Google and SpaceX Are in Talks to Put Data Centers in Space

The Wall Street Journal reported on May 13, 2026 that Google and SpaceX are in active discussions to launch AI compute infrastructure into orbit — data centers in space, designed to serve the surging demand for AI processing that terrestrial power grids and cooling infrastructure can no longer absorb fast enough. The talks are at an early stage, but involve placing GPU clusters aboard spacecraft using SpaceX's Starship launch vehicle, powered by solar arrays, cooled by radiative panels, and connected to Earth via Starlink's inter-satellite laser link network.

This is not a science fiction pitch. The economics that make it absurd today are the same economics that made cloud computing absurd in 2000. The question is whether Starship's cost-per-kilogram trajectory intersects with AI infrastructure demand fast enough to make orbital compute viable before terrestrial alternatives scale to meet demand.

What the WSJ Report Actually Says

The WSJ report describes early-stage discussions, not a signed deal. The contours of what is being discussed:

SpaceX would provide the launch capability via Starship, with a target cost-per-kilogram that makes large-scale orbital payload economically justifiable for commercial infrastructure. Google would bring the compute hardware — likely TPU-based or GPU-based clusters adapted for space environments — and the software and operational layer for running AI workloads in orbit.

The connection layer is Starlink. SpaceX's existing constellation already uses inter-satellite laser links to route data at near-light-speed between satellites. An orbital data center connected to Starlink can transmit inference requests and responses to any point on Earth with Starlink coverage — which, as of mid-2026, means most of the planet.

The immediate use case is not replacing AWS data centers in Virginia. It is serving the parts of the world where terrestrial cloud infrastructure does not reach fast enough, and handling the class of workloads — AI inference for remote sensing, satellite imagery processing, autonomous systems — where being in orbit is a direct advantage rather than a limitation.

The Physics: Why Space Is Actually Good for Data Centers

The instinct is to assume space is a terrible place for a data center. It is hostile, expensive to reach, and impossible to maintain in person. But several of the hardest problems in terrestrial data centers are trivially solved in orbit.

Cooling: Terrestrial data centers fight heat constantly. Water cooling, chillers, airflow management — a large portion of data center capex and opex is thermal management. In space, heat dissipation is purely radiative. A radiator panel facing away from the sun dumps heat directly to the 3 Kelvin background of space. The thermal sink in space is nearly perfect. An H100 cluster that generates 700W of heat per GPU can be cooled with a radiator panel sized appropriately — no water, no compressors, no cooling towers.

Power: A terrestrial data center pulls power from a grid and pays 3-7 cents per kWh, plus the carbon and reliability politics of grid dependency. A solar array in low Earth orbit receives approximately 1,400 watts per square meter of solar irradiance — higher than ground level because there is no atmosphere to absorb it. Modern solar cells at 30% efficiency generate about 420W/m². A 13 m² solar panel generates enough power for one H100 GPU continuously (accounting for the ~35% of the orbit spent in Earth shadow, which requires battery storage). The economics become compelling when launch costs fall far enough.

Location for specific workloads: Satellite imagery AI currently works like this — a satellite images the Earth, transmits hundreds of gigabytes of raw data to a ground station, which processes it in a terrestrial data center, then sends instructions or results back. Inserting an AI inference cluster directly into the satellite (or a companion data center in the same orbit) eliminates the downlink-process-uplink cycle entirely. For real-time applications — wildfire detection, military targeting, storm tracking — on-orbit processing is not a convenience, it is an architectural requirement.

The Physics: Why Space Is Hard for Data Centers

Radiation: Low Earth orbit sits below the Van Allen radiation belts for most of the orbit, but cosmic rays and solar energetic particles still cause bit-flips in consumer electronics. Standard GPU silicon accumulates single-event upsets at rates that require either error-correcting memory, radiation-hardened chips, or passive shielding. Radiation-hardened chips exist — the defense industry uses them — but they are typically 3-5× more expensive than commercial equivalents and several generations behind in performance. An H100-equivalent rad-hard GPU does not yet exist commercially.

The practical near-term answer is shielding: a spacecraft with sufficient mass to protect GPU clusters from most radiation events. This increases launch mass, which increases cost. The engineering tradeoff is well-understood — the question is where it lands in the cost model.

