Artificial intelligence has a power problem.
As demand for compute continues to increase, access to energy and cooling is becoming a critical constraint, along with the ability to scale both.
That pressure is prompting the industry to reconsider where compute lives, and one of the most extreme concepts is gaining serious attention: orbital data centers.
Building data centers in space is technically possible. Their viability, however, depends on whether they can provide enough value to offset the enormous engineering, operational, and economic penalties they introduce.
Futurum Research estimates global AI workloads may require 300 gigawatts of compute by 2030, far exceeding what existing grids can absorb. New hyperscale data centers can be designed and constructed in as little as 12 to 18 months, but securing grid-connected power often takes three to seven years.
These capacity shortfalls and barriers to expansion are driving interest in alternatives, and space offers distinct advantages.
In low-Earth orbit, solar arrays operate with near-continuous exposure, avoiding the intermittency that limits terrestrial renewables. Orbital systems also avoid many of the siting and resource constraints that slow large infrastructure projects and increase environmental pressures on Earth.
According to Futurum, orbital computing could represent a one trillion-dollar addressable market by 2030.
Orbital computing can range from satellite-based processors that provide onboard compute to massive facilities designed purely for power generation and data processing.
These are fundamentally different propositions, and scale changes everything: the economics, the engineering complexity, the physics, and ultimately the decision of whether to pursue it at all.
At one end of the spectrum, space-based edge computing is already happening. Satellites today generate enormous volumes of data, from high-resolution imagery and synthetic aperture radar to continuous sensor telemetry. Most of that data is still handled the way it was decades ago: collected in orbit and transmitted to ground stations for processing.
That model is increasingly creating bottlenecks. Downlink opportunities require line-of-sight access to ground stations, bandwidth is inherently limited, and modern satellites generate more data than they can reasonably transmit back to Earth.
“Your main bottleneck in space is data rate transmission,” says Andres Ramirez, principal thermal engineer at Aethero, which is building radiation-tolerant computing systems designed to process data in orbit. “If you can process, compress, and even make decisions on the data in space, you don't have to wait for it to come back to Earth.”
Aethero’s systems are not data centers in the terrestrial sense. Instead, they reflect the same architectural shift driving edge computing on the ground: moving compute closer to where data is created.
In orbit, that shift reduces latency, eases pressure on constrained downlinks, and minimizes unnecessary data movement. It also reduces security exposure by limiting how much data must traverse vulnerable communication links.
At the other end of the spectrum are the proposed gigawatt-scale orbital facilities that generate the boldest headlines — and the most daunting economics.
The cost gap between orbital and terrestrial infrastructure is significant. Based on current manufacturing and launch economics, building one gigawatt of orbital data center infrastructure would cost roughly $72.1 billion, Futurum estimates, excluding the cost of the compute hardware itself.
By comparison, a terrestrial off-grid data center comes in at about $16.2 billion per gigawatt. Grid-connected facilities remain the lowest-cost option at roughly $10 billion per gigawatt.
These cost realities explain why orbital computing is not positioned as a wholesale replacement for terrestrial, grid-connected data centers. The economics become more plausible, however, for niche use cases that would otherwise require off-grid deployment, face multi-year grid delays, or demand mission capabilities that ground-based infrastructure cannot provide. National security applications — where resilience, redundancy, and rapid decision-making may justify the premium — represent the most credible near-term use case.
Futurum estimates that roughly one-third of global AI compute spending could be economically justified for deployment beyond Earth by the end of the decade. That projection assumes continued reductions in launch costs and improvements in solar panel efficiency, and it primarily applies to the subset of workloads constrained by power availability, location, or mission requirements.
The economics of orbital computing cannot be separated from the engineering, because every design tradeoff cascades into cost.
Thermal management is perhaps the most daunting constraint. There is no convection in space, which means every watt of heat generated by a processor must be conducted through solid materials and radiated away. At gigawatt scale, orbital data centers would require radiator structures the size of football fields.
“Everything is thermal,” Ramirez says. “And this is a very critical thing in space and a very difficult hurdle to overcome.”
Radiation presents another layer of complexity. Electronics in orbit are exposed to solar particles and cosmic radiation that can cause bit flips, reversing the binary zeros and ones that underpin every computation. Protecting against this requires radiation-hardened components and additional shielding, which adds mass and therefore increases launch cost.
Then there is the question of maintenance. Terrestrial data centers are constantly upgraded; hardware that was state-of-the-art a decade ago has long since been replaced. In orbit, repair and upgrade options range from prohibitively expensive to nonexistent.
Whether the work is done by humans or robots, the logistics and cost of servicing orbital infrastructure remain unsolved problems.
Other challenges compound the picture: vibration and mechanical stress during launch, the growing risk of collisions with debris in an increasingly crowded low-Earth orbit environment, and the complexity of managing power during eclipse periods when solar access is interrupted and systems must rely on battery reserves.
“Orbital computing has to solve physics and economics at the same time,” says Brendan Burke, research director at Futurum Research. “The systems integration decisions being made today will determine who can actually scale it.”
None of these challenges exist in isolation. Thermal behavior cannot be separated from structure, materials, electronics, radiation exposure, or orbital mechanics. Every design decision made to manage heat changes mass. Every change in mass affects launch cost. Every tradeoff in shielding affects electromagnetic performance.
“Cost parity with terrestrial infrastructure means doubling solar panel specific power, getting satellite hardware costs below $5 per watt, and hitting launch costs of $100 per kilogram,” Burke explains. “These are interdependent constraints that converge into a single systems integration challenge.”
Multiphysics simulation helps engineers navigate that complexity. Models can evaluate how heat flows through densely packed electronics, how structures respond to launch vibration, how materials degrade under prolonged radiation exposure, and how closely integrated components interact electromagnetically in vacuum.
“Teams that can model that full complexity before hardware leaves the ground will have a real development advantage,” Burke says. “That’s where platforms like Synopsys, through Ansys, can help move this vision from concept to deployment.”
With compute demand outpacing our ability to scale capacity, it’s clear the future of AI infrastructure will not follow a single path. Traditional, grid-connected data centers will invariably be supplemented by new approaches.
Whether those approaches involve off-grid facilities or off-planet deployment, the underlying challenge is the same: designing, simulating, optimizing, and validating these systems against the full complexity of their physical environment — before they are built and deployed.
The teams that can do this effectively will turn novel concepts into economically viable alternatives.
An edited version of this article originally appeared in ENGINEERING.com