Business & Industry

Google's Orbital Data Centers Have a Rocket Now. The Economics Are Still Missing.

Google has a rocket for its plan to put AI hardware in orbit. What it doesn't have yet is a launch price, a named satellite builder, or a way around the physics that makes space a hard place to run a server.

Daniel HarrisDaniel Harris
Heat: 1,240
Three satellites with solar arrays above a planet horizon, linked by a docking rail

Dek: Google has a rocket for its plan to put AI hardware in orbit. What it doesn't have yet is a launch price, a named satellite builder, or a way around the physics that makes space a hard place to run a server.

The Project Suncatcher launch deal with Blue Origin

Blue Origin announced on Thursday 19 November that New Glenn will launch data center satellites for Google in 2029.

That's the whole deal, and it's worth being precise about the shape of it. This is a launch services agreement for a mission, not a contract to build and operate a constellation. Google is buying a ride, not a fleet.

The mission belongs to Project Suncatcher, the research program Google disclosed about a year earlier, built around solar-powered satellites carrying Tensor Processing Units.

Travis Beals, who runs the program at Google, described it as a long-term moonshot to scale AI infrastructure in orbit using the solar power available up there, and called the dedicated launch a milestone on the roadmap. That's the company's framing, and it's doing a lot of work in a short sentence.

What New Glenn carries: a rack of Google TPU compute

Each satellite is designed with computing capacity at roughly the scale of a data center rack, built around Google TPU chips.

The orbit is the interesting choice. The satellites go to a Sun-synchronous path, crossing the same part of Earth at the same local time on every pass, with the panels lit for almost the whole loop.

Power is the reason. A rack on Earth plugs into a grid and pays for the connection. A rack in orbit either sees the sun or runs on batteries. Google hasn't published a power budget for the satellites, so the numbers behind that orbit choice stay inside the company for now.

The weak part starts when the compute has to talk to anything else.

The 7x2 New Glenn has flown three times

The vehicle detail matters more than it sounds. The mission uses the 7x2 variant of New Glenn, the smaller configuration of Blue Origin's rocket.

New Glenn has launched three times to date. Three flights is a thin record for a rocket that's supposed to carry several tons of server hardware to a specific orbit in 2029, and Blue Origin's own schedule has slipped before.

The larger 9x4 variant, the one Blue Origin expects to be flying in the 2030s, isn't the vehicle named here. Google is booking the smaller rocket.

Project Suncatcher in 2029: a demo, and no named satellite builder

The first launch isn't scheduled until early 2029, and Ars Technica reads the mission as a technology demonstration: a test of networking solar-powered satellites that carry AI chips.

Google didn't say who is building the satellites. Ars Technica notes they may weigh on the order of several tons, which puts them in a class where the bus, the thermal design and the power budget are all custom work.

So the 2029 flight answers one question: can these satellites pass traffic between them while running real workloads? Everything else, including whether the idea survives contact with a launch manifest, comes later.

Why orbital data centers can't book a rocket in 2028

Here's the constraint that has nothing to do with Google's software. US commercial heavy-lift capacity is booked out through 2030, according to Ars Technica, including New Glenn's own manifest.

Google said a year ago that it was planning for the mid-2030s, when launch costs are projected to fall to as little as $200 per kilogram. Getting there probably requires a fully developed Starship V4 with roughly 200 tonnes of lift per flight, and SpaceX has been prioritizing its own payloads, including Starmind data centers.

That's the awkward part for orbital data centers as a category. The company with the cheapest heavy lift is also a competitor in the same business.

Blue Origin has its own plans too. The company is working on Project Sunrise, an orbital data center constellation of its own, which means Google may be helping to fund a rival's learning curve while it waits for a rocket.

Orbital data centers face heat, radiation and a $74 million launch

The physics hasn't moved since the last time anyone checked.

Cooling in vacuum doesn't work like cooling in a server room. Heat has to leave through pipes and radiators, and a radiator in space is limited by area, not by how fast you can spin a fan.

Radiation is the second problem. Energetic particles flip bits in memory, which means flight hardware needs error correction that ground hardware gets to skip.

Then there's the price of a ride. A single Falcon 9 launch runs about $74 million, per Engadget, and that's before the satellite. A rack of TPUs that has to be rad-hardened, thermally managed and networked to its neighbors costs more than the same rack in a leased building in Iowa.

Engadget also notes that more than 47,000 tracked objects are in low-Earth orbit, which is a sentence about debris rather than a prediction. It's the reason insurers ask questions.

Reasonable ones.

What to watch in the 2029 Suncatcher flight

Two numbers will tell you whether this is a program or a press release.

The first is inter-satellite throughput on the demo. If two satellites can pass work between them at a rate that makes a cluster useful, the networking argument survives. If the demo is one satellite doing a job alone, Google has proved less than the announcement implies.

The second is launch price. Until somebody quotes a per-kilogram figure near the $200 target, every orbital data center pitch is a physics paper with a rocket attached.

Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon, told The New York Times that going from one satellite to a network that operates like a giant data center "will take years and enormous funds". He's describing the gap between a demo and a service, and 2029 is on the demo side of it.

Google's October prototype, a single AI server satellite that can work for about 15 minutes at a time, is the honest baseline. The 2029 flight is meant to move past it. Whether it does is a question for a launch manifest that doesn't exist yet.

Share This Story

Sources

Related AI News

A falling candlestick chart with a magnifying glass over ledger lines below it
Business & Industry

Nvidia Lost $900 Billion in a Week and the SEC Is Now Reading the Books

The stock fell about 14% on Friday and has now dropped for seven straight weeks. The SEC examination reported the same day is about the books, not about the founder.

Heat: 1,950
A globe grid with data center blocks and a toll barrier drawn across one route
Business & Industry

Trump Wants a Tax on Data Centers Abroad. Nobody Has Written One Yet.

The president told reporters on Friday that companies building AI capacity overseas would face a "tremendous tax." No rate, no bill and no agency has been named to collect it, which turns the announcement into a story about intent.

Heat: 1,880
Five boardroom chairs with the middle one empty and a chip on the table in front of them
Business & Industry

Nvidia Ran a Week Without Jensen Huang and the Numbers Didn't Notice

Nvidia lost its founder, refused to take him back, and then shipped the fastest product ramp in its history inside seven days. The interesting question isn't whether the quarter was good. It's who decides what happens next.

Heat: 1,830
An envelope passing between two outlined blocks with a price marker dropping toward it
Business & Industry

Intuit Sold Mailchimp for a Third of What It Paid

Intuit bought Mailchimp for $12 billion in 2021 and announced on 19 November that it's selling the business for $3.68 billion. The gap is the story, and the structure of the deal is stranger than the headline number.

Heat: 1,310