Latency: Low Earth orbit is approximately 550 km above Earth's surface. The speed of light round-trip is about 3.6ms at that distance. Add satellite-to-satellite routing, ground station processing, and network overhead, and a realistic round-trip latency from a Starlink terminal to an orbital data center and back is 20-50ms. This is fine for batch AI inference jobs, model training runs, and satellite imagery processing. It is not fine for real-time interactive AI applications where humans expect sub-20ms response. Orbital data centers are built for the former workload category, not the latter.

Servicing: If a GPU fails in a terrestrial data center, a technician replaces it. If a GPU fails in a data center 550km above Earth, it stays failed until the next Starship servicing mission or end of life. This drives a design philosophy of massive redundancy and autonomous failure handling — the opposite of how dense GPU clusters are operated on the ground.

The Starship Economics: When Does This Work?

Everything about orbital data centers is gated on Starship's cost-per-kilogram trajectory.

The current state (mid-2026): Starship has completed 8 integrated flight tests. Full catch-and-reuse of the Super Heavy booster is operational. The upper stage (Ship) recovery via tower catch is operational on multiple flights. SpaceX has not publicly stated a price for commercial Starship launches, but internal targets have been reported at $10-50/kg to LEO when fully mature — compared to Falcon 9's current ~$2,700/kg.

The calculation at different cost points:

At $2,700/kg (Falcon 9 equivalent): A single H100 SXM5 server with 8 GPUs, cooling infrastructure, and power conversion hardware weighs approximately 600kg. Launch cost: $1.6 million per server. Hardware cost: $400,000. Total: $2 million per server to get to orbit vs $400K on the ground. Economics do not work at any AI workload.

At $500/kg (near-term Starship): $300K launch cost per 8-GPU server. Combined with hardware: $700K per server. For workloads where orbital placement provides direct value (remote sensing, satellite AI), this starts becoming viable — particularly for defense customers paying 10-20× terrestrial rates for mission-critical processing.

At $100/kg (mature Starship target): $60K launch cost per 8-GPU server. Combined with hardware: $460K — essentially the same as a well-provisioned terrestrial server. At this point, the power and cooling advantages of space make orbital compute cheaper than terrestrial for certain workload classes.

At $10/kg (theoretical fully-mature reusable Starship): $6K launch cost per 8-GPU server. Space becomes definitively cheaper than Earth for large-scale AI infrastructure when you include land, cooling, and power costs at terrestrial scale.

SpaceX's public timeline: commercial Starship cargo flights begin in 2027-2028. The $100/kg target is a 2030+ projection, not a 2027 reality.

What Google Brings to This

Google is not in these discussions because it is the only AI company — it is in these discussions because of three specific capabilities:

TPU infrastructure: Google's Tensor Processing Units are custom silicon designed specifically for AI workloads. The latest generation (Trillium/TPU v6) achieves significantly higher performance-per-watt than NVIDIA H100s for certain inference workload types. Lower power consumption matters enormously in space, where every watt of power requires ~3kg of solar panel mass. A lower-power custom AI chip is a structural advantage in the orbital environment.

Satellite data relationships: Google Earth Engine processes petabytes of satellite imagery today, using terrestrial data centers. Moving the first-stage processing of satellite imagery on-orbit is a direct extension of Google's existing business — not a new market, but a radically different infrastructure configuration for the same workload.

Starlink integration: Google and SpaceX have previously announced a deal for Pixel phones to connect directly to Starlink satellites for messaging. The relationship exists. An orbital data center served via Starlink is a natural extension of that integration.

Who Else Is Playing This Game

Google and SpaceX are not the only players thinking about orbital compute.

Amazon: Project Kuiper is Amazon's satellite internet constellation, currently in early deployment. AWS has been offering ground station as a service (AWS Ground Station) since 2019 — essentially cloud compute at satellite ground stations. The logical next step is compute in the satellite itself. Amazon has not announced Kuiper compute plans, but the infrastructure is being built for it.

Microsoft Azure Space: Microsoft launched Azure Orbital in 2021 — a service for connecting satellites to Azure cloud. In 2025, Microsoft partnered with Loft Orbital and others to offer hosted payload services on satellites. Azure Space is already selling compute co-located on satellites. It is a smaller scale than what Google and SpaceX are discussing, but the model is the same.

Defense contractors: Northrop Grumman, Raytheon, and Lockheed Martin have been developing space-based compute for defense applications for years — specifically for on-orbit processing of intelligence, surveillance, and reconnaissance (ISR) data. The classified programs in this area are significantly more advanced than anything Google and SpaceX are discussing publicly.

