Research · AI Infrastructure
The ground still wins the commodity-compute race through at least 2035, and remains the base case for 2040.
The ground remains the commodity-compute benchmark through at least 2035 and remains the base case for 2040. The decisive comparison is not orbit versus today's worst grid bottleneck. It is orbit versus the terrestrial frontier that is already adapting.
The takeaway
Space-based compute has real niches: orbital edge processing, defense and sovereignty workloads, grid-constrained premium compute, space-to-space services, and brand or strategic optionality. But it does not yet clear the cost-per-watt hurdle for bulk AI compute. The terrestrial frontier is moving on several fronts at once: secondary power-rich markets, utility-backed transmission, large-load tariffs, behind-the-meter generation, batteries and flexible load, direct-to-chip cooling, dry/hybrid heat rejection, reclaimed water, modular construction, and rapid silicon refresh. The orbital case has to beat that moving baseline, not a stalled substation request in a saturated market.
01
The orbital-compute question exists because the terrestrial compute buildout has become an energy-infrastructure problem.
LBNL's 2024 data-center energy report estimates U.S. data-center electricity consumption rose to 176 TWh in 2023 and could reach 325–580 TWh in 2028, 6.7%–12% of U.S. electricity. LBNL's Queued Up work separately reports nearly 2,600 GW of generation and storage capacity seeking grid connection at the end of 2023, with most projects that enter the queue ultimately withdrawn. That is not a marginal facilities-management problem; it is a generation, transmission, equipment, rate-design, and land-use problem.
Equipment lead times make the constraint tangible. CISA's National Infrastructure Advisory Council reports 80–210 week lead times for large transformers; DOE's large-power-transformer resilience report frames extended replacement lead times as a grid-resilience issue with maximum lead times reaching as much as 60 months in some contexts. The implication is the same: transformer procurement can set the schedule.
Local politics are tightening too. Loudoun County moved to designate data centers as conditional or Special Exception uses in areas where they had previously been allowed by right. Virginia's JLARC has identified data-center load as a driver of generation and transmission buildout difficulty, customer-cost exposure, and backup-generator emissions.
This is the strongest version of the orbital argument: terrestrial compute needs land, water, generators, substations, transmission, transformers, local permits, local political consent, and sometimes years of queue exposure. Orbit offers continuous solar flux, no local cooling-water withdrawal, no county zoning hearing for the compute platform, and a new venue for capacity growth. That strongest version still has to answer the only question that matters: can orbit deliver usable compute watts, at the relevant hardware vintage, below the cost and schedule of the improving ground-side frontier?
02
Three orbital scenarios against a 620 MW terrestrial baseline. The frontier moves on both sides, but terrestrial keeps winning through 2040.
The reference case is a 1 GW facility-load AI campus with roughly 620 MW usable IT load, 85% utilization, a three-year build, and 12 operating years. The model assumes approximately $24.8 billion of initial terrestrial capex, about $18.6 billion for IT-side (accelerators, servers, storage, networking, racks, integration at ~$30/W usable IT) and about $6.2 billion for land, building, power, cooling, and interconnection (~$10/W usable IT). Annual operating cost is modeled around $1.6 billion, with major IT refreshes in years 7 and 11. Modeled terrestrial PV cost: $46.5 billion, or roughly $12.6 per present-value usable IT-watt-year.
The orbital 2030 near-term case assumes near-term launch economics, high spacecraft manufacturing cost, and conservative mass for the full orbital compute module, compute hardware, power generation and conversion, radiators, bus, communications, station-keeping, shielding, redundancy, deorbit hardware, ground segment, plus spares, insurance, and refresh penalties. Modeled PV cost: $296.1 billion, or roughly $80.3 per usable IT-watt-year, about 6.4× terrestrial.
The 2035 steelman uses aggressive reusable-launch assumptions and spacecraft mass-production learning. Modeled PV cost: $147.5 billion, or roughly $40.0 per usable IT-watt-year, about 3.2× terrestrial. This is the most important case because it gives the orbital bull thesis real credit. It still does not win.
The 2040 radical case uses very low launch cost, lighter mass per usable IT watt, lower spacecraft dollars per kilogram, and lower operating and insurance cost. Modeled PV cost: $87.2 billion, or roughly $23.7 per usable IT-watt-year, about 1.9× terrestrial. The radical 2040 case still needs an additional roughly 47% PV-cost reduction to match the terrestrial baseline. The frontier can move on both sides, but terrestrial also benefits from chip efficiency, power-market adaptation, cooling gains, modular construction, and global siting arbitrage.
Bearing Labs scenario model
03
What the public record makes clear about orbit-versus-ground today.
