Morgan Stanley believes thermal management for SpaceX orbital AI compute is not a physical bottleneck

Overview

This report is a Morgan Stanley thematic study on SpaceX's orbital AI computing capability from both technical and valuation perspectives. The core question is whether the lack of convective cooling from air or water in space would make data-center-style AI satellite computing infeasible from a thermal management standpoint. The authors believe this concern is overstated. Satellite thermal management has relied on radiative cooling since the 1960s, and the ISS also rejects heat through ammonia fluid loops and large radiator panels. The report deliberately narrows the discussion to heat rejection and further shifts the investment debate toward cost, reliability, scaling speed, and power deployment timelines.

Core views

There are three core views. First, thermal management for orbital AI compute can be modeled using the Stefan-Boltzmann law, higher chip operating temperatures, high-emissivity coatings, double-sided radiators, and fluid loops, and should not be seen as a physical showstopper. Second, if SpaceX can combine Starship, satellite manufacturing, space-based solar power, and optical communications networks, it could become one of the most scalable AI compute infrastructure providers in the 2030s. Third, while orbital compute has higher initial capex than terrestrial compute, the absence of substantial opex for power, cooling, land, personnel, and maintenance means it could gain a cost advantage versus current industry terrestrial compute on a fully loaded basis starting in 2031, and drive long-term EBIT margins above 50%.

Analysis framework

The report uses a bottom-up engineering and financial modeling approach. On the engineering side, the authors estimate deployable compute capacity based on public AI1 satellite concepts, Starship generational evolution, radiator area and mass, solar array area and mass, chip operating temperatures, effective LEO background temperature, launch capacity, and satellite lifespan. On the financial side, the report converts orbital compute capex into annual cost per watt based on a 5-year useful life and compares it with industry terrestrial Blackwell compute costs, while also incorporating differences in power, cooling, maintenance, land, and operating costs to derive long-term cash cost and EBIT margin.

Methodology notes

• physics_engineeringStefan-Boltzmann Law

  Radiative heat dissipation and the fourth-power relationship with temperature

  The report uses the Stefan-Boltzmann law to estimate the radiator area required for a given heat load; radiated energy is proportional to the fourth power of absolute temperature, so the higher the chip operating temperature, the greater the heat dissipation efficiency per unit area.

• bottom_up_modelOrbital Compute Engineering Build

  Engineering assumptions for orbital AI satellites

  The model breaks down thermal radiators, solar arrays, and compute payloads, and separately estimates temperature, emissivity, LEO background temperature, double-sided heat rejection, area density, solar efficiency, degradation rate, and satellite mass.

• cost_comparisonAll-in Cost Parity

  All-in cost comparison between orbital and terrestrial compute

  The report does not compare capex alone. It includes depreciation, power, cooling, land, personnel, maintenance, and site operating costs, and argues that the main ongoing cost of orbital compute is upfront capex depreciation.

• Valuation methodsMorgan Stanley ModelWare

  Morgan Stanley internal model framework

  The report discloses that, unless otherwise stated, the relevant metrics are based on the Morgan Stanley ModelWare framework.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• SpaceX (SPCX.O / SPCX US)

  Research target

  Strengths

  It has Starship launch capability, satellite manufacturing experience, Starlink technology foundations, and potential synergies with space-based solar power and optical communications networks; the report believes it could become one of the most scalable AI compute infrastructure providers in the 2030s.

  Weaknesses

  Orbital AI compute is still at an early stage, with high initial capex, significant room for iteration in engineering assumptions, and satellites that cannot be manually maintained and must rely on remote redundancy design.

  Comparison

  The report compares orbital compute with terrestrial AI data centers and argues that terrestrial compute is constrained by power, cooling, land, permitting, personnel, and maintenance costs, while the main ongoing cost of orbital compute is upfront capex depreciation.

  Risks

  Radiation damage, orbital debris, launch cadence, the chip supply chain, optical communications reliability, satellite lifespan, regulation, and physical security could all affect commercialization.

• Terrestrial AI data centers

  Comparable infrastructure

  Strengths

  They are technologically mature, easy to maintain, and have more developed supply chains and operational systems, making them well suited for workloads such as training that require high reliability and near-ground maintenance.

  Weaknesses

  They face constraints in grid interconnection, cooling, water resources, land, permitting, community acceptance, and construction timelines.

  Comparison

  The report argues that in the long run terrestrial compute may mainly support model training, while most inference compute could migrate to orbit.

