Orbital data centers: Google and SpaceX in talks

Orbital data centers are the subject of secret talks between Google and SpaceX. The project is codenamed “Low Latency,” and according to internal documents obtained by this reporter, the two companies have

📊 Market Context: According to Dataintelo, the global orbital data center market was valued at $1.8 billion in 2025.

been in active negotiations for the past three months. The goal is nothing short of launching a fully functional, rack mounted data center into low Earth orbit by 2027. If it works, it will rewrite the economics of cloud computing, geopolitics, and even the physical limits of the internet. If it fails, it will be the most expensive debris field in history.

Here is the part they did not put in the press release. These are not rumors. We have verified the memo dates, the code names, and the financial structures. A person familiar with the talks told us that Google Cloud’s infrastructure team and SpaceX’s Starlink engineering group have held four closed door meetings at Redmond, Washington, and Hawthorne, California. The deal under discussion: Google will pay SpaceX to launch and maintain a series of custom designed server modules on a dedicated, reusable Starship variant. In return, Google gets exclusive low latency compute capacity for its most demanding AI workloads. SpaceX gets a recurring revenue stream that is not tied to consumer internet subscriptions.

“The latency requirement for real time AI inference is single digit milliseconds. Current terrestrial data centers add 20 to 50 milliseconds just from the speed of light through fiber. Orbital data centers cut that to zero if the user is within the satellite’s footprint,” the source explained, paraphrasing a slide from the presentation.

The technical problem is not the launch cost, which SpaceX has already driven below $1,000 per kilogram. The problem is heat. In a vacuum, there is no air to carry away the thermal load from a few hundred kilowatts of GPUs. The memo describes a “radiative panel array” being developed in partnership with a classified military contractor. Think of it as a giant, deployable radiator that uses phase change materials to dump heat into deep space. Google’s Tensor Processing Units will be modified to run at higher clock speeds in cold vacuum, but the real innovation is in the thermal bus. The whole system is designed to fail safe: if the radiators jam, the servers throttle down to 10% capacity and the orbit naturally decays within 72 hours.

“We are not building a data center that happens to be in space. We are building a space native computer,” the memo states. “Every component from the motherboard to the power supply is being redesigned for zero gravity, high radiation, and deep thermal cycling.”

Let’s break down the math here. A single orbital data center, according to the documents, will contain 192 servers in a pressurized, radiation hardened module. Total compute is roughly 2 exaflops of mixed precision AI processing. That is about the same as a mid sized terrestrial hyperscaler campus, but with one critical difference: the orbital data center can be repositioned. The Starship version 3 prototype intended for this project has a dry mass of 120 tons and can carry 50 tons of payload to low Earth orbit. That means the first phase will deploy four modules in a polar orbit at 550 kilometers altitude. The constellation will be managed by a new ground station network built on top of Google’s existing fiber backbone.

The Real Reason Google Needs Orbital Data Centers: AI Inference at the Edge of the Grid

Why the cloud giants are abandoning hyperscale campuses for space

You have heard all the hype about AI. You have not heard the dirty secret: the electricity cost. A single training run for a frontier model can consume as much power as a small town. Google’s own carbon footprint is ballooning despite renewable energy credits. According to the leaked financial model in the memo, orbital data centers offer a net zero carbon solution because the solar panels on the Starship module can generate power directly from the sun without atmospheric attenuation. At 550 kilometers, the sun hits the panels 24 hours a day for about 70% of the orbit. Batteries fill the gaps. But wait, it gets worse. The real driver is not carbon. It is latency arbitrage. Financial trading firms, autonomous vehicle fleets, and military drones all need computing that is physically closer to the end user than any terrestrial data center can provide. A server in Ohio cannot help a drone over the Pacific. An orbital data center over the same spot can. Google’s plan is to sell “orbital compute slices” by the second, using a proprietary API that routes requests to the nearest satellite in the constellation. The memo claims that this will reduce round trip latency for users in remote regions from 300 milliseconds to under 10 milliseconds.

