“I wish Elon luck.” That was OpenAI CEO Sam Altman’s response on the TBPN podcast to the growing enthusiasm for orbital data centers. Altman argued that putting compute underwater is a simpler engineering challenge than launching it into space, according to Benzinga.
“I feel like none of the ocean guys working on ocean robotics are thinking about putting this in at the bottom of the ocean,” Altman said. “It feels easier than constructing space data centers.”
The remark positions OpenAI directly against an orbital computing push that now includes SpaceX, Google, and Blue Origin. The Wall Street Journal reported that SpaceX and Google have discussed space-based computing projects. Google has explored a prototype initiative called Project Suncatcher, with small computing satellites potentially launching by 2027, according to Google Research. Blue Origin submitted FCC plans in March for 51,600 data center satellites under “Project Sunrise,” per CNBC.
The Cooling Problem
The infrastructure debate is fundamentally about heat. AI inference at scale generates enormous thermal loads, and cooling represents one of the fastest-growing cost lines for hyperscalers. Orbital proponents argue that space offers unlimited solar energy and eliminates land permitting battles. Altman’s counter: the ocean solves the cooling problem with less complexity.
“The space is free. The land is free. You don’t run into the regulatory stuff,” Altman acknowledged on TBPN, per Benzinga. But he added that energy generation and infrastructure remain “difficult challenges in extreme environments,” whether that environment is orbit or ocean floor.
Jeff Bezos, whose Blue Origin is actively pursuing orbital compute, offered a measured view on CNBC on May 20. “Some of the timelines we hear are very short,” Bezos told Andrew Ross Sorkin. “People would talk about two or three years. That’s probably a little ambitious.” Bezos identified energy costs and chip prices as key barriers, noting that launch costs still need to drop before orbital compute becomes viable.
Why This Matters for Agent Infrastructure
The debate is not academic for teams deploying autonomous agents at scale. Agent workloads are inference-heavy, long-running, and unpredictable in burst patterns. They stress power grids and cooling systems differently than batch training runs. An agent orchestrating a multi-step workflow might sustain GPU utilization for hours rather than milliseconds, making thermal management a persistent cost rather than a peak-demand problem.
JP Morgan forecasts $200 billion in AI spending from top US cloud providers, according to Benzinga. Deloitte projects US defense AI spending will grow 3.5x by 2029, with contracts increasingly flowing to AI companies. Where that compute physically runs, and how it stays cool, will shape pricing, latency, and availability for every agent framework competing for enterprise adoption.
Three Bets, No Consensus
The industry is placing three incompatible bets on the physical future of AI compute.
SpaceX merged with xAI in February partly to pursue orbital data centers, and its upcoming IPO could value the combined entity at $1.75 trillion or higher, per CNBC. Google is prototyping satellite compute. Blue Origin has filed for a 51,600-satellite constellation.
Meanwhile, Amazon, Microsoft, and Google continue expanding land-based facilities. And Altman is quietly proposing a fourth path: subsea compute that nobody in ocean robotics has seriously attempted yet.
“Everyone has diversified into every company at this point, basically,” Altman observed. The comment captures a market where the physical infrastructure question remains genuinely unsettled, and where the answer will determine not just cost structures but which geographies and regulatory regimes control AI capacity in the next decade.