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. … it’s credible that space could house much or even the majority of compute buildout throughout the 2030s.
(Footnote 1: We hope to do more analysis on how transformative AI might change this picture in the future. Speculatively, our initial thinking is TAI could accelerate the timeline over which compute transitions to space but this is not necessarily the case. In particular, during an industrial explosion pressure to grow rapidly might be so strong as to incentivize aggressive usage of non-renewables on Earth like oil and gas. If so, transition to space might be delayed for a one time boost on Earth, in which case the picture may look similar to the one we outline here, but with the added prologue of a large-scale terrestrial buildout.)
Below we give two example scenarios, one conservative and one bullish.21 (Footnote 21: In particular, these correspond to our rough estimates for the 10th and 90th percentile fastest cases. Though we think there could be as much as 12-18 months of variance for most major milestones.) 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.
Slow ODC takeoff scenario. In this scenario, Starship reaches rapid reusability mid-2030 having suffered many setbacks. SpaceX builds just one additional gigabay over the ~4.5 year development cycle of Starship, then begins to ramp production. By 2031, 3 gigabays have produced a total of 18 Starships but turnarounds are slow, requiring 6 days of downtime after each flight. The first meaningful tranche of orbital compute is deployed throughout 2031 and totals around 1,095 launches amounting to roughly 4 GW of orbital compute. Meanwhile forecasts expect perhaps 100 GW of terrestrial data center buildout by 2030. Thus, orbital compute makes a negligible fraction of total compute until perhaps later throughout the 2030s.
Fast ODC takeoff scenario. In this scenario, Starship begins deploying Starlink V3 satellites in 2026 and reaches rapid reusability by 2027. SpaceX then builds out 2 new gigabays by EOY 2027 and ramps production variously across 4 facilities, producing at least 48 Starships by early 2028. By this point Starships are produced at one Starship per month on average, as contrasted with Boeing’s ~38 planes per month or Tesla’s production of ~10,000+ cars per month from a single factory floor. Throughout the year, the growing fleet manages a total of 10,000+ launches. By EOY 2030, up to 100GW of orbital compute is operational.22
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.
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
(Footnote 18: The remaining milestones, roughly in order are: Block 3 debut with Raptor 3 engines (Flight 12, targeting May 2026; the first Block 3 booster exploded during pressure testing in November 2025 and its replacement passed cryo testing in February); catching the Ship at the tower (likely Flights 13 to 15), which is the single most important step since every launch currently discards a ~$30M upper stage; an orbital refuelling demo (targeting mid-2026); a second launch site at Kennedy Space Center (targeting late 2026), roughly doubling annual launch capacity; and rapid turnaround of both stages (likely late 2027 or 2028), which would be the point where launch costs fall dramatically.)
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.
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. At the ambitious end, two broad approaches are being pursued for data-center-scale compute, each with distinct tradeoffs.
Constellation model (SpaceX, Google, Blue Origin). A fleet of individual satellites, each equipped with GPUs, orbiting independently. This leverages SpaceX’s existing mass-production capability and could be relatively quick to deploy. The unsolved limitation is bandwidth: today’s commercial inter-satellite optical links deliver on the order of 1 to 100 Gbps over thousands of kilometres, which is orders of magnitude short of what tightly coupled ML training clusters require. Google’s Project Suncatcher proposes packing satellites into tight clusters (kilometre-scale or sub-kilometre-scale separations rather than thousands of kilometres) and using optical links similar to terrestrial fibre technology. At short distances, the physics of free-space optics improves somewhat, and Suncatcher suggests bandwidths on the order of 1–2 Tbps per transceiver pair may be achievable. This is an interesting idea, but it shifts the challenge to maintaining dense satellite formations with precise pointing and tracking.
Modular station model (Starcloud). A central spine structure with docking ports, to which compute modules (roughly 40 MW each) are launched and physically attached, connected to a large solar array. This sidesteps the bandwidth problem entirely, since modules are physically connected. However, it is further from what existing manufacturing and launch capabilities can deliver so this is a longer term vision, and introduces engineering challenges around structural integrity at scale. We found Starcloud’s published cost estimates to be overly optimistic compared to our assessment, though it’s possible Starcloud has one or more clever engineering tricks in mind while we did not try hard to optimize beyond today’s space hardware. The concept may also become more feasible after continued investment in space-based infrastructure.
