Will we ever find ourselves saying a sentence like this? Can we even imagine that, 10 years from now, thousands of processors capable of performing in orbit what they currently do in refrigerated, purpose-built facilities on Earth could be launched into space? In this edition of INESC TECWatch, we set off in search of answers with the help of our researchers in advanced computing.
When we look up at the night sky, we may not realise it, but what we see is no longer what we saw before, or during those astronomy sessions long ago – when, standing on tiptoe, we peered through a telescope at the star-filled canvas extending before our eyes. In a classic “before and after” comparison, the feeling would be much like watching skin gradually acquire new freckles over time.
Some of those spots are satellites – thousands of them. We sent them there, and there they remain. We see them at night as tiny white dots scattered across a dark blanket. But in a decade or two, will we notice even more bright points overhead? Could we one day look up and see a constellation of… data centres? It’s difficult to say, although some people are pretty sure about it. And yes, we are talking about structures the size of industrial parks being launched into orbit, where they would rely on vast solar arrays for power; and they could be far more visible than most of today’s satellites.
The first edition of The New York Times in 2026 echoed this vision – one that owes remarkably little to even the boldest science fiction. The newspaper highlighted a growing number of statements from leaders in the Artificial Intelligence and Space sectors. All point in the same direction and follow what appears to be a straightforward line of reasoning: data centres are consuming ever more energy, placing increasing pressure on the communities that must coexist with these new “neighbours”, while also requiring large areas of land. The proposed solution? Move them into space. The main argument is compelling at first glance: plentiful energy, near-constant access to sunlight, and no clouds standing between the Sun and solar panels.
The leading voices behind this vision come as no surprise. In November last year, Google unveiled their plans: by 2027 (through the Suncatcher programme), the goal is to begin testing satellites equipped with Google TPUs powered by solar energy. Elon Musk has also argued that “global demand for electricity to power AI cannot be met with Earth-based solutions alone, even in the short term, without causing difficulties for communities and the environment”. SpaceX’s ambition is to place in orbit the kinds of data centres currently operated on Earth by the usual technology giants, including Google, Amazon, NVIDIA and OpenAI.
Closer to home, the European Space Policy Institute (ESPI) estimates that the global market associated with data centres by 2030 – regardless of where they are located – could reach €535 billion. According to ESPI, Europe cannot afford to miss out on a new golden age capable of generating significant multiplier effects across the wider economy.
All of this sounds promising. Yet the crucial question remains: would orbital data centres be viable? Can data-centre processors operate effectively in space? There are also financial questions left unanswered. Detailed calculations fall beyond the scope of this article, but the costs associated with the vision promoted by major technology CEOs are, unsurprisingly, astronomical. Launching a single kilogram of material into space can cost between €2,000 and €8,000 – and these facilities would be measured in stadiums rather than square metres.
Vacuum, gravity and extreme temperatures
Let’s begin by examining the issue from a purely computational perspective. Ricardo Macedo, an INESC TEC researcher specialising in advanced computing, explained: “A GPU operating in orbit could deliver performance comparable to that of a GPU in a terrestrial data centre, because the underlying architecture and processing power are essentially the same.”
However, he stated that “many GPU workloads, like training or inference for AI models, depend on frequent communication between GPUs, CPUs and storage systems”. In orbit, available bandwidth and the latency of connections to other systems are significantly more constrained than in a ground-based data centre, which can affect the overall performance of applications.
That challenge alone is significant. Yet there is another. According to INESC TEC researcher Miguel Peixoto, a GPU operating in space is exposed to radiation capable of causing logical errors, including bit flips and other soft errors, which must be corrected by software. This correction process consumes computing resources of its own. “As a result, the same GPU may deliver lower performance in space than it would on Earth,” he explained.
