Google has begun testing artificial intelligence chips in orbit, the company's first confirmed experiment aimed at answering a question the hyperscale industry has circled for years: can data centres work in space?

The tests form part of Google's broader research programme into space-based data centres, an approach that proponents argue could eventually tap abundant solar power and continuous cooling in the vacuum of orbit. Google has not disclosed which chips are flying, on which spacecraft, or in what quantities — but the fact that real silicon is now operating above the atmosphere moves the idea from conference slideware into engineering reality.

Why put compute in orbit at all?

The commercial logic is straightforward. Terrestrial AI training and inference consume gigawatts, and grid constraints now shape where hyperscalers can build at all. In orbit, a solar array receives unfiltered sunlight around the clock, with no clouds, no night cycle and no competition for land or water. Radiating waste heat into space is harder than on Earth — vacuum has no convection — but engineers argue that large radiator surfaces can manage it.

For Google, which designs its own Tensor Processing Units (TPUs) for AI workloads and operates one of the world's largest compute estates, the experiment is a natural extension of its infrastructure roadmap rather than a side project. If orbital compute proves viable, the company that owns the full stack — chips, software, and cloud distribution — would capture most of the value.

What do the tests actually involve?

Details remain scarce. What is confirmed is the core fact: Google is testing AI chips in orbit as part of its space data centre research. The trials come as the cost of launching mass to low Earth orbit has fallen sharply, driven by reusable rockets — a prerequisite that simply did not exist when earlier space-compute concepts were floated.

Several engineering hurdles stand between a test payload and an operational orbital data centre:

• Radiation tolerance: commercial AI accelerators are built for data centres, not for the radiation environment of orbit, which can cause bit flips and degrade silicon over time.
• Thermal management: high-power chips in the tens to hundreds of watts per package must shed heat radiatively.
• Maintenance: failed boards cannot be swapped; architectures must be fault-tolerant by design.
• Data links: feeding models and returning results requires high-throughput, low-latency ground-to-space connectivity.

Each test flight is designed to measure how commercial-grade accelerators behave against these constraints.

Who else is chasing orbital compute?

Google is not alone in probing the concept. A cluster of startups has formed around space-based data processing in recent years, and the Defence Advanced Research Projects Agency (DARPA) has run programmes examining distributed computing in orbit. What distinguishes Google's effort is scale of intent: a hyperscaler with its own silicon, its own AI workload demand, and the balance sheet to fund multi-year infrastructure bets.

The experiment also lands amid a wider reordering of the AI compute supply chain. TSMC's leading-edge capacity remains the industry bottleneck, advanced packaging is booked out, and every major cloud operator is scrambling to secure accelerators. A future in which some fraction of AI inference — particularly latency-tolerant batch work — runs in orbit would add a new, unconventional tier to that supply chain.

What happens next?

The immediate significance of Google's orbital chip tests is informational, not industrial: the company is collecting flight data that will tell it whether space-qualified AI compute is an engineering problem or a physics problem. No operator has yet committed to deploying production data centres off-planet, and any such move would depend on launch costs continuing to fall and on radiation-hardened or radiation-tolerant accelerators proving their reliability over multi-year missions.

For now, the chips are up, running, and phoning home — and the data they return will shape whether the next decade of AI infrastructure is built only on the ground.