Musk Refutes Concerns Over Orbital AI Cooling
Elon Musk stated that those who deem orbital AI unfeasible due to cooling issues lack the foundation to make such conclusions if they do not understand heat dissipation per unit area of heat sinks, coolant temperatures, and maximum GPU operating temperatures.
The core of orbital computing is not whether space is cold, but whether the waste heat generated by the computing load can be stably transferred from the GPU, HBM, network, and power components to the external heat sink via cold plates, thermal structures, or pumped liquid circuits, and radiated into space through infrared radiation. Vacuum eliminates convection but does not prevent conduction, heat pipes, and liquid cooling circuits from functioning inside the equipment.
The heat dissipation capacity of heat sinks mainly depends on effective radiative area, surface emissivity, and temperature, approximately following q=εσ(T4−Ts4). This means that an increase in heat sink temperature will raise the heat dissipation per unit area to the fourth power, allowing for a reduction in the required heat sink area; the trade-off is that the coolant, packaging, chip junction temperature, leakage, and long-term reliability face higher pressure.
Taking the NVIDIA Vera Rubin NVL72's configuration of about 136 kW as an example, an idealized calculation based on a 45°C liquid cooling supply temperature and 0.90 emissivity shows that a single-sided effective heat dissipation area of about 260 square meters is required; considering solar radiation, Earth's infrared, albedo, field of view, pollution, and lifespan decay, the actual design margin will continue to expand.
SpaceX and NVIDIA have disclosed the development of the space-optimized Vera Rubin NVL72 system, with Musk previously stating that the first batch of related AI satellites is targeted for launch in Q4 2027; reports mention a peak computing power of up to 250 kW, averaging about 175 kW, but the heat sink area, cooling circuit, and sustained peak load capacity have not yet been fully disclosed by SpaceX.
In market mechanisms, buyers of orbital AI are not merely purchasing GPU computing power, but are betting that solar power, launch capabilities, inter-satellite laser links, and thermal management can collectively reduce marginal computing costs for operators and capital. Beneficiaries include companies that master launch, satellite assembly, proprietary networks, and high-temperature heat dissipation materials; those under pressure are large data center projects relying on ground power grids, cooling water, land, and permitting expansion. Project value is determined by the quality of the heat sink, deployment reliability, hardware lifespan, network replenishment frequency, and unit effective computing cost, rather than the narrative of "natural cooling in space."
Source: Public Information
ABAB AI Insight
SpaceX's historical path is not to first prove a single technology's perfection before entering large-scale deployment, but to bind repeatable production, low-cost launches, and in-orbit iteration into a system: Falcon 9 reduces the frequency cost of entering space through first-stage reuse, while Starlink rapidly lays out a communication network by reusing high launch cadence. Orbital AI continues this same logic—if launch, satellite manufacturing, and replacement are internalized by the company, the expensive and large heat sink is no longer just a burden on the spacecraft, but can become infrastructure that is diluted with the mass production and replenishment of satellites.
The key to the capital path is not to move a ground server into orbit, but to replace the ground data center's power grid access, substations, diesel backup, cooling towers, water rights, land, and local approvals with launch payloads, solar wings, energy storage, thermal control circuits, and laser communication links. NVIDIA provides high-density computing modules, while SpaceX masters launch, satellite platforms, and the Starlink backhaul network; this combination compresses capital expenditures originally divided among utilities, real estate developers, data center operators, and network operators into a vertically integrated launch and network system.
Historically, Iridium went bankrupt due to high satellite deployment costs and limited user scale, later restructuring to form a specific communication market; Starlink has improved this constraint by relying on reusable rockets and large-scale consumer broadband subscriptions. Orbital AI is more demanding than communication constellations: the value of communication satellites mainly comes from spectrum coverage and terminal access, while computing satellites must continuously address high heat flux density, radiation damage, GPU update cycles, non-repairable components, and workload backhaul. The current stage is closer to "system feasibility validation," rather than a mature expansion period that can directly compete with ground AI clusters for unit computing power.
Essentially, this is about industrial chain restructuring: the profits and bottlenecks of ground AI focus on electricity, land, grid connection, cooling, and data center construction, while orbital AI attempts to rewrite scarce factors as launch capability, satellite production capacity, heat dissipation materials, space networking, and in-orbit replacement capability. Heat dissipation is not a physical paradox sufficient to negate a solution, but it is a core constraint determining system quality score, sustained load, satellite lifespan, and economics; those who can reliably dissipate every watt of waste heat at lower quality and higher temperatures are more likely to convert launch scale into sellable computing power.
ABAB News · Cognitive Laws
Space does not lack cold sources; it lacks low-cost heat dissipation paths.
Technical infeasibility often arises from unknown parameters, not physical prohibitions.
The deeper the vertical integration, the more infrastructure can become a product.