01
The terrestrial bottleneck is real, but it is not static
LBNL reports U.S. data centers consumed about 176 TWh in 2023 and could consume 325–580 TWh in 2028, or 6.7%–12% of U.S. electricity, and shows nearly 2,600 GW of generation and storage seeking grid connection at year-end 2023. Those constraints explain the orbital pitch. They do not prove it. Terrestrial is responding through secondary-market siting, large-load tariffs, behind-the-meter generation, batteries, grid-enhancing technologies, virtual power plants, and early transformer reservation.
02
Modeled terrestrial parity is roughly $12.6 per usable IT-watt-year
A 1 GW facility-load terrestrial frontier campus modeled as roughly 620 MW of usable IT load produces a PV cost intensity of about $12.6 per usable IT-watt-year under the stated assumptions. A near-term orbital case at $80.3 is about 6.4× the terrestrial baseline; an aggressive 2035 orbital steelman at $40.0 is about 3.2×; even a radical 2040 low-cost case at $23.7 remains about 1.9×. An orbital stack has to deliver the same usable compute for less than roughly $46.5 billion PV cost in this reference case, launch, spacecraft, power, radiators, communications, spares, refresh, insurance, deorbit, and ground segment included.
03
Launch cost matters, but it is not the whole problem
A NASA paper cites a Starship-style aspiration around $10/kg to orbit, a figure that should be treated as aspirational rather than bankable. Once launch price falls below roughly the low-hundreds of dollars per kilogram, non-launch costs dominate: flight-qualified compute, solar power, power conversion, radiators, spacecraft bus, optical communications, station-keeping, insurance, spares, ground systems, and hardware refresh. The cheapest credible orbital story is therefore a mass-manufactured module story, repeatable compute, power, thermal, communications, and bus units produced at satellite-factory scale, not merely a cheaper-rocket story.
04
Orbit's cleanest advantage is water; its hardest problem is heat
A 1 GW terrestrial IT load using a hybrid evaporative profile of 0.40–1.20 L/kWh-IT consumes roughly 3.5–10.5 billion liters per year, or about 2.5–7.6 million gallons per day. Orbit avoids that local water burden. But in orbit, essentially all IT power still becomes heat, and heat has to be radiated. A first-order Stefan–Boltzmann sizing for 1 GW of IT heat gives roughly 2.6 million m² of radiator at 300 K, 2.0 million m² at 320 K, 1.4 million m² at 350 K, or 0.8 million m² at 400 K, before mass, deployment, degradation, redundancy, orientation, solar absorption, albedo, and planetshine penalties.
05
Space is regulatory and political arbitrage, not escape
Terrestrial data centers face zoning, stormwater, wetlands, air permits, utility service, transmission, and community politics. Orbital compute shifts the forum to FCC Part 25 satellite licensing, ITU coordination, FAA Part 450 launch licensing, FCC orbital-debris rules, NOAA remote-sensing licensing if imaging is involved, and ITAR/EAR export controls. Those are not trivial substitutes for county hearings, they are different regulatory regimes with different timelines: 9–18 months for a small demo payload on a conventional spectrum and licensed vehicle, 24–60 months for a commercial LEO compute constellation, and longer for novel high-power, large-constellation architectures with contested spectrum and debris concerns.
04
Three waves of orbital compute today, status, not hype.
Wave 1 · Orbital edge processing is real
In-orbit computing is not speculative. NASA describes HPE sending a high-performance computer to the International Space Station in 2017, and documents SpaceCube onboard processors and the SCENIC system launched to the ISS in March 2023. This matters for Earth-observation filtering, onboard autonomy, defense and intelligence workloads, and space-to-space services. It does not prove that orbital hyperscale AI training is economic, the difference between onboard edge processing and a 1 GW AI data-center substitute is orders of magnitude.
Wave 2 · Limited orbital GPU / hosted-compute is emerging, not proven
Axiom Space is a credible LEO infrastructure actor: NASA has described its commercial station pathway, including a Payload, Power, and Thermal Module that may support a free-flying Axiom Station as early as 2028. Kepler Communications is relevant to orbital communications infrastructure rather than compute capacity, the FCC granted U.S. market access for a 140-satellite LEO fixed-satellite-service system. Lumen Orbit surfaced in FCC experimental-license materials for Lumen-1, but the public record does not yet verify meaningful deployed GPU capacity, commercial pricing, SLAs, routine refresh, or a scaled fleet.
Wave 3 · GW-scale orbital data-center concepts are filing-stage
Starcloud's FCC public notice states the company sought authority for up to 88,000 satellites in sun-synchronous orbit between 600 km and 850 km as a distributed data center in space, material because it defines an architecture at a scale that could matter to hyperscale compute. SpaceX filed for an NGSO orbital data-center system of up to one million satellites at 500–2,000 km with high-bandwidth optical inter-satellite links. SpaceX is the actor most capable of invalidating a conservative ground-ahead forecast. But a filing is not a factory, a power system, a radiator architecture, a spectrum grant, an insurance market, a deorbit solution, or a refresh cadence.
