Orbital Compute Capital Cost

What is the capital cost per kW_IT for orbital compute?

What is the capital cost per kW_IT ($/kW_IT) for deploying AI compute in orbit, including launch, platform manufacturing, and GPU hardware?

Answer

Orbital capex ranges from 54,993 $/kW_IT in 2026 (optimistic) down to 29,718 $/kW_IT by 2040 (optimistic), with the central estimate falling from 138,614 $/kW_IT (2026) to 49,122 $/kW_IT (2040). The conservative scenario starts at 263,383 $/kW_IT and declines to 98,787 $/kW_IT by 2040.

The capex composition undergoes a dramatic structural shift over time. In 2026, launch cost dominates the central and conservative scenarios, while GPU hardware already dominates the optimistic scenario. By the mid-2030s, declining launch costs cause GPU hardware to become the dominant capex component in all scenarios -- a transition that fundamentally limits further capex reduction, since GPU cost is shared with terrestrial deployments and cannot be eliminated through space-specific innovation.

Inputs

Analysis

Capex Formula

Total orbital capex is the sum of three components:

1. Launch cost = total_satellite_mass (kg/kW_IT) x launch_cost_per_kg ($/kg)
2. Orbital GPU cost = gpu_cost_per_kw ($/kW_IT) x gpu_cost_premium (multiplier)
3. Platform manufacturing cost = manufacturing cost for non-compute satellite hardware ($/kW_IT)

Launch Cost: The Declining Giant

Launch cost is the most time-sensitive component and the one with the widest scenario range:

In the central scenario, launch cost falls by 97% from 83,239 $/kW_IT (2026) to 3,022 $/kW_IT (2040), driven by Starship achieving increasing reuse rates and flight cadence. This decline is the single most important trend in the entire orbital compute cost model.

In the optimistic scenario, launch cost becomes negligible by 2030 (1,655 $/kW_IT), reflecting Starship at 20+ reuses with internal transfer pricing. By 2040, launch is just 358 $/kW_IT -- less than 1% of total capex.

In the conservative scenario, launch remains a major cost even in 2040 at 20,437 $/kW_IT, reflecting limited Starship reuse and the combination of high $/kg pricing with the heavy satellite mass of the conservative mass budget (65.3 kg/kW_IT).

GPU Cost: The Irreducible Floor

Orbital GPU cost is the product of baseline GPU cost and the space adaptation premium:

GPU cost is time-invariant in the model. While GPU architectures improve rapidly (3-4x performance per generation), the cost per kW_IT has remained roughly stable because each generation increases both performance and power density while maintaining similar $/kW pricing. This is a critical observation: moving to space does not reduce GPU cost, which represents the single largest cost component of AI compute both on Earth and in orbit.

The Capex Composition Shift

The most important structural finding is how capex composition changes over time:

2026 Central: Launch (83,239) dominates at ~60% of total capex (138,614). GPU cost (37,375) is ~27%, and platform manufacturing ($18,000) is ~13%.

2030 Central: Launch drops to 19,728 (~28% of total), GPU rises to ~52% (37,375 of 71,778), and platform manufacturing ($14,675) is ~20%.

2040 Central: Launch is just 3,022 (~6% of total). GPU is the dominant cost at ~76% (37,375 of 49,122). Platform manufacturing ($8,725) accounts for ~18%.

This shift has a profound implication: as launch costs decline, further orbital capex reduction becomes increasingly difficult because the dominant remaining cost (GPU hardware) is inherent to AI compute and identical whether deployed on Earth or in orbit. The space-specific costs (launch + platform) compress from ~73% of capex in 2026 to ~24% by 2040 in the central case, leaving GPU cost as an irreducible floor.

Total Capex Trajectories

The optimistic capex converges toward ~29,718 $/kW_IT by 2040, essentially reaching a floor set by GPU cost (27,000) plus the 2040 optimistic platform manufacturing cost ($2,360) plus negligible launch (358). This floor is modestly above the GPU cost alone, reflecting the irreducible overhead of putting compute hardware into orbit — though the premium shrinks as platform manufacturing costs decline through learning curves.

Comparison to Terrestrial Capex

Terrestrial compute capex is GPU cost plus infrastructure:

• Optimistic: $25,000 (GPU) + $8,000 (infra) = $33,000/kW_IT
• Central: $32,500 + $12,500 = $45,000/kW_IT
• Conservative: $40,000 + $20,000 = $60,000/kW_IT

By 2040, orbital capex in the optimistic scenario (29,718) approaches the terrestrial optimistic capex ($33,000), while the central orbital capex (49,122) exceeds the central terrestrial capex ($45,000) by ~14%. The conservative orbital capex (98,787) remains nearly double the conservative terrestrial figure ($60,000).

The key insight is that orbital capex includes a structural premium over terrestrial capex driven by three factors: (1) the GPU cost premium (8-30%), (2) platform manufacturing cost ($8,000-$35,000) which replaces but does not fully offset terrestrial infrastructure cost ($8,000-$20,000), and (3) residual launch cost which has no terrestrial equivalent. Only in the optimistic long-term scenario does this premium shrink to a level where operational savings (zero terrestrial electricity purchases in orbit) could potentially offset it.