damage that permanently degrades chips over time, and transient bit flips that can corrupt data in memory.32 At the altitudes relevant to the first orbital data centers (around 500-600 km), both effects appear manageable based on the limited testing data available, though relatively little research exists. We don’t include any loss of compute efficiency due to errors or chip loss due to radiation damage in our model.
Operating at scale in low Earth orbit introduces several related risks around debris, collisions, and crowding. Currently these issues are manageable, but they’ll grow with the scale of ODC deployment.
No mass-production line for space-rated data center hardware exists today, and building one would take time and investment. SpaceX has demonstrated that space hardware can be manufactured at volume: Starlink launched roughly 2,500 satellites in 2025, and much of the production infrastructure (clean rooms, testing facilities, automated assembly) could in principle be adapted. Some commentators describe the path from here to orbital data centers as straightforward, essentially a matter of fitting more compute onto existing satellite platforms.
We think this understates the challenge. Communication satellites and compute-dense data center modules have different engineering requirements: higher heat loads requiring new thermal integration, power distribution architectures serving dense GPU arrays rather than modest communication payloads, and compute packaging at densities that have not yet been attempted in space. On the other hand, we’ve argued that viable launch-costs are at least 2-3 years away, and at least 3-5 years from reaching scale. Given those lead times, and the several companies facing massive economic incentives to solve these challenges, we think building out the ODC component supply chain is unlikely to be a major bottleneck.
The case for orbital data centers rests on accepting one of two distinct claims: that terrestrial buildout may eventually hit limiting constraints (land, permitting, grid capacity, water, community opposition), or that space may offer flatly superior unit economics once launch costs fall far enough. Both could be true simultaneously. In scenarios where the economics are roughly competitive rather than clearly favouring one side, several non-cost factors could tip the balance.
First, there is a straightforward strategic case for SpaceX in particular to invest in orbital data centers even at an initial loss: once Starship achieves full reusability, SpaceX will likely enjoy a multi-year window as the world's sole low-cost heavy-lift provider, and locking in the orbital data center market during that window — proving the architecture, building the supply chain, securing orbital slots — would create a structural moat that could persist long after competitors develop comparable vehicles.
As discussed, one key factor will be the speed at which new data centers can be built and launched, versus constructed on Earth. Permitting and regulation play a large role here. If the regulatory environment and build times on Earth look roughly like they do today, space may soon be the fastest place to build, whether or not it’s cheaper.
Another key consideration is physical and cyber security, especially as AI technologies become more strategically valuable. Orbital data centers would effectively eliminate physical access and insider threats, but they would become unusually easy to observe and destroy with anti-satellite weapons. As such, we conclude that orbital data centers would be substantially harder to exfiltrate data from, but easier to destroy.
For AI model weights, where theft is the primary concern, orbital hosting offers a genuinely stronger posture. For critical infrastructure where availability matters most, fragility against space weapons is a bigger liability.
Orbital data centers look more viable than we expected going in and the case rests primarily on launch cost. The engineering challenges seem significant but none appear intractable. Commercial off-the-shelf solar, passive radiators, and standard error-correcting codes seem sufficient for a baseline architecture. We expect limited demos and shipments of compute to orbit in the coming years but our best estimate is that orbital data centers are unlikely to represent a meaningful share of compute before 2030. They might scale quickly thereafter however, and are more likely to be bottlenecked by chip production rates rather than by launch cadence. A generally important dynamic is that terrestrial data centers likely face headwinds that worsen as compute scales up while orbital data centers face constraints that relax with scale.
| Factor | Terrestrial Data Centers | Orbital Data Centers | Favors |
|---|---|---|---|
| Cost per GW | Lower today | Higher today; lower with Starship and investment | Earth today; Space later |
| Frontier training | Well suited | Potentially bandwidth-constrained | Earth |
| Inference | Well suited | Well suited | Neutral |
| Cooling | Mature but costly | Immature, likely modestly cheaper at scale | Neutral |
| Reliability / maintenance | Repairable and well understood | No practical servicing; true failure rates uncertain | Earth |
| Permitting / local opposition | A sizable and likely growing burden with scale | Currently very favorable; opposition may grow with orbital congestion | Space |
| Land use | Fairly small footprint | Minimal terrestrial footprint | Space |
| Carbon / local environmental burden | Can be low in principle, but likely depends on gas-backed buildout | Favorable vs gas-backed terrestrial buildout | Space |
| Security from exfiltration | Hard; insider and physical-access problem | Stronger | Space |
| Security from destruction | More robust | More vulnerable to counterspace attack | Earth |
| Long-run scaling | Earth-bound constraints tighten with scale | Can scale nine orders of magnitude higher than Earth ceiling | Space |
All in all we think a transition to space-based compute is credible and could happen soon but almost surely not before the decade is out. There are a few key uncertainties which we expect to get answers for within the next few years. What would need to be true for ODCs to become a substantial fraction of total compute?