05
Six grid-side levers that improve the ground-side comparison every year, the moving baseline orbit has to beat.
01
Secondary-market siting
Hyperscale development is already moving toward markets with cheaper power, available land, utility willingness, and large tax bases. The Meta Richland Parish record, including Entergy Louisiana's Mount Olive–Sarepta transmission materials, illustrates the timing mismatch between 18–24 month data-center builds and three-to-five-year generation or seven-to-ten-year transmission development, and identifies secondary markets with reliable cheap power as an explicit response.
02
Large-load tariffs and cost allocation
AEP Ohio's data-center tariff settlement (PUCO, July 2025) and Dominion Energy Virginia's GS-4 schedule illustrate the conventional large-load service framework once measured demand reaches or exceeds 500 kW in at least three billing months. The direction is clear: large compute loads will increasingly pay for commitment, exit risk, and system upgrades directly, not through socialized rates.
03
Behind-the-meter and bridge power
Gas turbines, reciprocating engines, batteries, backup fleets, staged utility service, and dedicated substations are not elegant, but they are deployable. They are also politically exposed, especially where air permits and emissions become visible. They buy time while transmission catches up.
04
Grid-enhancing technologies
DOE states that dynamic line ratings can allow lines to deliver 50% more energy than labeled limits under favorable conditions, and identifies dynamic line ratings, power-flow control devices, and analytical tools as grid-enhancing technologies. These tools will not create a 1 GW interconnection by magic, but they can improve the terrestrial frontier at the margin and defer or optimize upgrades.
05
Virtual power plants and flexible load
DOE describes virtual power plants as potentially providing 80–160 GW of flexible capacity by 2030, addressing 10%–20% of peak load and saving about $10 billion per year in grid costs. AI training is not infinitely flexible, but some workload scheduling, UPS/BESS dispatch, and non-critical curtailment can reduce grid stress.
06
Early transformer and equipment reservation
The transformer problem is solvable only by early procurement, long-horizon utility planning, and sponsor credit. It is a severe terrestrial disadvantage versus a purely conceptual orbital facility, but orbital compute still needs power electronics, flight hardware, ground stations, and launch cadence. The supply chain changes; it does not disappear.
06
Five workloads where orbital compute can earn its place without beating terrestrial commodity compute.
In-orbit data processing
Earth-observation, defense, space-domain awareness, and scientific payloads benefit from processing data near the sensor. NASA's SpaceCube and HPE Spaceborne Computer evidence supports this category. This is where space already wins on its own terms, not as a substitute for terrestrial hyperscale.
Defense, intelligence, and sovereignty
Certain workloads value physical separation, orbital vantage point, denial resilience, or sovereign architecture more than lowest cost per watt. Orbit is a strategic asset here, not a commodity substitute. Different buyers, different procurement logic.
Grid-constrained premium compute
If a buyer's alternative is waiting five years for a terrestrial interconnection in a constrained market, orbital compute may price as schedule insurance. That is not a commodity-cost win; it is a scarcity product. Real demand, narrow market.
Space-to-space services
Lunar, cislunar, station, satellite-servicing, and autonomous spacecraft operations could value compute where terrestrial round-trip dependence is operationally costly. Different problem, different cost stack, different customer set.
Brand and sustainability optics
Space-based solar-powered compute will attract attention. But brand value is not a substitute for cost parity, thermal closure, deorbit compliance, or refresh economics. It is real demand at the margin, not the spine of a hyperscale market.
07
Orbital compute should be priced as a strategic option until it proves a working cost stack on these dimensions.
01
GW-class deployable radiators and thermal control
Stefan–Boltzmann sizing for 1 GW of IT heat gives 0.8–2.6 million m² of radiator depending on operating temperature, before mass, deployment, degradation, redundancy, orientation, and solar-absorption penalties. None of this exists at scale today, and higher radiator temperature reduces area by the fourth power but raises equipment, coolant, reliability, material, and operating constraints.
02
Mass per usable compute watt, after power, shielding, bus, and redundancy
The relevant unit is not kilograms launched. It is usable compute watts per launched kilogram once the full stack (compute, power, radiators, bus, comms, shielding, redundancy, deorbit) is integrated. No public dataset proves this number at hyperscale AI scale.
03
Radiation-tolerant accelerator performance and refresh
Can a flight-rated accelerator refresh cycle approach terrestrial GPU refresh economics? Without it, the orbital silicon lags by a vintage or two, and the cost-per-watt comparison tilts further toward ground every refresh cycle.
04
High-cadence launch without long standdowns and satellite-factory throughput
The cheapest credible orbital story depends on launch cadence remaining high and on mass-manufactured compute, power, and thermal modules at satellite-factory scale. Throughput is unproven at GW-equivalent volumes, and any extended standdown, debris, anomaly, range constraint, disrupts deployment, refresh, insurance, and contracts.