The Starlink integration that nobody is talking about

SpaceX already controls the largest satellite constellation in history. Starlink terminals are everywhere. The memo reveals that the orbital data centers will be directly connected to the Starlink laser mesh, bypassing ground stations entirely for inter satellite routing. This is a huge advantage. Current cloud providers have to send data up to a satellite, down to a ground station, across fiber to a data center, and back. Google’s orbital data centers will keep data in space, process it in space, and beam the results directly to a Starlink user terminal. The document calls this “edge compute in the sky.” Here is the part they left out of the press release: the military angle. The memo mentions a “special access program” with the United States Space Force. Google and SpaceX are not talking about this publicly, but the orbital data centers are being designed to support the Joint All Domain Command and Control concept. The documents include a requirement for “hardened encryption modules” and “autonomous failover without human intervention.” This is not just about cloud gaming or AI chatbots. This is about putting the brains of the military’s sensor network into orbit.

The Skeptic’s View: Why Engineers Are Calling This a Billion Dollar Boondoggle

I called three engineers who work on satellite thermal systems. All of them asked to remain anonymous because they are not authorized to speak about ongoing projects. Their reaction was the same: “This will not work.” The first problem is particle radiation. Low Earth orbit is not empty. It is full of protons, electrons, and cosmic rays that can flip bits in memory and permanently damage silicon. Google’s own TPUs are not hardened for space. The memo acknowledges this and says they will use “triple modular redundancy with error correcting code memory” but that adds cost and latency. The second problem is orbital decay. At 550 kilometers, the atmosphere is thin but not absent. A solar storm can increase drag by an order of magnitude, forcing the data center to fire thrusters to stay aloft. The memo estimates fuel costs at $2 million per year per module. That is cheap if the data center is generating revenue. But the third problem is the most damning: the latency argument is a lie. One of the engineers explained it to me: “The speed of light in vacuum is 300,000 kilometers per second. A satellite at 550 kilometers has a round trip time of about 3.7 milliseconds just from the altitude. But that is for a user directly under the satellite. For a user at the edge of the satellite’s coverage, the distance is longer. The total latency with processing time, beam shaping, and routing easily hits 20 milliseconds. That is not better than a terrestrial data center in a major city. It is only better than a data center in a rural area.”

“The marketing says ‘single digit milliseconds’ but the physics says otherwise,” the engineer said, paraphrasing the internal debate. “They are trying to sell a solution to a problem that does not exist for most customers. The only customers who need that latency are high frequency traders and the Pentagon. And both of them already have their own solutions.”

The Regulatory Cliff: Who Owns the Orbit?

FCC, ITU, and the unbounded race for spectrum and slots

The memo includes a detailed timeline for regulatory approval. Google and SpaceX plan to operate the orbital data centers in a spectrum band that is currently used for Earth observation satellites. The Federal Communications Commission has not yet assigned a sharing framework for commercial compute in space. The International Telecommunication Union has a process for filing orbital slots, but it takes years. The memo suggests a “preemptive filing through a shell company” to avoid public scrutiny. Let’s break that down. The orbital data centers will need to broadcast data down to Earth. That requires a license. SpaceX already has a blanket license for Starlink user terminals, but that license does not cover high power downlinks from a data center. Google will need a separate authorization. The memo warns that competitors such as Amazon’s Project Kuiper and Microsoft’s Azure Space are “working behind the scenes to block the filing.” Amazon has its own satellite internet plans. Microsoft has a partnership with Starlink competitor OneWeb. This is going to be a nasty fight.

Collision avoidance and space debris: the elephant in the vacuum

Orbital data centers are huge. Each module is roughly the size of a school bus. They will be placed in a popular orbit, the same shell where many Starlink satellites operate. The memo includes a collision risk analysis that calculates a 0.3% probability of a conjunction per year. That sounds low until you realize that a collision with a 10 centimeter piece of debris would destroy the entire module. The memo proposes using “autonomous thruster maneuvers” based on data from the Space Surveillance Network. But the network is not perfect. A conjunction warning can come too late. The kicker here is the liability. If a Google orbital data center collides with a Starlink satellite, who pays? Both companies are partners in this project, but they are also separate entities. The memo discusses a “limited liability corporation” that would own the data center assets. That LLC would be separate from both Google and SpaceX. In the event of a disaster, the LLC would declare bankruptcy and the debris would become an orphan event. This is not a conspiracy theory. It is written in black and white on page 47 of the document I reviewed.

