AI infrastructure spending has never been larger. Efficiency gains are eroding the demand assumptions that make orbital data centers worth building
YASUYOSHI CHIBA / AFP via Getty Images
AI data centers are running out of room on Earth. Power grids are strained, permits take years, and the land required keeps growing. Space looks like the answer, but only if AI needs so much compute, for so long, that the conventional grid can't keep up. If it doesn't, the orbital thesis will collapse before it gets off the ground.
So far, the numbers are large. Global data center electricity consumption is projected to more than double by 2030. The four largest technology companies have committed $725 billion in capital expenditure for 2026 alone. Analysts have underestimated actual spending for two years running.
But the push for orbital data centers is still predicated on assumptions about how much computing power AI will require, how much enterprises will pay for it, and how long both trends will hold. Each of those assumptions is now being tested.
The International Energy Agency projects that global data center electricity consumption will rise from 415 terawatt-hours in 2024 to about 945 TWh by 2030. A terawatt-hour is roughly the amount of electricity the United Kingdom uses in a month. AI is the primary driver, with electricity demand from AI-optimized data centers projected to more than quadruple over that period.
McKinsey estimates that meeting this demand will require $6.7 trillion in capital expenditures worldwide by 2030, with $5.2 trillion of that going to AI workloads alone. Goldman Sachs $GS Research forecasts global data center power demand will rise 165% by the end of the decade compared with 2023 levels. Google $GOOGL, Amazon $AMZN, Microsoft $MSFT, and Meta $META have planned $725 billion in combined capital expenditures for 2026, a 77% increase from 2025's record, with actual growth exceeding 50% in both 2024 and 2025 against initial estimates of about 20%.
If terrestrial power and permitting can't keep pace with demand this large, space becomes a plausible overflow valve.
The amount of computing power needed to achieve a given level of AI performance has been falling fast enough to complicate long-term demand projections. Epoch AI found that the compute needed to reach a given performance level in language models has halved about every eight months. The same AI task that required a warehouse full of chips in 2022 requires far fewer today, and that pace of improvement is accelerating.
The cost of running AI models has followed the same trajectory. The Stanford HAI AI Index Report 2025 found that the cost of running a model at the performance level of GPT-3.5 dropped more than 280-fold between November 2022 and October 2024. Epoch AI's analysis found that prices to reach specific performance levels have fallen between nine and 900 times per year, with a median of 50 times per year. A task that cost $1,000 to run in 2022 might cost $20 today.
The trend is being pushed further by a technique called model distillation, in which a smaller, cheaper model is trained to mimic a larger one, producing similar results at a fraction of the cost. DeepSeek-V3 matched the performance of leading models from OpenAI and Anthropic at a training cost of about $5.6 million, a figure that would have seemed impossibly low two years ago. Researchers at Dropbox demonstrated that a similar process applied to smaller models cost between $3 and $18 per run while still delivering meaningful performance gains. If models keep getting cheaper to run, the total compute demand that would legitimize orbital data centers could be far lower than today's forecasts suggest.
The revenue side is equally uncertain. About 95% of enterprise AI pilot programs fail to achieve measurable financial returns, a report from MIT's NANDA initiative found. Most companies remain stuck in experimentation rather than generating returns from AI, Bain's 2025 technology report found. The IEA's own projection that data center electricity use will double by 2030 is made, as a Brookings analysis noted, "in spite of anticipated efficiency gains." The gap between investment and return has widened year over year, as infrastructure spending has scaled faster than revenue. If enterprise AI adoption stalls, the hyperscalers could pull back, reducing the very compute demand that makes orbital infrastructure necessary.
The hyperscalers are already feeling the cash pressure. The four major tech giants generated $200 billion in combined free cash flow in 2025, down from $237 billion in 2024. Morgan Stanley $MS analysts project Amazon could face negative free cash flow of almost $17 billion this year. Pivotal Research forecasts an almost 90% decrease in Alphabet's free cash flow.
Efficiency bulls have a ready response. When Microsoft CEO Satya Nadella was asked about DeepSeek's efficient AI model, he invoked the Jevons paradox. The 19th-century economic principle argues that cheaper resources tend to drive higher total consumption, not lower, because falling costs open up uses that weren't previously viable. Applied to AI, the argument is that falling model costs will generate entirely new applications, consuming more total computing power than the efficiency gains saved.
History offers some support. When steam engines became more efficient, Britain burned more coal, not less, because cheap energy made possible industries that hadn't existed before. If AI inference becomes cheap enough, demand could multiply across sectors that can't justify the cost today.
But the Jevons paradox only works if lower prices generate dramatically more usage. For AI, that's still an open question. Citigroup $C projects that launch costs won't fall to the level needed for orbital data centers to be economically viable until around 2040. That's a 15-year window in which demand has to hold — and keep growing.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