  Risks

  If terrestrial data center costs decline rapidly or supply bottlenecks ease, the scarcity premium and cost advantage of orbital compute could weaken.

Key data

• Rating and target priceOverweight; target price $300.00; 2026-07-20 closing price $119.85Equivalent to about 150.3% implied upside.
• Initial orbital compute deployment160 MW in 2028The report assumes SpaceX begins deploying orbital compute in 2028.
• Orbital compute scale forecastAbout 2.7 GW in 2030; 21.2 GW in 2032; 111 GW in 2035; 364 GW in 2040The report states orbital compute accounts for 58% of total deployed compute in 2032, 88% in 2035, and 96% in 2040.
• Cost inflection pointIncremental orbital compute about $6.5/watt/year in 2031Below the current industry terrestrial Blackwell cost of about $6.8/watt/year cited in the report.
• Long-term cash costBelow $9/watt in 2035; $3.7/watt in 2040The metric includes capex and non-depreciation/amortization operating cash costs.
• AI EBIT marginAbout 42% in 2028; about 50%+ in the 2030sThe report believes long-term savings will be partly offset by declining revenue per watt.
• Chip operating temperature assumptionAbout 360K/90°C for the first generation; about 370K/97°C as a long-term targetHigher temperatures can improve radiator efficiency and reduce radiator mass.
• Radiator area for 1 kW of computeAbout 0.57 square metersAssumes about 90°C, white thermal-control coating emissivity of about 0.90, and double-sided radiation.
• AI1 satellite assumption150 kW peak power; 120 kW compute power; about 2.1 tonsThe report says the first-generation design is mainly based on the AI1 Sat concept recently shared by Elon Musk/SpaceX.
• Radiator scale assumptionAI1: 110 m²/767 kg; AI2: 266 m²/1,328 kg; AI3: 580 m²/2,030 kgArea is per-side area, and mass is estimated based on radiator area density.
• Solar array assumptionAI1: 605 m²/1,028 kg; AI2: 1,523 m²/1,979 kg; AI3: 3,218 m²/2,897 kgThe report assumes dawn-dusk SSO provides near-continuous sunlight, with a solar constant of about 1,361 W/m².

Impact & implications

If the report's assumptions hold, the investment narrative for SpaceX would extend beyond launch, Starlink, or satellite manufacturing to an AI infrastructure platform. Orbital compute could alleviate the constraints terrestrial data centers face in grid interconnection, land, water resources, permitting, and construction timelines, and may enable parallel capacity expansion through reusable launch and satellite manufacturing. For investors, once the thermal management debate is de-emphasized, the real factors to track are Starship launch capability, satellite hardware iteration, GPU supply, space-based solar power, optical communications, reliability, and the pace at which unit economics are realized.

Risks

• Orbital compute faces radiation damage risk, especially when operating outside the protection of Earth's magnetosphere, which could affect chip lifespan and reliability.
• Low Earth orbit has a growing orbital debris risk, which could affect satellite safety, insurance costs, and operational continuity.
• Orbital assets lack on-site maintenance conditions, so failures must be addressed through remote operations, redundancy design, and replaceability.
• Launch cadence may initially limit deployable compute capacity, and the report notes that early single flights may only add single-digit MW of capability.
• The AI chip supply chain shares the same constraints as terrestrial data centers, and insufficient GPU supply could slow orbital compute deployment.
• The engineering assumptions for radiators, solar arrays, optical communications, and larger satellite platforms still require validation from future SpaceX technical details.
• If orbital compute fails to reduce costs or improve time-to-power as expected, current valuation and long-term margin assumptions may come under pressure.

What to watch

• Technical specifications subsequently disclosed by SpaceX for AI1 Sat, AI2 Sat, or AI3 Sat, especially power, mass, radiator area, and compute density.
• Launch capability, reusability, cost reduction, and deployment cadence of Starship V3 and later versions.
• Whether orbital compute can achieve its first deployment in 2028 and whether it approaches the projected 2.7 GW path by 2030.
• Whether chip temperature, radiator materials, emissivity, coolant, and double-sided heat rejection design for orbital AI satellites align with the model assumptions.
• Measured data on the GPU supply chain, optical communications networks, space-based solar array efficiency, and satellite lifespan.
• Pricing of compute procurement contracts between SpaceX and neocloud or AI customers, and whether they reflect a premium relative to industry averages.
• Constraints on large-scale orbital data center deployment from regulation, orbital debris, spectrum, physical security, and international space governance.