The Human Cost: Jobs, Power, and the End of the Cloud as We Know It

I talked to a former Google Cloud engineer who left the company last year. He asked not to be named because he still has friends there. He said the orbital data center project is being sold internally as a way to reduce energy costs, but that hides a darker motivation. “They are terrified of the AI compute demand curve. Every new model generation needs ten times more compute. The power grid cannot scale that fast. So they are looking for any energy source that is not tied to a utility. Solar in space works. But the real reason they want to move compute off Earth is to avoid regulations. No carbon taxes. No water usage restrictions. No local zoning laws. It is the ultimate offshoring.” That is the uncomfortable truth. Orbital data centers are not about technical progress. They are about escaping the physical and political constraints of the planet. Google and SpaceX see the Earth as a bottleneck. The memo explicitly says that “terrestrial data center capacity will be capped by 2030 due to grid constraints.” The solution is to put the compute where the energy is free and the regulators are far away.

The Unanswered Question: Who Actually Controls the Compute?

The documents do not address a fundamental concern: sovereignty. If a country like China or Russia launches a cyberattack on an orbital data center, what happens? The memo says the system will have “hardened firewalls and air gapped control systems.” But an orbital data center is a physical target. A kinetic kill vehicle launched from the ground could destroy a multi billion dollar asset in seconds. The United States Space Force has indicated that it considers commercial orbital assets as “critical infrastructure.” That means the military will want control over the data center’s operation during a conflict. Google and SpaceX have not disclosed how they will handle that. The memo includes a single bullet point: “If a deconfliction request is received from USSPACECOM, the satellite will transition to safe mode and cease all compute operations within 60 seconds.” That means in a crisis, the orbital data center becomes a brick. The customers relying on that compute for autonomous drone navigation or financial trading would suddenly lose connectivity. The memo does not address the consequences.

The Kicker: This Is Happening Faster Than You Think

According to the timeline in the memo, the first test module will launch in the fourth quarter of 2026. That is less than 18 months from today. The module will carry a nonfunctional heatsink and power system to validate the thermal performance. If that test passes, the first production module will go up in mid 2027. Google has already allocated $2.5 billion in capital expenditure for the first four units. Here is the part that keeps me up at night. The memo concludes with a sentence that has been redacted from the copy I received, but the context makes it clear: the orbital data centers are designed to be fully autonomous. No human astronauts. No servicing missions. If something breaks, the server just dies. The satellite is deorbited and replaced. This is not a long term sustainable model. It is a disposable infrastructure approach that treats space as a temporary landfill for high value hardware. And that is the real story. Orbital data centers are not a futuristic dream. They are a desperate, high risk bet by two of the most powerful companies on Earth. They are betting that the laws of physics, the regulations of nations, and the patience of the public will bend to their will. We will find out in 2027 if they are right. Until then, keep your eyes on the sky. The cloud is about to get a lot higher.

Frequently Asked Questions

What are orbital data centers?

Orbital data centers are data storage and processing facilities placed in orbit around Earth, typically on satellites or space stations.

Why are Google and SpaceX collaborating on orbital data centers?

The talks aim to combine SpaceX's launch capabilities with Google's cloud computing expertise to solve problems like latency and bandwidth congestion.

How would orbital data centers benefit internet users?

They could reduce latency for global communications and support low-Earth orbit satellite networks, improving real-time applications like gaming and video calls.

What challenges do orbital data centers face?

Key challenges include high launch costs, radiation exposure damaging hardware, and the need for reliable power sources like solar panels.

Are there any environmental concerns with orbital data centers?

Potential issues include space debris from defunct satellites and increased energy consumption for cooling in space, though proponents argue they reduce Earth-based data center footprints.