Somewhere between science fiction and a spreadsheet that actually closes, a handful of startups are making a case that the next great frontier for computing isn’t a warehouse in Iowa or a hyperscale campus in northern Virginia. It’s low Earth orbit.
The idea sounds absurd on first hearing. Launch thousands of tons of server racks into space, where they’d orbit the planet processing artificial intelligence workloads, cooled by the vacuum itself, powered by unfiltered sunlight. No land permits. No water rights battles. No angry neighbors complaining about the hum of diesel generators at 2 a.m.
And yet.
A growing number of companies — Lumen Orbit, Axiom Space spinoffs, and several stealth-mode ventures — are arguing that the economics of orbital computing are not only plausible but potentially inevitable, driven by the explosive demand for AI training capacity that terrestrial infrastructure simply cannot satisfy fast enough. The question isn’t whether the physics works. It does. The question is whether the dollars work, and whether they work soon enough to matter.
As Ars Technica explored in the first installment of a detailed series on the subject, the instinctive reaction from most technologists is disbelief. “There’s no way this is economically viable, right?” — the literal title of their report — captures the prevailing skepticism. But the publication’s analysis, which walks through launch costs, power generation, thermal management, and latency constraints with unusual rigor, arrives at a more nuanced conclusion than the headline suggests.
The core argument rests on a few converging trends that, taken individually, seem incremental but together start to look like something genuinely different.
First, launch costs. SpaceX’s Falcon 9 already brings the price of putting a kilogram into low Earth orbit down to roughly $2,700, according to industry estimates. Starship, if it achieves even a fraction of its reusability targets, could push that figure below $100 per kilogram. That single variable changes the math on orbital infrastructure more dramatically than anything else. When it cost $54,500 per kilogram on the Space Shuttle, nobody was seriously proposing shipping HP ProLiant servers to orbit. At $100 per kilogram, the conversation shifts from “why would you?” to “why wouldn’t you, given certain constraints?”
Those constraints are real and they’re substantial. But they’re not the ones most people assume.
Latency is the objection that comes up first in virtually every conversation about space-based computing. And it’s the easiest to dispatch. Low Earth orbit sits roughly 550 kilometers above the surface. The round-trip signal time is approximately 4 to 8 milliseconds, depending on orbital altitude and ground station placement. That’s worse than a co-located server, obviously. But it’s comparable to — and in some cases better than — the latency between, say, a company’s offices in Chicago and a cloud region in Virginia. For AI training workloads, which are batch-processed and latency-tolerant, the delay is essentially irrelevant. Nobody’s proposing to run high-frequency trading from orbit. Not yet, anyway.
Power is where the argument gets genuinely interesting. A solar panel in space receives roughly 1,361 watts per square meter of solar irradiance — the solar constant, unfiltered by atmosphere, clouds, or night cycles (with the caveat that orbital night does occur for portions of each 90-minute orbit in LEO). On Earth, a solar panel in the Arizona desert might average 6 to 7 peak sun hours per day. In orbit, with proper orientation, you can capture energy for roughly 60 of every 90 minutes. The energy density advantage is enormous, and it comes without the land acquisition, permitting, and grid interconnection delays that are currently bottlenecking terrestrial data center expansion across the United States and Europe.
Those bottlenecks are not trivial. According to reporting from multiple outlets in recent months, data center developers in Virginia’s Loudoun County — the densest concentration of data centers on the planet — are facing 4- to 7-year waits for new grid connections. Dominion Energy, the regional utility, has publicly acknowledged that demand from data centers is outstripping its ability to build generation and transmission capacity. Similar constraints are emerging in Dublin, Amsterdam, Singapore, and Frankfurt, where local governments have imposed moratoriums or severe restrictions on new data center construction due to power grid strain and water usage concerns.
Water. That’s the other variable that rarely makes headlines but dominates the private conversations of data center operators. Evaporative cooling — the standard thermal management approach for large-scale facilities — consumes staggering volumes of water. A single large data center can use millions of gallons per day. In drought-prone regions, this has become a political and environmental liability. Google, Microsoft, and Meta have all faced public scrutiny over water consumption at their facilities.
Space solves this problem with elegant brutality. The vacuum of space is, thermodynamically, an almost perfect heat sink — not because it’s cold in the conventional sense, but because radiative cooling works extraordinarily well when there’s no atmosphere to conduct heat back. A properly designed radiator panel in orbit can reject waste heat far more efficiently than any terrestrial cooling system. No water. No cooling towers. No chillers consuming 30 to 40 percent of a facility’s total energy budget.
The Real Question: Can the Numbers Actually Close?
This is where the analysis sharpens. The skeptics aren’t wrong to focus on cost — they’re just sometimes wrong about which costs matter most.
The capital expenditure of launching hardware to orbit is significant but, as noted, declining rapidly. The more important figure is the total cost of ownership over the operational life of the hardware. And here, orbital proponents make a case that deserves serious scrutiny rather than reflexive dismissal.
Consider the fully loaded cost of a new terrestrial data center. Land acquisition. Construction, which now routinely takes 2 to 4 years for a large campus. Grid interconnection, including the cost of dedicated substations and transmission upgrades that utilities increasingly pass through to customers. Backup generation — diesel or natural gas generators that sit idle 99 percent of the time but must be provisioned for 100 percent of critical load. Cooling infrastructure. Water rights or municipal water contracts. Ongoing energy costs, which in many markets now exceed $0.10 per kilowatt-hour for industrial customers and are rising. Property taxes. Physical security. Staffing.
