AI Is Leaving Earth

The motivation is simple: AI is hungry, and Earth can't feed it forever. Data centers already consume over 4% of U.S. electricity, projected to reach 12% by 2028. Microsoft reportedly has warehouses full of GPUs sitting idle because there isn't enough power to run them.

Traditional data centers also consume 2 to 5 million gallons of water per megawatt annually for cooling. As droughts intensify globally, this is becoming both an environmental and political liability. The AI boom is straining power grids, driving up electricity bills, and forcing tech companies to get creative.

Google's Project Suncatcher, announced November 4, 2025, is arguably the most ambitious initiative in the space. The plan: deploy clusters of 81 satellites orbiting in formation within a 1-kilometer radius, connected by laser links capable of 1.6 terabits per second, each carrying Google's proprietary Trillium TPU chips.

In partnership with satellite-operator Planet Labs, Google will launch two prototype satellites in early 2027. Each will be equipped with four TPUs to test radiation tolerance, thermal management, and inter-satellite communication.

Elon Musk, characteristically, isn't waiting for the economics to become obvious. In December 2025, he confirmed SpaceX's plans to scale Starlink satellites for AI workloads, tweeting: "Starlink V3 will be 20kW and launched at scale around Q4 next year. No problem to scale that to >100kW if the satellite mass is shifted towards solar arrays and radiators for AI compute."

SpaceX's advantage is staggering scale. With over 9,400 satellites already orbiting (65% of all active satellites globally) and high-speed laser links connecting them, Starlink is already the world's largest distributed network infrastructure.

While Google and SpaceX plan for the future, Starcloud has already done it. The Redmond, Washington startup launched Starcloud-1 on November 2, 2025, aboard a SpaceX Falcon 9, carrying a single NVIDIA H100 GPU into low Earth orbit. This was the first data-center-class chip ever to operate in space.

The 12-person team, which includes ex-SpaceX Starlink engineer Adi Oltean, has since demonstrated what most thought impossible. They trained NanoGPT (developed by OpenAI co-founder Andrej Karpathy) on Shakespeare's complete works, producing eerily Elizabethan outputs. More impressively, they deployed Google's Gemma LLM and received its first orbital transmission.

This isn't just an American competition. China launched its "Three-Body Computing Constellation" in May 2025 (a nod to Liu Cixin's science fiction trilogy), with 12 satellites from Zhejiang Lab and ADA Space. The plan calls for 2,800 satellites with combined processing power of 1,000 petaflops, connected by 100 Gbps laser links.

The European Union's ASCEND project, led by Thales Alenia Space under a Horizon Europe grant, published positive feasibility results in June 2024. Their roadmap targets a 50kW proof of concept by 2031 and 1 gigawatt deployment by 2050. It's more conservative than American timelines but comes with significant government backing focused on European digital sovereignty.

For investors, this emerging sector presents both opportunity and extreme risk. The total addressable market is compelling: global data center spending is projected to reach $517 to $691 billion by 2030, while AI electricity consumption grows 50% annually. If space can capture even 5 to 10% of new capacity, that represents tens of billions in potential revenue.

The economics remain uncertain. Google estimates that space data centers need launch costs below $200/kg to achieve cost parity. That's a 7 to 8x reduction from today's $1,500 to $2,900/kg. SpaceX's Starship could theoretically achieve this, but hasn't yet demonstrated operational reliability.

Critical analysts note substantial hurdles:

• Radiation: Consumer-grade chips need hardening at scale
• Thermal management: Massive radiators required (1 to 3 m² per kilowatt)
• Satellite handoffs: Computation must continue during orbital transitions
• Upfront costs: Could exceed $50 to $100 million per megawatt initially

The race to space-based computing represents a fundamental shift in how we think about infrastructure constraints on AI. For the past decade, progress in AI has been gated by three factors: algorithms, data, and compute. Space potentially removes the ceiling on the third.

If Starcloud's economics prove out, and if Starship achieves the launch frequency that Google and SpaceX are banking on, we could see AI training runs of unprecedented scale. Models with trillions of parameters trained on distributed orbital networks, with no grid capacity or permitting constraints. The implications for AI development, and for the companies that control this infrastructure, are enormous.