AI Compute Demand May Not Justify Orbital Data Centers

The race to build data centers in space rests on a single assumption: that artificial intelligence will demand so much computing power that Earth's infrastructure simply cannot keep up. But that assumption is now facing a stress test from two directions—plummeting model costs and uncertain enterprise adoption.

The demand projections are massive

The International Energy Agency forecasts global data center electricity consumption will climb from 415 terawatt-hours in 2024 to roughly 945 TWh by 2030. For context, a terawatt-hour approximates what the United Kingdom consumes in a month. AI workloads are the primary driver, with electricity demand from AI-optimized facilities expected to more than quadruple during that span.

McKinsey estimates meeting this demand will require $6.7 trillion in capital expenditures worldwide by 2030, with $5.2 trillion dedicated to AI workloads. Google, Amazon, Microsoft, and Meta have collectively planned $725 billion in capital spending for 2026 alone—a 77% increase over 2025's record levels.

If terrestrial power grids and permitting processes cannot accommodate growth at this scale, orbital data centers start to look viable as an overflow solution.

Efficiency is collapsing costs faster than expected

The computing power required to achieve a given AI performance level has been halving approximately every eight months, according to Epoch AI. The Stanford HAI AI Index Report 2025 found that running a model at GPT-3.5 performance levels became 280 times cheaper between November 2022 and October 2024. Epoch AI's analysis shows prices to reach specific performance benchmarks have fallen between nine and 900 times per year, with a median of 50 times annually.

Model distillation—training smaller models to mimic larger ones—is accelerating this trend. DeepSeek-V3 matched leading models from OpenAI and Anthropic at a training cost of approximately $5.6 million, a figure that would have seemed impossible two years prior. Dropbox researchers demonstrated similar processes costing between $3 and $18 per run while maintaining meaningful performance.

Enterprise adoption remains uncertain

About 95% of enterprise AI pilot programs fail to achieve measurable financial returns, according to MIT's NANDA initiative. Bain's 2025 technology report found most companies remain stuck in experimentation rather than generating returns from AI investments.

The financial pressure is already visible. The four major hyperscalers generated $200 billion in combined free cash flow in 2025, down from $237 billion in 2024. Morgan Stanley analysts project Amazon could face negative free cash flow of nearly $17 billion this year. Pivotal Research forecasts an almost 90% decrease in Alphabet's free cash flow.

The Jevons paradox counterargument

Microsoft CEO Satya Nadella has invoked the Jevons paradox in response to efficiency concerns—the 19th-century economic principle suggesting that cheaper resources drive higher total consumption by enabling previously unviable uses. Applied to AI, falling model costs could generate entirely new applications that consume more total computing power than efficiency gains save.

But this only holds if lower prices generate dramatically more usage. Citigroup projects launch costs won't fall to levels needed for orbital data centers to be economically viable until around 2040—a 15-year window during which demand must not only hold but continue expanding.

Why it matters

The orbital data center thesis depends on sustained, exponential growth in AI compute demand outpacing terrestrial infrastructure capacity. If efficiency gains continue reducing costs at current rates while enterprise adoption stalls, the economic justification for space-based infrastructure collapses before launch costs become viable. The next few years will determine whether we're building toward an orbital future or overbuilding for a demand peak that never materializes.

These details were first reported by Quartz.