Thanks to Margot Stakenborg, Anders Sandberg, Tom Davidson, Max Dalton, Stefan Torges, Linchuan Zhang, Gregory Lewis, Charity Weeden, Madeleine Chang, Jackson Wagner and many other reviewers for helpful comments and discussion.
If you notice an error in our analysis or an important missed consideration we’d love to correct it. Please contact avi@forethought.org with any suggestions or leave a public comment.
Two effects matter. Total ionising dose (TID) causes cumulative charge buildup in transistor oxide layers, leading to threshold voltage shifts, increased leakage current, and eventual device failure. This is largely irreversible. Single event upsets (SEUs) flip bits in memory and registers but leave hardware physically undamaged, and are correctable via error-correcting codes or rewrite.
Altitude is the dominant variable. Trapped proton flux drops roughly 10× per ~300 km altitude decrease in the low Earth orbit range, where you're on the lower flank of the inner Van Allen belt (which peaks around 3,000–4,000 km). The choice of orbital altitude is critical.
TID reference numbers at different altitudes, assuming 10 mm aluminium shielding:
The limited chip testing data we have suggests comfortable margins at 550 km. Google's Trillium TPU showed first high-bandwidth memory irregularities at 2 krad (2.7× the 5-year shielded dose at 550 km) but no hard faults up to 15 krad (20× margin), tested via 67 MeV proton beam irradiation. NASA tested NVIDIA GPUs at 6 krad with no permanent failures. The key vulnerability appears to be high-bandwidth memory, not the logic/compute die itself. Modern sub-7nm CMOS has actually improved in TID tolerance because thinner gate oxides trap less charge. That said, these are small-sample tests on specific chips, and orbital data centers would deploy hardware at much larger scale than has been radiation-tested.
At 800–1,000 km, the picture looks much less favourable: you'd likely hit degradation thresholds within 1–2 years even with shielding. But shielding is relatively cheap compared to the compute it protects: roughly $100,000 in launch cost (at $250/kg) for a 20 mm shell around the satellite bus, versus ~$9,000,000 for the chips inside.
SEU rates appear manageable, but data is limited. An H100-class GPU in low Earth orbit would see roughly hundreds of raw bit flips per day in high-bandwidth memory. Standard single-error-correcting, double-error-detecting codes handle virtually all single-bit errors. The estimated uncorrectable (double-bit in the same code word) rate is ~0.01–0.1 per year per GPU. Google measured roughly 1 uncorrectable error per 10 million inferences on their TPU, which is operationally negligible. These numbers come from a small number of tests and could look different at fleet scale.
Solar cell degradation
Solar cell degradation depends on which cells you use, and we think commercial crystalline silicon cells are the right choice. Traditional space missions use radiation-hardened triple-junction GaInP/GaAs/Ge cells, which are extremely durable but expensive (roughly $50–100/W). Starlink instead uses commercial off-the-shelf silicon cells similar to terrestrial panels, at a fraction of the cost (roughly $0.20–0.50/W). These degrade faster under radiation, but the cost difference is enormous.
For orbital data centers, we think the Starlink approach is correct. Chip depreciation (new chips deliver 5–10× performance per watt every few years) sets an effective lifespan of roughly 5 years on any given module. Put simply, paying a large premium for solar cells that last 15–20 years doesn't make sense when the rest of the hardware is being replaced on a 5-year cycle.
In principle, the array oversize could be reduced by co-optimizing against chip bleed: since a fraction of GPUs fail each year, end-of-life power demand is lower than beginning-of-life, and the solar array only needs to cover surviving chips at time t rather than the original total. We haven't modelled this interaction however it could offer modest savings.
Other forms of wear
Space weather represents a tail risk. A handful of satellites (out of roughly 15,000) have been damaged or lost to space weather over the last 40 years, mostly in geostationary orbit, which is far more exposed than low Earth orbit.
Three main effects arrive in sequence after a solar event:
The most concrete threat to orbital data centers is atmospheric drag from thermosphere expansion after a coronal mass ejection. In February 2022, a moderate geomagnetic storm caused 38 of 49 newly-launched Starlink satellites to reenter and burn up. Critically, these were at a vulnerable 210 km staging altitude, not their operational 550 km. At operational altitude the drag increase was manageable with active station-keeping.
A severe storm (e.g. a Carrington-class event, estimated recurrence every 150–500 years) could cause several simultaneous problems: a burst of single event upsets potentially overwhelming error correction temporarily, accelerated solar cell degradation, and significant drag increases. In the case of such an event global satellite networks and orbital data centers could power down and orient edge-on in their direction of motion to minimize drag but it is unclear how devastating versus endurable this would be.
