Where Compute Goes for Power

On 1 October, a refrigerator-sized satellite carrying four Google Tensor Processing Units entered low Earth orbit.

The satellite is part of Project Suncatcher, Google’s attempt to answer a question that would have sounded unnecessarily extravagant not long ago: could large-scale machine-learning infrastructure eventually operate in space?

For now, the experiment is modest. Google is not building an orbital data centre. It wants to know whether the same kind of specialised chips used for AI workloads in terrestrial facilities can survive launch, radiation and the thermal conditions of orbit. The company has confirmed contact with the satellite and says it is functioning as expected. Over the coming weeks, engineers will collect data about how the hardware behaves in an environment where replacing a failed component requires considerably more effort than walking across a server hall.

Yet Google’s research programme is not modest in its destination. Project Suncatcher imagines groups of solar-powered satellites carrying AI accelerators and communicating through optical links fast enough for the machines to work together. Its published research describes an illustrative cluster of 81 satellites flying within a kilometre-wide formation and eventually operating with the sort of high-bandwidth connections required by modern machine-learning systems.

There are obvious reasons to treat this as a curiosity of the space industry. There are also good reasons to doubt that an orbital data centre will soon compete with a warehouse full of servers in Virginia, Ireland or Singapore.

The more interesting part of Project Suncatcher is not space.

It is electricity.

Google’s argument begins with an unusual property of orbit: sunlight is much easier to collect there. A solar panel in an appropriate low-Earth orbit can receive near-continuous sunlight and, according to Google’s analysis, potentially generate up to eight times as much energy over a year as a similar panel at a mid-latitude location on Earth. Instead of generating electricity in space and trying to transmit it back to the ground, an old and technically difficult idea, Suncatcher proposes moving the computation to where the energy is available.

That reversal tells us something about what computing is becoming.

For much of the digital era, geography appeared to matter less. Software moved from machines sitting under desks into remote data centres. Those data centres became “the cloud”, a metaphor that was useful precisely because it concealed location. Applications could be consumed almost anywhere without the user needing to know where their computation physically occurred.

But the abstraction was always incomplete. The cloud lives somewhere.

It needs substations, transmission lines, cooling systems, backup power, fibre routes, land and enormous collections of physical machines. The infrastructure became less visible to users at roughly the same time that it became more concentrated and industrial.

AI is making that physical layer harder to ignore.

The International Energy Agency estimates that worldwide data-centre electricity consumption could more than double to around 945 terawatt-hours by 2030. That is slightly more electricity than Japan consumes today. AI is expected to be the largest single contributor to the increase, although conventional computing and other digital services remain important.

Globally, data centres would still account for only a few percent of electricity demand. That qualification matters. Claims that AI is about to consume the world's electricity are not supported by current forecasts.

The difficulty is geographical concentration.

A steel plant, a residential neighbourhood and thousands of electric cars distribute their electricity demand differently. A large data centre may require hundreds of megawatts or eventually several gigawatts at one site. The IEA notes that this concentration can make data centres much harder for electricity systems to accommodate even when their share of global consumption appears modest. In the United States, data-centre expansion is expected to account for around half of additional electricity demand through 2030.

The constraint is therefore not simply whether the world can generate enough electricity in aggregate. It is whether enough electricity can be generated and delivered at a particular place, on a particular timetable, with sufficient reliability.

That changes where computing goes.

The data centre is following the power

The traditional geography of digital infrastructure was shaped heavily by telecommunications. Data centres clustered around fibre routes, major population centres and established internet exchanges. Latency mattered. Being physically closer to customers reduced the time required to move data between them and the machines processing it.

Energy is increasingly joining connectivity as a first-order site-selection constraint.

This explains why technology companies are entering parts of the energy system that once sat well outside what would have been considered a computing business. Google has signed agreements supporting advanced geothermal and nuclear generation. Microsoft, Amazon and other technology companies have pursued their own nuclear and renewable-energy arrangements. The intention is not simply to improve environmental credentials. Reliable long-term electricity supply is becoming part of the capacity required to expand computing.

