Satellite InferenceAI Inference in OrbitGet in touch ECONOMICS / COST MODEL
Year by year to 2045, against the same GPUs in a US data centre. Launch prices follow the experience curve of space launch, measured over more than 4,400 launches; the node's mass comes from our sizing model and every cost from our cost model. The simulator opens on our case; its scenarios show what happens if launch stops getting cheaper, satellites stay costly or launch grows faster or slower, and every assumption stays open for engineers.
In orbit the GPUs cost more per hour than on the ground: 15% are spares that can never be repaired, and the node computes 96% of the year. Launch and satellite include 5% a year for operations, ground stations and insurance.
Teal: the change that makes the GPU-hour cheaper; coral: dearer. Availability cannot rise above 100%, the node computing through every shadow.
| Launcher | 2026 design41 kg/kW | 2030 design26 kg/kW | 2035 design17 kg/kW | 100 kW per tonne10 kg/kW |
|---|---|---|---|---|
| Falcon 913 t to this orbit | 0.3 | 0.5 | 0.8 | 1.3 |
| Starship V369 t to this orbit | 1.7 | 2.7 | 4.0 | 6.9 |
Falcon 9 carries about 13 t to this orbit (cost model v3.1). Starship V3 is designed for 100 t to low orbit fully reused (SpaceX, Form S-1, 2026); to this higher orbit we assume 69 t, the 1.45 price ratio of our cost model. SpaceX states a target of over 100 kW of compute per tonne for its own satellites.
01 / LAUNCH
Terzi and Nicoli (PNAS Nexus, 2026) gathered the cost of every orbital launch since 1960. Over the whole period, each doubling of the mass launched cut the cost per kilogram by 21%. From 2025's average of $3,868 to low orbit, that rate gives $3,134 in 2030 and $2,048 in 2040.
Since 2010 the curve is far steeper, and the paper's projections use it. With payload growing 13.3% a year, a kilogram to low orbit costs $1,568 in 2030 and $272 in 2040. Growth of 7.5% or 21% a year gives $488 or $113 in 2040.
The curves describe cost per kilogram of capacity; customers pay list prices. SpaceX's rideshare price to sun-synchronous orbit went from $6,000 per kilogram in 2024 to $7,000 in 2026, and $7,500 for 2027. Google's Suncatcher study reaches under $200 per kilogram around 2035 only with about 180 Starship launches a year.
02 / AFTER LAUNCH
On the central launch path, with satellites mass-produced and our node designs, an orbital GPU-hour costs 4.3 times the ground's in 2030, 1.9 times in 2035 and 1.3 times in 2045. It does not come level by 2045, and neither does the most favourable combination in the simulator. These are single nodes: our whole cluster on the design page, with spare nodes, radiation shielding and fixed equipment, costs 5.9 times the ground in 2030.
The GPUs cost the same in both places, but in orbit 15% are spares that cannot be repaired and the node pauses in shadow. That adds $0.31 per GPU-hour, against the $0.49 the ground spends on its building, electricity and upkeep. To break even, launch and satellite together would have to cost under $0.19 per GPU-hour; in 2045 they cost $0.82.
That is why we do not sell orbit on price. Data already in space becomes an answer on board, minutes after capture. New capacity in orbit brings its own power, while new grid connections on the ground take years. Each larger step goes ahead only when launch and spacecraft prices are contracted rather than forecast.
03 / NODE SIZE
Every node carries mass that does not shrink with its power: avionics, attitude control, propulsion and radios. A cluster also needs a spare node, so that losing one does not lose capacity. Small nodes repeat the fixed mass; large nodes need a large spare. Turyshev (JPL, 2026) writes this trade as an objective whose best node size grows with the square root of the cluster's size.
At 300 kg per node, a 10 kW node carries 30 kg of it per kilowatt and a 100 kW node 3 kg. Below 10 kW the fixed mass outweighs the arrays, radiators and compute.
A 1 MW cluster built from 500 kW nodes needs a 500 kW spare: half its mass again. Large nodes also outgrow launchers: one Starship V3 can carry a node of about 2.7 MW with our 2030 design.
With our 2030 design, the lightest node for a 1 MW cluster is about 107 kW, at 31.9 kg per kilowatt; 20 kW nodes are 31% heavier. For 100 MW it is about 1.1 MW. Node-1 is planned at 10–20 kW because it flies alone, on a purchased bus, within rideshare and medium-lift capacity.
Fixed mass: avionics, attitude control, propulsion and radios that every node carries whatever its power; Turyshev (2026) puts these at 135–780 kg for a 100 kW node. The spare node keeps the cluster's capacity when one node fails. Diamonds: published spacecraft for scale.
04 / METHOD AND LIMITS
The cost of a GPU-hour follows our cost model v3.1. Launch and satellite costs are annualised at 10% over five years, with 5% a year added for operations, ground stations and insurance, and divided by 96% availability, the share of the year the node computes. GPUs cost $30,000 per kW in both places, with 15% spares in orbit; each GPU draws 1.2 kW and works 70% of the time. The ground pays $12 per watt for its building over 15 years, 9.0 cents per kWh (US EIA, industrial, 2026) at a PUE of 1.25, and $250 per kW-year of upkeep. Launch to the 650 km dawn–dusk orbit costs 1.45 times the low-orbit price per kilogram. The model reproduces the cost model's published cases, $29.64 per GPU-hour with 2026's best launch and satellites, $7.46 around 2030 and $3.14 at $200 per kilogram against $2.04 on the ground; automated tests check this on every build.
The launch paths reproduce Terzi and Nicoli's projections from their published data to within $2 per kilogram: $1,568 in 2030 and $272 in 2040, with 9,148 t launched in 2030 and 31,889 t in 2040. Their prices are 2024 dollars per kilogram of low-orbit capacity. The satellite cost paths, from $200 to $10 per watt in the mass-production case, and the target of 100 kW of compute per tonne are scenarios, not forecasts. Mass per kilowatt comes from our sizing model on the power-and-heat page: 41, 26 and 17 kg per kW for 2026, 2030 and 2035 designs. Node size follows Turyshev's objective with one spare node per cluster; the fixed mass per node is the main uncertainty, and the scenarios span its published range.
Not modelled: taxes and subsidies, insurance pricing, launch failures, debris and end-of-life disposal, the price customers pay for inference, demand, and the value of speed or sovereignty.
Sources: Terzi and Nicoli, From Sputnik to Starship, PNAS Nexus 5(7), 2026, with its Data S3; Google, Towards a future space-based, highly scalable AI infrastructure system design, 2025; Turyshev, Orbital data centers: spacecraft constraints and economic viability, 2026 and Operations, maintenance and industrial scaling of MW-class orbital data centers, 2026; SpaceX, Form S-1 (June 2026) and rideshare price list (October 2026); US EIA Short-Term Energy Outlook, September 2026; GPU rack prices from DatacenterDynamics (2024) and Wing VC (2026). No launch provider or chip maker is represented as a partner.