The implications of the bandwidth constraint depends on whether ODCs are being used for training, or inference:
Inference is a natural fit for orbital compute. Modestly equipped satellites could serve inference workloads where each request is independent. Latency from LEO to the ground is also modest at a few milliseconds.
Training looks harder. Frontier runs synchronize gradients across thousands of GPUs at every step, demanding sustained bandwidth between nodes that dwarfs current inter-satellite links. Even Suncatcher’s proposed 1–2 Tbps, if achieved, would need to connect enough satellites to match the bisection bandwidth of a terrestrial cluster. The modular station model avoids this by physically connecting modules, but at the cost of in-space assembly no one has yet attempted.
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
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
(Footnote 19: Notably, many other analyses give all-in flight costs ranging from $1,000/kg to $10/kg (our median of $100/kg is in the middle of this range) and it can be difficult to tell what motivates a given assumption. SpaceX shares target numbers for launch costs but relatively little information about the economics of the vehicle, we think the third-party analysis The Starship Report done by Payload gives the best breakdown and we use their $90M build cost to estimate the cost of a Starship flight in various reuse regimes. Further sources include: The Basic Economics of Starship, and this aggregation from NextBigFuture.)
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.
Some notes and charts from Forethought’s report (model). Man, their reports are so aesthetic.
Some other quotes and links
… 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.
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
(Footnote 18: The remaining milestones, roughly in order are: Block 3 debut with Raptor 3 engines (Flight 12, targeting May 2026; the first Block 3 booster exploded during pressure testing in November 2025 and its replacement passed cryo testing in February); catching the Ship at the tower (likely Flights 13 to 15), which is the single most important step since every launch currently discards a ~$30M upper stage; an orbital refuelling demo (targeting mid-2026); a second launch site at Kennedy Space Center (targeting late 2026), roughly doubling annual launch capacity; and rapid turnaround of both stages (likely late 2027 or 2028), which would be the point where launch costs fall dramatically.)
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.
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. At the ambitious end, two broad approaches are being pursued for data-center-scale compute, each with distinct tradeoffs.
Constellation model (SpaceX, Google, Blue Origin). A fleet of individual satellites, each equipped with GPUs, orbiting independently. This leverages SpaceX’s existing mass-production capability and could be relatively quick to deploy. The unsolved limitation is bandwidth: today’s commercial inter-satellite optical links deliver on the order of 1 to 100 Gbps over thousands of kilometres, which is orders of magnitude short of what tightly coupled ML training clusters require. Google’s Project Suncatcher proposes packing satellites into tight clusters (kilometre-scale or sub-kilometre-scale separations rather than thousands of kilometres) and using optical links similar to terrestrial fibre technology. At short distances, the physics of free-space optics improves somewhat, and Suncatcher suggests bandwidths on the order of 1–2 Tbps per transceiver pair may be achievable. This is an interesting idea, but it shifts the challenge to maintaining dense satellite formations with precise pointing and tracking.
Modular station model (Starcloud). A central spine structure with docking ports, to which compute modules (roughly 40 MW each) are launched and physically attached, connected to a large solar array. This sidesteps the bandwidth problem entirely, since modules are physically connected. However, it is further from what existing manufacturing and launch capabilities can deliver so this is a longer term vision, and introduces engineering challenges around structural integrity at scale. We found Starcloud’s published cost estimates to be overly optimistic compared to our assessment, though it’s possible Starcloud has one or more clever engineering tricks in mind while we did not try hard to optimize beyond today’s space hardware. The concept may also become more feasible after continued investment in space-based infrastructure.
The implications of the bandwidth constraint depends on whether ODCs are being used for training, or inference:
Inference is a natural fit for orbital compute. Modestly equipped satellites could serve inference workloads where each request is independent. Latency from LEO to the ground is also modest at a few milliseconds.
Training looks harder. Frontier runs synchronize gradients across thousands of GPUs at every step, demanding sustained bandwidth between nodes that dwarfs current inter-satellite links. Even Suncatcher’s proposed 1–2 Tbps, if achieved, would need to connect enough satellites to match the bisection bandwidth of a terrestrial cluster. The modular station model avoids this by physically connecting modules, but at the cost of in-space assembly no one has yet attempted.
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