Far from the stable conditions of Earth, temperatures in space can plunge to around -250°C, while direct sunlight remains almost constant. This creates an additional challenge for computing infrastructures. Miguel Peixoto explained: “Data centres on Earth can employ a range of cooling systems, often combining multiple technologies. In space, however, the absence of convection as a heat-dissipation mechanism, the effects of microgravity, and extreme temperatures – very low in shadow and extremely high under direct solar exposure – require solutions specifically adapted to the environment, greatly limiting the available options.”
Modern data centres primarily rely on air, water or oil-based cooling systems. “Air cooling is not feasible in space because convection does not occur in a vacuum,” said Ricardo Macedo. “Liquid-based solutions are also more difficult to implement and operate because of pressure conditions and engineering constraints specific to the space environment.” The consequence, he argued, is clear: “thermal management systems would need to be fundamentally rethought, requiring further research before practical and efficient solutions can be deployed at scale”.
Cooling the back room
Given all these constraints, could we realistically see a fully operational orbital data centre within the next decade? Perhaps not. Ricardo Macedo provided a more plausible scenario: the launch of small-scale proof-of-concept missions over the coming years. NVIDIA has already taken a first step by sending the H100 into space – a chip with around 100 times the processing capability of any processor previously deployed beyond Earth. “This is the first time a terrestrial-grade data-centre GPU will be sent into space and operated in orbit,” stated the CEO of Starcloud at the time of the launch, in November 2025. The company aims to place a 40-megawatt data centre in orbit within the next decade, processing data at costs comparable to those of terrestrial facilities.
In orbit, the H100 will attempt tasks it has performed countless times on Earth, including AI-processing applications such as analysing Earth-observation imagery and running large language models (LLMs). These early missions can be perceived as the equivalent of pre-season testing: an opportunity to refine technologies and address critical challenges before entering a far more demanding phase. Among those challenges are “energy constraints, thermal dissipation, radiation tolerance and communication limitations”, stated Ricardo Macedo. Solving this equation will be essential if such systems are ever to be certified and deployed at a scale comparable to that of modern terrestrial data centres.
Miguel Peixoto pointed out the International Space Station (ISS) as an example. The ISS relies on a “highly sophisticated cooling system” capable of “dissipating up to 70 kW of heat”. Today, that capacity would be sufficient for “only two racks within a conventional data centre”. As Peixoto emphasised, modern facilities “contain anything from dozens to hundreds – or even thousands – of racks”. While the ISS’s Active Thermal Control System (ATCS) has demonstrated that cooling high-performance computing infrastructure in space is possible on a very small scale, “terrestrial HPC and AI data centres operate at a scale far beyond anything currently supported by the ISS,” he explained.
Then there is the issue of launch itself. “Current estimates suggest that approximately 20 tonnes of fuel are required for every tonne of material sent into space,” said Miguel Peixoto. Many terrestrial data centres already draw at least part of their energy from renewable sources. “That means we are still a long way from achieving a lower carbon footprint in space.”
The environmental impact of increasingly frequent launches must also be considered. Continuous launch campaigns by companies such as SpaceX can generate hundreds of tonnes of CO₂ emissions; add to that the potential growth of space tourism and mega-constellations such as Starlink and Kuiper, and the cumulative impact could increase substantially over the coming decades. “The overall impact will depend heavily on the type of workload involved,” said Ricardo Macedo. “For applications requiring continuous communication with Earth, system efficiency may be reduced, and the overall carbon footprint could increase because of the environmental cost associated with space launches.”
There may, however, be situations in which orbital computing offers genuine advantages. If data processing takes place directly on information generated in space – such as Earth-observation or astronomical data – and is carried out locally rather than transmitted back to Earth, the need for large-scale data transfers “could be significantly reduced, potentially delivering gains in efficiency”.
The balance between these advantages and disadvantages will shape the debate in the years ahead. The question that remains is whether one day we will look up at the night sky and say: “That data centre wasn’t there yesterday.”
News, current topics, curiosities and so much more about INESC TEC and its community!