05
Spectrum and optical backhaul at compute-scale data rates
Hyperscale AI data movement needs spectrum and optical backhaul architectures that are not yet a commodity. Spectrum is regulated; optical backhaul is contested with broadband constellations.
06
Fleet-scale collision avoidance, debris compliance, and insurance
FCC's five-year LEO deorbit rule, debris compliance, and insurance capacity for very large compute constellations are all still maturing. Insurance markets for FOAK-scale orbital compute fleets do not exist today.
07
Ground segment, cybersecurity, and customer workload migration
Even an orbital fleet still needs ground stations, fiber backhaul, cybersecurity, and operations. Customer workload architecture for latency-tolerant training versus latency-sensitive inference is unsettled, inference for terrestrial users is poorly served from orbit absent specialization, caching, or space-native demand.
08
Five recommendations for the players shaping the orbit-versus-ground comparison.
01
Hyperscalers: treat orbital compute as an option portfolio, not a substitute for terrestrial infrastructure
The core program should remain staged terrestrial campuses in power-rich markets, with early transformer reservations, utility cost-sharing, BTM/bridge power where politically viable, direct-to-chip and dry/hybrid cooling, reclaimed-water strategies, and flexible-load design. Orbital pilots should focus on workloads where orbit has structural value: edge processing, defense, sovereignty, and space-native services.
02
Infrastructure investors: underwrite orbital compute like FOAK infrastructure, not like a data-center lease-up
The relevant questions are not only launch price and capex; they are spectrum, debris, insurance, refresh, manufacturing yield, launch cadence, and customer willingness to pay for non-cost attributes. A filing-stage constellation should not receive the same risk treatment as a permitted, utility-served, terrestrial powered shell.
03
Launch providers: stop selling orbital compute primarily as a launch-cost story
The investment case improves when launch providers show integrated module costs: watts per kilogram, radiator kilograms per kilowatt rejected, deployed thermal area, replacement cadence, fairing packing efficiency, launch manifest reliability, and deorbit cost. Cheap rockets without cheap modules do not close the gap.
04
Utilities and grid operators: assume AI load growth remains a terrestrial problem
Even if orbital compute succeeds in niches, terrestrial AI load growth is the load-bearing capacity question. Large-load tariffs, direct cost assignment, long-term commitments, flexible-load credits, grid-enhancing technologies, and secondary-market transmission planning are the practical response, DOE's grid-enhancing technology and VPP materials show there is still room to expand effective grid capability.
05
Regulators: avoid both extremes
Terrestrial regulators should not use data-center impacts as a reason to freeze economically valuable infrastructure, they should require transparency on water, power, generators, transmission, and ratepayer risk. Space regulators should not treat orbital compute as just another broadband constellation: compute fleets could multiply spectrum, debris, power, brightness, and launch-cadence concerns.
In closing
The space-based-compute bull case is strongest when it argues that terrestrial constraints are real. It is weakest when it assumes those constraints are static and that orbital constraints are merely engineering details. The correct comparison is not orbit versus a delayed Ashburn interconnection. It is orbit versus a global terrestrial frontier that can move to power-rich markets, build dedicated transmission, reserve transformers, use large-load tariffs, deploy BTM generation, add batteries, optimize transmission, shift workloads, improve PUE/WUE, refresh silicon, and use modular construction.
The answer to the central question, what best satisfies power, schedule, connectivity, and cost priorities, is therefore: best overall for general AI compute is terrestrial frontier data centers; best strategic hedge is limited orbital compute pilots for edge processing, defense, sovereignty, and space-native workloads; not yet bankable as a commodity substitute is GW-scale orbital data centers. Starcloud and SpaceX filings are important, but filings are not deployed capacity, and announced satellite counts are not usable compute watts.
The single biggest risk to this recommendation is not that terrestrial power becomes easy. It will not. The biggest risk is that SpaceX or a similar vertically integrated actor compresses launch cadence, spacecraft manufacturing, power, thermal rejection, optical networking, and deorbit into one industrial machine, and proves a cost per usable compute watt close to terrestrial before 2040. Until that happens, orbital compute should be priced as a premium strategic product, not the next default hyperscale platform.
Full brief · 16 pages · PDF
Read the complete brief
Standing note
Independent analysis based solely on publicly available federal-agency documents, FCC and NASA filings, national-lab reports, regulatory dockets, and industry publications. Scenario economics are deliberately conservative toward orbit and use stated assumptions for order-of-magnitude comparison; they are not audited bids. For educational and discussion purposes only. Does not constitute investment, legal, or engineering advice and does not represent any of the named operators, launch providers, agencies, or regulators discussed.