Reflect Orbital: A startup that raised funding this year to put reflective satellites in orbit that direct sunlight to ground-based solar farms at night. Not a data center play, but indicative of the trend: space infrastructure is increasingly being evaluated for terrestrial energy and compute problems.

The Use Cases That Actually Work

Not every AI workload belongs in orbit. The ones that do:

Earth observation AI: Processing satellite imagery on-orbit to detect wildfires, floods, troop movements, crop failures, illegal fishing, and infrastructure changes — without transmitting raw imagery data to the ground. The data reduction is enormous: a raw synthetic aperture radar (SAR) image of a 100km² area is 50GB. The AI output ("no change detected" or "anomaly at coordinates X,Y") is 1KB. Moving the AI to orbit reduces downlink bandwidth requirements by 50,000:1 for this class of application.

Starlink edge AI inference: Running AI inference nodes close to Starlink's inter-satellite routing network allows Starlink terminals in remote locations — ships, aircraft, rural areas, disaster zones, military forward operating bases — to access AI capabilities at lower latency than routing to terrestrial cloud regions. This is a direct product extension for Starlink's commercial and military customer base.

Global AI coverage without geopolitics: A terrestrial AI data center in the US is subject to US export control law. Serving Chinese or Russian users from US cloud infrastructure is restricted. An orbital data center in international space is in a different legal environment — one that satellite operators have been navigating for decades with broadcast services. The geopolitical dimensions of orbital AI infrastructure are complex and not yet fully mapped.

Autonomous systems coordination: Autonomous vehicles, drones, and robots operating in areas without terrestrial coverage need AI inference from somewhere. Orbital AI compute, connected via Starlink, provides global coverage for the autonomous systems AI stack that no terrestrial cloud region can match.

The Timeline: What Is Realistic

The WSJ report describes early-stage talks. Based on the underlying economics and engineering constraints, a realistic development timeline:

2026-2027: Discussions and feasibility studies. SpaceX continues Starship commercial certification. Google adapts TPU hardware for radiation-hardened packaging. No hardware in orbit yet.

2028-2029: First demonstration payload. A small cluster (8-16 TPU/GPU equivalents) on a Starship manifest, connected to Starlink, running inference for internal Google research or a defense customer. Proves the architecture works. No commercial product.

2030-2032: First commercial orbital AI compute product — likely a specialised offering for Earth observation customers (governments, agricultural tech, climate monitoring) and Starlink commercial subscribers. Still expensive per-FLOP relative to terrestrial, justified by the specific workload advantages.

2033-2035: Cost parity with terrestrial for specific workload classes, assuming Starship achieves $100/kg. Orbital AI compute becomes a mainstream infrastructure option for latency-insensitive batch workloads.

Key Takeaways

• WSJ confirmed talks: Google and SpaceX in early discussions for orbital data centers — not a signed deal, but real enough to report. Starship provides the launch vehicle; Google provides the AI compute and software stack; Starlink provides the connectivity.
• Why space works for compute: Near-perfect radiative cooling (no water, no chillers), uninterrupted solar power above the atmosphere, on-orbit positioning for satellite imagery AI workloads, and global coverage via Starlink for remote AI inference.
• Why space is hard: Radiation bit-flips require rad-hard chips or shielding; latency adds 20-50ms (fine for batch, bad for real-time); servicing requires Starship missions not technicians; current cost at Falcon 9 rates is 4-5× terrestrial.
• The Starship cost gate: Everything hinges on Starship reaching $100/kg to LEO. At that price, orbital compute approaches cost parity with terrestrial. Current reality: $500-2,700/kg. Target: $10-100/kg by 2030+.
• Who benefits now: Defense and intelligence (already paying premium for on-orbit processing), Earth observation (50,000:1 bandwidth reduction by processing imagery in orbit), and Starlink commercial customers needing AI inference in areas without terrestrial coverage.
• Competitors: Amazon Kuiper + AWS Ground Station, Microsoft Azure Space, defense contractors with classified on-orbit compute programs all pursuing the same architecture.
• Realistic commercial product: 2030-2032 earliest for a commercial offering, 2033-2035 for cost parity with terrestrial at specific workload classes. Today's talks are a 7-10 year infrastructure investment, not a 2026 product announcement.

For the Google I/O 2026 preview covering Gemini and Android announcements, read Google I/O 2026 Preview: Gemini 3.2 Flash, Android 17, What Developers Get. For the xAI merger into SpaceXAI that already combines Musk's AI and space operations, read xAI Dissolved Into SpaceXAI: Musk Merges Grok and Colossus Into SpaceX.