Elon Musk’s vision of moving large-scale artificial‑intelligence computation off Earth has ignited fresh debate about the future of cloud infrastructure. Over the past months Musk has linked SpaceX, his AI outfit xAI and revived Dojo3 ambitions into a narrative that positions orbit as the next frontier for hyperscale compute. But Matt Garman, CEO of Amazon Web Services, used a high‑profile Cisco AI summit in San Francisco to deliver a stark rebuttal: while alluring in theory, orbital data centres remain impractical and prohibitively expensive today.
The commercial and technical rationale for the concept is straightforward. AI models consume enormous power and generate prodigious heat; on Earth this drives soaring energy demand, costly cooling systems and local limits on further scaling. Proponents argue that space offers perpetual solar power, low ambient temperatures for radiative cooling and relief from terrestrial constraints such as land use and local grid capacity. High‑profile bets from figures such as Musk and Jeff Bezos, plus initiatives by Google and a handful of startups, have converted hubris into headline‑grabbing prototypes and proposals.
Garman’s pushback, however, focuses on the arithmetic and engineering realities. He pointed to current launch cadence limitations — there are nowhere near enough rockets to loft the hundreds of thousands or millions of payloads required — and he emphasised the astronomical per‑kilogram cost of sending heavy, delicate GPU clusters into orbit. The economics, he argued, do not yet support substituting terrestrial racks with orbital equivalents, a point echoed privately by other industry leaders such as Nvidia’s Jensen Huang.
Beyond launch economics, the technical barriers are formidable. Space exposes hardware to wide thermal swings, high radiation doses and micrometeoroid and debris risk; GPU accelerators and modern interconnects were not designed for those conditions. Effective radiative cooling for megawatt‑scale GPU clusters would demand deployable radiator surface areas and structures far larger than any flown to date. Achieving robust, high‑bandwidth Earth links, autonomous on‑orbit maintenance and debris‑avoidance systems would multiply complexity and cost.
Despite those obstacles, companies are experimenting. Google’s “sunlight catcher” concept, a demonstration slated for the late 2020s, and startups like Starcloud — which lofted an Nvidia H100–equipped 60‑kg satellite — are testing the edges of feasibility. Bezos’s Blue Origin and Musk’s SpaceX both figure prominently in the narrative because only actors with integrated rockets, launch scale and capital can contemplate the logistics of orbiting large compute legs. Yet the prototypes so far are small, niche and far from the gigawatt‑scale systems some futurists envisage.
The debate matters because it reframes how hyperscalers and national governments think about the limits of compute growth and infrastructure strategy. If orbital compute ever becomes economical it would reshape energy markets, data sovereignty and military‑civil dynamics in space. For now, though, the sensible conclusion is incrementalism: expect demonstration missions, specialty edge uses and incremental advances in radiation hardening and deployable structures rather than wholesale migration of cloud datacentres into orbit within a two‑to five‑year horizon.
Policymakers and investors should treat grandioise pronouncements as what they are — strategic signalling intertwined with product marketing — while tracking a narrow set of indicators that would make orbit viable: a sustained plunge in launch costs and cadence, credible breakthroughs in on‑orbit assembly or modular fabrication, and economically viable radiation‑hardened accelerators. Absent those, the story of space data centres will remain an expensive curiosity rather than the next chapter in cloud computing.