The distinction matters because buying electricity is different from ensuring electricity exists.

A rapidly growing data centre can sign a contract with an electricity supplier, but somebody still has to generate the energy. Transmission capacity must connect the generator to the site. A substation has to handle the load. Equipment has to be manufactured. Planning permission has to be obtained. In many electricity systems, these processes move much more slowly than the technology industry is accustomed to building servers.

The IEA points to precisely this mismatch. A data centre can sometimes be developed within two or three years; major parts of the electricity system require much longer planning and construction periods.

Compute can therefore encounter a strange constraint. The chips exist. The capital exists. Customers may exist. What does not exist is a grid connection capable of powering the machines.

At that point, computing strategy becomes infrastructure strategy.

Project Suncatcher is an extreme version of the same response. If energy becomes difficult to bring to the machine, change where the machine lives.

Space solves one constraint by creating several more

The elegance of this idea disappears as soon as the engineering begins.

A terrestrial data centre has an enormous advantage: Earth.

It sits on foundations. Engineers can visit it. Broken hardware can be removed. Fibre can be buried into the building. Cooling systems can move heat into water or air. New chips can replace old ones when computing technology improves.

Orbit removes many of those conveniences.

Cooling is one of the least intuitive examples. Space is cold, but a vacuum is a difficult place to cool high-performance electronics. On Earth, heat can be transferred into moving air or cooling liquid. In a vacuum, there is no surrounding air to carry it away. Heat ultimately has to be radiated.

Google is experimenting with heat pipes and radiator surfaces, and describes thermal management as one of the central unresolved engineering problems for Suncatcher.

That matters because modern AI accelerators concentrate enormous amounts of heat into a small area. A scaled orbital computing system would not simply require solar panels and chips. It would need large physical structures devoted to getting rid of the heat those chips produce.

Networking presents another problem. Training a large AI model does not involve thousands of independent processors quietly doing separate jobs. They exchange enormous amounts of information while computing together. Google's terrestrial systems use specialised networks with extremely high bandwidth and low latency.

Ordinary satellite communications are nowhere near sufficient.

Google estimates that a useful orbital cluster would require links on the order of ten terabits per second between satellites. Its laboratory work has demonstrated a 1.6-terabit-per-second bidirectional optical connection over a short free-space path, but translating that into a moving constellation requires satellites to remain in precise formations while lasers maintain links between them.

Then there is reliability.

Google's ground testing found that its Trillium TPUs could withstand cumulative radiation levels above those expected during a five-year mission, although radiation-induced errors remain something the company must manage. The problem becomes harder at system scale. A terrestrial cluster can tolerate failures partly because a technician can replace a failed board. Sending a technician to a compute satellite is not an ordinary maintenance call.

The architecture therefore has to assume failure differently. Redundancy, fault tolerance and hardware lifetime become part of the economic calculation.

And eventually there is the simplest problem of all: everything has to be launched.

Google’s own paper does not claim that this is economically viable under present conditions. Its optimistic scenario depends on launch prices falling dramatically, potentially towards $200 per kilogram by the middle of the 2030s. That projection assumes continued improvements in reusable launch systems and very high launch volumes.

An independent 2026 analysis of orbital data-centre economics is considerably more cautious. It concludes that abundant solar energy alone cannot make the system competitive: radiator mass, spacecraft construction, communications, utilisation, hardware lifetime and launch expense all have to work simultaneously. Under its representative assumptions, general-purpose orbital computing remains economically difficult, while workloads that originate in space or require significant processing before data is sent to Earth look more credible as early applications.

That distinction is useful. The first successful space-based data centre may not be a replacement for one in Arizona.

It may be a computer that happens to be near the data it needs.

Infrastructure changes what counts as “near”

This is where the Suncatcher experiment becomes more interesting than its futuristic appearance.

