Why Is Google Sending an AI Chip Into Space?

5 things you need to know

1. Google's first orbital AI chip launched on a SpaceX rocket.
2. The test will show if a TPU can get 1 kilowatt of power, stay cool and run models in space.
3. A two-satellite demo follows next year, linked by laser.
4. The long-term plan is an orbital data center of 81 satellites.
5. The plan needs Starship to fly about 180 times a year. It has never flown more than five.

The Big Number: 180. That's roughly how many Starship flights a year Google's paper implies it needs (our derived figure from about 1,800 launches over 10 years). Starship's record is five.

The Money Map: where the cost sits

• Chip: Google's own TPU, so no Nvidia dependence for this test.
• Satellite: a standard Planet Labs platform for the prototype. Purpose-built ones come next year.
• Launch: the biggest unknown. The paper assumes about $200/kg by 2035, but the authors say they aren't claiming economic feasibility.
• Power and cooling: the point of the test. Power is the thing orbit is meant to make cheaper, and cooling is the thing orbit makes harder.

What changed

• Before: orbital data centers were mostly a concept paper.
• Now: hardware is in orbit, and the paper has passed peer review.
• Also changed: Google redid its radiation tests after finding its first setup was too protective. The corrected tests showed slightly more errors.

What caught our attention

• Google calls this a long-term moonshot and openly says the rockets it needs don't exist yet.
• The paper says it isn't an economic feasibility study, yet it carries a detailed launch-price forecast.
• Google is a major investor in SpaceX, its launch provider and a company that stands to gain if launch demand grows.
• Radiation errors are fine for everyday AI use (about one in a million operations, per Google), but were already a concern for months-long training runs.

Finance Speak → Human Speak

• TPU: Google's custom chip for AI, like a specialised engine.
• Inference: using a trained AI model to answer questions, as opposed to training it.
• Learning curve (20% a year): each time you build more, the cost drops, and here it's assumed to drop about 20% a year.

Rainingmoney's reading (not Google's): the hardware test is likely to succeed in some form. The harder-to-forecast part is whether launch costs fall fast enough, which is why this is a decade-long project and not a product.

Comprehension check

• Why does Google run the chip in 15-minute bursts? To avoid straining the satellite's power and heat systems.
• What would Starship need to do for the plan's cost assumptions to hold? Fly roughly 180 times a year.

The Bottom Line: Today's launch is a small test of a very large idea. The chip is the part Google controls, while the cost of getting to orbit is the part that decides whether the idea works.