The Terrestrial Crisis

Data Center Electricity Supply vs Demand (TWh)

Source: IEA Key Questions on Energy and AI (2026) — demand 485 TWh (2025) → 945 TWh (2030); generation serving data centres ~537 TWh (2025) → ~1,050 TWh (2030)

Rising Terrestrial Electricity Prices ($/kWh)

Source: EIA (historical) + AEO2025 (projected), US all-sectors retail average

⚡ IEA: Key Questions on Energy and AI (2026)

The IEA’s April 2026 follow-up to its landmark Energy and AI report confirms data-centre electricity use reached ~485 TWh in 2025 (~1.5% of global electricity) and is on track to roughly double to ~945 TWh by 2030 — comparable to Japan’s total electricity use today. Demand grew 17% in 2025; AI-focused data centres grew even faster at ~50%.

  • 2025: ~485 TWh (~1.5% of global electricity); 2030 Base Case: ~945 TWh (~3%)
  • Demand grew ~17% in 2025 — AI-focused facilities ~50%
  • Generation serving DCs projected to exceed 1,000 TWh by 2030
  • Renewables meet nearly half of incremental supply to 2030; gas and coal fill the near-term gap
  • Grid connection queues and supply-chain bottlenecks tightening
  • Builds on IEA Energy and AI (2025) sensitivity range: ~700 TWh (Headwinds) to ~1,700 TWh (Lift-Off) by 2035

💧 Nature: Data Centre Water Consumption

A peer-reviewed study in Nature found that US data centres consumed 1.7 billion litres of water per day in 2014 — before the AI boom. A single 15 MW facility uses as much water as three hospitals. Fewer than one-third of data centre operators even measured their water usage at the time of the study.

  • A 1 MW data centre can consume ~25.5 million litres of water per year
  • Google drew 1.9M litres/day from South Carolina aquifers
  • Up to 57% of cooling water drawn from potable (drinking) sources
  • These figures pre-date the AI boom — current consumption is significantly higher

🌌 Scientific American: Data Centers in Space

Scientific American examines why data centres, once welcomed by towns for their jobs and tax revenue, are now facing community opposition over power, water, and land demands. Data centres could account for nearly half of US electricity demand growth through 2030 — which is why some tech executives now see orbit as the escape hatch.

  • Data centres projected to account for nearly half of US electricity demand growth by 2030
  • Global data centre power requirements could double by end of the decade
  • Local permitting is stalling builds across Texas, Oklahoma, and the Pacific Northwest
  • Space-based alternative: near-continuous solar and zero water use

The Orbital Solution

🌍 The Impact of Computing Data Centres Orbiting Earth (MNRAS)

A paper in the Monthly Notices of the Royal Astronomical Society computing the brightness, size, and appearance of proposed orbital data centers. Large solar arrays spanning several kilometers would be necessary to power gigawatt-scale compute, creating significant light pollution visible from Earth. Raises important environmental concerns about orbital congestion and astronomical observation interference that counterbalance the ‘clean energy in space’ narrative.

  • Gigawatt-scale ODCs need multi-kilometer solar arrays — highly visible from Earth
  • Brightness could interfere with astronomical observations
  • Raises orbital debris concerns alongside compute benefits
  • Analyzes SpaceX (1M satellites), Starcloud (88K), Blue Origin (51.6K) proposals
  • Provides crucial environmental counterpoint to orbital solution narrative
  • Design mitigations exist but add mass and cost

🤖 Why We Should Train AI in Space

This whitepaper by Starcloud (then LumenOrbit) makes the engineering case for running AI training in orbit. A sun-synchronous orbit provides solar power ~96% of the time. The vacuum of space enables passive radiative cooling — heat radiates into deep space, eliminating water-cooled chillers entirely.

  • ~96% solar duty cycle in dawn/dusk sun-synchronous orbit
  • PUE of ~1.05 vs 1.2–1.5 for typical Earth data centers (lower = less energy wasted on cooling)
  • Zero operational water use — heat rejected via radiation into deep space
  • Hardware must be radiation-hardened to survive cosmic rays and solar flares

☀️ Project Suncatcher

Google Research’s Project Suncatcher proposes clusters of TPU-equipped satellites flying hundreds of meters apart in LEO, connected by laser links capable of 1.6 Tbps. Google tested their Trillium TPU chips in a proton beam — they survived 3x the radiation expected over a 5-year mission.

  • Satellites fly in tight clusters (hundreds of meters apart) to enable high-speed laser data links between nodes
  • 1.6 Tbps inter-satellite links — comparable to fiber optic backbone speeds
  • Google's TPU chips survived radiation doses 3x higher than expected for a 5-year mission
  • Economics become competitive when launch costs fall below ~$200/kg, projected mid-2030s

🇪🇺 ASCEND Feasibility Study

ASCEND is a European Space Agency-funded feasibility study for an orbital data center that would serve Europe’s AI infrastructure needs — without depending on US or Chinese cloud providers. Facilities would be assembled robotically in orbit from hundreds of launches.

  • Data sovereignty: Europe storing its data in space rather than on US/Chinese-owned clouds
  • Robotic assembly in orbit — too large to launch in one piece
  • Targets operational by 2030, scaling to thousands of tons of infrastructure
  • Designed to achieve net-zero carbon footprint by keeping energy generation in space

Visionaries & Timelines

🚀 Elon Musk: Cheapest AI in Space

In early 2026, Elon Musk claimed that space will be the cheapest place to run AI within 3 years, driven by the economics of Starship reusability and the cost advantage of solar power in orbit over grid electricity on the ground.

  • Starship targets <$100/kg to LEO — 10x cheaper than current rockets — making orbital compute economically viable
  • Solar in orbit available ~96% of the time vs. paying grid electricity rates on Earth
  • No land, no permits, no utility infrastructure needed
  • Most analysts consider the 3-year timeline aggressive; 5-10 years is more widely cited

🔍 Sundar Pichai: Data Centers in Space Will Be the New Normal

Google CEO Sundar Pichai publicly endorsed orbital data centers in late 2025, citing Project Suncatcher as Google’s bet. He predicted space-based compute will become the ’new normal’ within a decade — a notable signal given Google’s $50B+ annual infrastructure spend.

  • Google plans to launch prototype TPU satellites with Planet Labs by early 2027
  • Orbital compute lets Google scale AI without competing for scarce grid capacity
  • 'New normal' prediction reflects a ~10-year horizon, not an immediate shift
  • Significant given Google spends more on data center infrastructure than most countries' space budgets

🛒 Jeff Bezos: Gigawatt-Scale Data Centers in Space

At Italian Tech Week in 2025, Amazon founder Jeff Bezos predicted gigawatt-scale data centers in space within 10-20 years. Blue Origin’s New Glenn rocket — now operational — is his vehicle for making it happen, alongside the proposed Orbital Reef commercial space station.

  • A gigawatt is roughly the output of a nuclear power plant — Bezos envisions that scale in orbit
  • Blue Origin's New Glenn rocket can lift ~45 tonnes to LEO, enabling large orbital infrastructure
  • 10-20 year timeline is more conservative than Musk's — and more widely accepted by analysts
  • Bezos sees solar power in orbit as permanently cheaper than building new grid capacity on Earth

🎤 Philip Johnston: Let's Build AI Data Centers in Space

Starcloud CEO’s TED talk on using commercial hardware in orbit to escape terrestrial power/water/land limits for AI.

  • First company to train AI models entirely in orbit with COTS H100 GPUs
  • Demonstrates commercial chips survive radiation without full hardening
  • Economic viability improves rapidly as launch costs fall
  • Urgent call to action for space data centers to meet exploding AI demand