Computing has always been located near whatever resource is hardest to move.

In the early internet, that resource was often connectivity. Data centres accumulated around network exchanges and fibre. For real-time consumer services, they still do.

For some forms of AI, energy may become increasingly important.

If a workload can tolerate delay and does not continuously exchange huge amounts of information with users on Earth, it has more freedom to move. Training, simulation, scientific computation and processing information generated by satellites are different from serving an interactive request that needs an answer in milliseconds.

This suggests that the future geography of computing may become more specialised.

Some computation will remain close to people because latency matters. Some will remain near corporate data because moving sensitive or enormous datasets is difficult. Some may locate near cheap generation because electricity dominates the economics. If orbital infrastructure ever becomes viable, the earliest workloads are likely to be those for which its unusual combination of plentiful solar energy and difficult communication actually makes sense.

That would be a reversal of how people normally think about the cloud.

Instead of computation being everywhere, computation becomes increasingly deliberate about where it is.

The terrestrial consequence arrives first

It would be easy to finish this argument by imagining enormous constellations of AI satellites.

That is probably the least useful implication for someone living through the next decade.

The terrestrial consequence is already here.

Electricity availability is becoming an input into decisions once regarded primarily as technology decisions. Data-centre developers are looking not simply for suitable buildings but for megawatts. Technology firms are negotiating directly with power producers. Grid operators must plan around large industrial loads arriving faster than traditional electricity-demand forecasts assumed.

Those choices have consequences because the infrastructure is shared.

A new transmission line can support homes, factories and data centres together. New generation built for a large technology customer can increase capacity elsewhere in the system. Large guaranteed buyers may also make projects such as advanced nuclear or geothermal energy easier to finance.

But scarcity can work in the other direction. Where generation and grid capacity do not expand quickly enough, new large loads compete for limited connections. Costs associated with network reinforcement have to be allocated somewhere. Communities decide which developments receive land and planning approval. Governments decide whether digital infrastructure deserves preferential access to energy.

Those are not questions about artificial intelligence in isolation.

They are questions about infrastructure priorities.

This is the part of the AI expansion that disappears when the discussion stays at the level of models and applications. A larger model looks like a software achievement from one direction and a request for electricity, cooling and capital from another.

Both descriptions refer to the same system.

What the satellite tells us

Google's satellite is currently testing four TPUs.

That is very far from a functioning orbital data centre. Google still has to demonstrate sustained thermal management, radiation tolerance in real operating conditions, high-bandwidth links between moving satellites, reliable ground communication and an economic model that survives the cost of launching and replacing hardware.

Project Suncatcher may ultimately prove that some of those problems are harder than the terrestrial constraints it is trying to escape.

If so, the experiment will still have revealed something important.

Google did not put AI chips into orbit because space is a convenient place to operate computers. Almost every aspect of operating them becomes more difficult there.

It did so because space possesses one resource that increasingly matters to computing: extraordinary access to energy.

For decades, digital infrastructure became easier for ordinary users to ignore. We moved our documents, businesses and communications into a cloud whose physical location rarely entered the conversation.

Now the physical layer is reappearing.

The question raised by Suncatcher is not whether tomorrow’s chatbot will answer from a satellite.

It is why the location of the machine answering it has started to matter so much again.

References

Google Research, Project Suncatcher prototype satellite is in orbit, 1 October 2026. blog.google

Google Research, Exploring a space-based, scalable AI infrastructure system design, November 2025. Google Research

Agüera y Arcas et al., Towards a future space-based, highly scalable AI infrastructure system design, Google Research / arXiv, 2025; subsequently peer-reviewed in Joule. arXiv

International Energy Agency, Energy and AI, data-centre electricity-demand analysis. IEA

International Energy Agency, Electricity 2026, electricity-demand outlook through 2030. IEA

Turyshev, Orbital Data Centers: Spacecraft Constraints and Economic Viability, 2026. https://arxiv.org/abs/2604.27197?