We model two space solar configurations, both using commercial off-the-shelf silicon cells (the Starlink approach, not traditional radiation-hardened triple-junction cells). We use 20% cell efficiency (AM0), 1,361 W/m² solar constant, 95% capacity factor (for dawn-dusk SSO), with 8% per year compound degradation over a 5-year design life.
Oversize factor: To deliver 1 kW continuous at end of life, the array must compensate for both capacity factor and degradation:
S1: Starlink V3-class (91 W/kg derated). Back-calculated from Starlink V3 specifications (50,400 W from 554 kg array, 257 m² area). System derating 0.721 (empirical, includes packing efficiency, wiring, temperature, MPPT). Hardware cost: $0.75/W derated (commercial Si + 1.5× NewSpace qual premium). Mass per kW-continuous: 17.56 kg. Hardware cost per kW-continuous: $1,198.
S2: Optimised stripped Si (189 W/kg derated). Component-level build: 0.85 kg/m² cell assembly + 40% structure overhead = 1.19 kg/m² total. System derating 0.828 (better packing on flat substrate). Hardware cost: $0.471/W derated. Mass per kW-continuous: 8.43 kg. Hardware cost per kW-continuous: $752.47
Similarly, we model two off-grid terrestrial solar setups, calculating the cost to deliver 1 kW continuously as in the space case. We do not add an oversize factor for the Earth case beyond what is needed to overcome the reduced CF and don’t account for solar cell degradation which is less severe than in space, but this somewhat disadvantages space solar.
Baseline A (100% renewable off-grid, $20,307/kW-continuous): Solar at 28% CF, $1.10/W installed, 4,643 W nameplate per kW-continuous ($5,107). Battery storage 80 kWh at $175/kWh ($14,000 or 69% of total). Other $1,200. 5-year LCOE: $463/MWh. Battery storage dominates because 99.9%+ uptime without fossil backup requires multi-day autonomy.
Baseline B (bullish solar + gas turbine, $5,590/kW-continuous): Solar at 32% CF, $0.85/W, 3,500 W nameplate ($2,975). BESS 12h at $125/kWh ($1,500). Gas turbine at $562/kW (EIA 2023 actual, simple-cycle; xAI Colossus uses Solar Turbines Titan 350) with 10% utilization and $153 fuel/O&M. Other $400. 5-year LCOE: $128/MWh.
Battery cost is the largest lever on Baseline A (69% of cost). If we allow for cost reductions to $50/kWh batteries (well below any 2025 projection), crossovers remain above $519/kg (S1) and $1,133/kg (S2). The space advantage is robust because the 95% CF advantage is structural.
Baseline B is less sensitive to battery cost since it leans on gas backup in cases of poor weather (storage is only 27% of total). In general, if building off grid to avoid grid interconnect wait times a reliable backup energy source can give large savings and should be the default for builders. We include Baseline A however, since backlogs for turbines versus the abundance of solar and battery available suggest that it may be the fastest way to quickly scale energy production on Earth.
In vacuum, heat can only leave a surface via radiation. The Stefan-Boltzmann law gives the total power radiated per unit area from a surface at temperature T
P=εσT4P = \varepsilon \sigma T^4
A radiator panel in orbit radiates from both sides, but also absorbs energy from three sources: direct sunlight, reflected sunlight (albedo) from Earth, and thermal infrared from Earth. The net power rejected per square meter is:
Pnet=2εσT4−αS−αFalbedo(Al⋅S)−εFIRTearth4P_{net} = 2\varepsilon\sigma T^4 - \alpha S - \alpha F_{albedo}(Al \cdot S) - \varepsilon F_{IR} T_{earth}^4
Where ε = 0.92, α = 0.09 (AZ-93 white paint), S = 1,366 W/m², Al =0.3, and Tearth = 253 K. Then given a two-sided radiator at a worst case altitude of 550km, FIR = 0.50, Falbedo = 0.25.
| Conservative | Mid / Bull | |
|---|---|---|
| Material | 1.5 mm CF composite | 1.0 mm CF honeycomb |
| Panel areal mass | 2.40 kg/m² | 1.60 kg/m² |
| System areal mass (incl. plumbing) | 3.60 kg/m² | 2.40 kg/m² |
| Operating temperature | 25°C | 45°C |
| Net rejection | 585 W/m² | 830 W/m² |
| System W/kg | ~163 | ~346 |
The environmental comparison between orbital and terrestrial data centers is often billed as favoring space unanimously, but the picture we found is a bit messier. Here we note some points of comparison.
Texto de forethought.org tal como lo publicó, leído de su página y mostrado aquí con el formato de El Autómata. El original, en forethought.org · Volver a la noticia de El Autómata