An orbital data center eliminates or radically reduces many of these line items. No land. No construction permits. No grid interconnection. No water. No cooling energy penalty. No property taxes. The energy source — solar — is free after the capital cost of the panels, and the panels themselves benefit from the higher irradiance and lack of atmospheric degradation.
What you get instead is a different set of costs: launch, satellite bus construction, radiation hardening of components, orbital maintenance, ground station networks for uplink and downlink, and the eventual cost of deorbiting hardware at end of life. The Ars Technica analysis walks through these in detail, noting that radiation hardening — once a prohibitively expensive requirement — has become significantly cheaper as commercial satellite operators have driven demand for radiation-tolerant components. Modern server-grade chips from companies like AMD and Nvidia are already designed with error-correction capabilities that provide a baseline of radiation tolerance, though additional shielding and redundancy would still be required for orbital deployment.
The operational life of server hardware is another factor that cuts in favor of orbital deployment, counterintuitively. On Earth, data center operators typically refresh server hardware on a 3- to 5-year cycle, driven partly by efficiency improvements in new generations of chips and partly by physical degradation. In orbit, the absence of humidity, dust, and corrosive atmospheric gases could extend hardware life, though radiation damage to semiconductors works in the opposite direction. The net effect is debatable, but several analyses suggest that a 3- to 5-year orbital life is achievable with current technology.
Lumen Orbit, one of the most visible companies in this space, has been making the rounds with investors and potential customers. The company’s pitch, as reported by multiple technology publications, centers on AI training workloads — specifically, the large-scale matrix multiplication operations that dominate the training of large language models and other foundation models. These workloads are embarrassingly parallel, latency-tolerant, and energy-hungry. They’re the perfect candidate for a computing platform that offers abundant power and cooling but imperfect connectivity.
The business model that’s emerging looks less like a traditional colocation or cloud provider and more like a specialized compute utility. A company needing to train a massive AI model would essentially rent orbital compute capacity for a defined period — days, weeks, or months — uploading training data and model parameters, then downloading the trained model weights when the job completes. The data volumes involved are large but manageable: a state-of-the-art language model’s training dataset might be tens of terabytes, and the resulting model weights a few hundred gigabytes. Optical inter-satellite links and high-bandwidth ground stations — technologies already proven by SpaceX’s Starlink constellation — could handle these transfers.
But let’s not get carried away. The engineering challenges remain formidable.
Vibration during launch is a serious concern for densely packed server hardware. Thermal cycling as satellites pass in and out of Earth’s shadow — roughly every 45 minutes — creates mechanical stress on solder joints and component interconnects. Micrometeorite impacts, though statistically rare for any individual satellite, become a real actuarial consideration when you’re talking about constellations of hundreds or thousands of compute nodes. And then there’s the regulatory environment: orbital debris mitigation requirements, spectrum allocation for data links, and the still-evolving international framework for commercial space operations.
There’s also the question of who builds the hardware. Today’s servers are designed for terrestrial data centers — standard 19-inch rack form factors, air or liquid cooling, components rated for benign thermal and vibration environments. An orbital server would need to be fundamentally redesigned: different form factors, different thermal interfaces, different power distribution, different mechanical packaging. This isn’t impossible — the satellite industry has been building space-rated electronics for decades — but it adds cost and complexity that the optimistic projections sometimes gloss over.
So where does this leave us?
The honest answer is that orbital data centers are probably not viable today, in 2025, for general-purpose computing. The launch costs haven’t fallen far enough. The hardware isn’t purpose-built. The ground station infrastructure for high-bandwidth data transfer isn’t dense enough. And the demand for AI compute, while genuinely enormous, is still being met — barely, expensively, with long wait times — by terrestrial facilities.
But the trajectory matters more than the snapshot. Every variable in the equation is moving in the right direction for orbital computing. Launch costs are falling on a steep curve. Solar panel efficiency is improving. Radiation-tolerant chip design is advancing. Optical communication bandwidth is increasing. And terrestrial constraints — power grid limitations, water scarcity, permitting delays, community opposition — are tightening, not loosening.
The AI demand curve is the wild card. If the appetite for training compute continues to grow at anything like the pace of the last three years — and most credible forecasts suggest it will — then terrestrial infrastructure alone may genuinely prove insufficient within the next decade. Not because we can’t build more data centers on Earth. We can. But because we can’t build them fast enough, in enough places, with enough power and water, to keep pace with demand.
That’s the opening for orbital computing. Not as a replacement for terrestrial data centers but as a supplement — an overflow valve for the most energy-intensive, latency-tolerant workloads that are currently constrained by the physical limitations of earthbound infrastructure.
It’s a narrow wedge. But narrow wedges have a way of widening.
The investors putting money into this sector — and there is real money flowing, though exact figures are closely held — are betting on a future where the cost curves cross within 5 to 10 years. They may be wrong. Space ventures have a long and colorful history of overpromising and underdelivering. But they may also be early, which in venture capital is functionally the same as being right, just with worse cash flow.
One thing is certain: the conversation has shifted. Five years ago, orbital data centers were a punchline. Today they’re a business plan. Whether they become a business remains to be seen. But the physics is real, the demand is real, and the terrestrial constraints are real. That’s more than most ambitious technology concepts can claim at this stage.
And for those of us who grew up watching shuttle launches on grainy televisions in living rooms across the Midwest, wondering what all that expensive rocketry would eventually be good for — well. Maybe it’s good for training the AI models that will define the next era of computing. Stranger things have happened. Though not many.
WebProNews is an iEntry Publication