Amid the planetary polycrisis – which is, among other things, a crisis of capital – AI data centers appear to be one of the last places offering opportunities for secure capital accumulation. In this context, the cloud, which has long been mistakenly portrayed as an immaterial entity, is also being redefined as a factory. Miriam Fahimi reflects on the political struggles at stake.
*
March 18 is a day of remembrance that has not yet been given the importance it deserves, explains Chancellor Friedrich Merz in his government statement to the European Council on that day. He is referring to the Day of Democratic History: On March 18, 1848, around 300 people died in Berlin in the struggle against the Prussian monarchy and for democratic rights. Exactly 142 years later, on March 18, 1990, the first free election to the People’s Chamber of the GDR took place.
Day of democracy?
Yet the connection Merz draws from history to the present bears little resemblance to the political struggles of the past. Artificial intelligence (AI), in particular, takes center stage in his speech: to enable more “artificial intelligence made in Europe,” Europe must dismantle regulations and give the technology more leeway. After all: “Europe regulates too much, even in the field of artificial intelligence.” With this, Merz is alluding to the European AI Regulation, the world’s only legal framework that has introduced bans on and protective mechanisms against AI – such as against algorithmic discrimination – based on five risk levels.
While one searches in vain in the Chancellor’s speech for a connection between the ‘more AI’ rhetoric and democracy, in a side room of the Bundestag, there is no longer any mention of democracy at all. For there, Digital Minister Karsten Wildberger and Economy Minister Katharina Reiche, both of the CDU, are presenting Germany’s first National Data Center Strategy. The general sentiment is that more ‘AI made in Europe’ also requires more data centers ‘made in Germany.’ Over the next five years, data center capacity in Germany is set to double, and quadruple for AI data centers. The federal government won’t even need to provide much additional support, because according to a study, data center capacity in Germany is expected to at least triple, reaching 5,000 to 5,500 megawatts.
Inside a data center
Although there are an estimated 3,000 data centers in Germany, they constitute a well-protected, rather inconspicuous infrastructure that the vast majority of internet and AI users have never set foot in. Their history began with the first mainframe computers in the late 1940s and evolved in the 1990s – with the development of microchip processors, server-based networks, and the Internet – into what we now understand by the term: places where not only computation is performed, but data is also stored and transmitted, and AI is trained.
Data centers are therefore indispensable for all digital services, whether it’s writing and sending emails, online shopping, streaming, gaming, or AI. They are usually long, single-story, box-shaped buildings with excellent connections to the road and fiber-optic networks, cooling towers on the roof, and a substation on the grounds.
Inside a data center, servers are literally stacked one on top of another: in the white space, they sit in racks, which in turn are housed in enclosures that are well-sealed against dust and dirt and cooled to 21 degrees. AI data centers require chips produced specifically for this purpose, which are optimized for parallel computing and thus not only consume a great deal of energy but also place new demands on cooling. Because of their power density, heat can no longer be dissipated by air alone and is therefore cooled with fresh water applied directly to the chip. Because chips and servers in AI data centers can be added modularly – allowing computing power to be expanded flexibly – these facilities are also called hyperscalers. But they are also ‘hyper’ in the truest sense of the word; for example, META’s planned Hyperion is roughly the size of Manhattan.
For data centers in the U.S., Berkeley Lab has calculated that they consume 628 million gallons (equivalent to 2,377,238,610 liters) per day. However, the majority of data centers in the U.S. rely on evaporative cooling, whereas in Germany, circular water-based cooling systems are predominantly being considered. In circular systems, groundwater is not consumed directly, but it must still be made available to data centers first. If capacities are to be doubled or quadrupled, as planned by the current federal government, this means not only more buildings, servers, chips, and electricity, but also more coolant and water for cooling, more diesel and more batteries for emergency power supply, more gas for the fire suppression system, and more CO2 in the air.
New strategies for regressive politics
Let’s take a closer look at the data center strategy. The three overarching themes – ‘Energy and Sustainability,’ ‘Location and Space,’ and ‘Technology and Sovereignty’ – address key demands of the data center industry. These range from accelerated and more flexible connections to the power grid and faster, simplified permitting processes, to the rezoning of brownfield sites and the strengthening of the German IT sector.
While sustainability is defined exclusively in terms of renewable energy, just a few days ago these very criteria were watered down in an amendment to the Energy Efficiency Act. Germany was the only EU country to impose minimum requirements on data center operators in its national implementation of the Energy Efficiency Act.
Now, the goal of powering data centers 100% with renewable energy by 2027 has been postponed to 2030. What’s more, to solve Germany’s energy problem, Economy Minister Katharina Reiche – who previously served as chairwoman of the E.ON subsidiary Westenergie – is pushing for new fossil-fuel-powered projects. It is thus not surprising that it is E.ON, which is currently building Germany’s first gas-fired power plant for a U.S. data center in Frankfurt am Main.
Critical scholars have already spoken out against the plans for ‘technology and sovereignty.’ With regard to the data center strategy, sovereignty can best be described as regressive-territorial. This refers to a form of sovereignty that is exclusively concerned with locating data centers on one’s own soil, without addressing questions of ownership and dependencies at other levels of the supply chain – such as those involving the U.S. chip manufacturer NVIDIA. What is overlooked, however, is that the majority of the 1,000 hyperscalers are owned by just three U.S. companies: Google, Microsoft, and Amazon. And these companies are also investing heavily in Germany: Of the 15 billion euros invested in Germany’s data centers last year, half came from U.S. companies.
Worsening energy crisis
“The current energy crisis is worse than any other crisis before it,” said Fatih Birol of the International Energy Agency (IEA) recently. This refers not only to dwindling fossil fuel reserves but also to the geopolitical instability resulting from Russia’s war of aggression against Ukraine, as well as the military aggression by Israel and the U.S. against Gaza, Iran, Lebanon, and the rest of the Gulf region. The already precarious state of energy supply in Europe and around the world is being further exacerbated by the war over the Strait of Hormuz, which is a mandatory passage point for global oil and gasoline supplies.
The German government’s expansion plans are taking place in the midst of this energy crisis. According to IEA projections, global electricity consumption by data centers will rise by 15% per year through 2030 – more than four times faster than the growth in total electricity consumption from all other sectors. If all data centers were currently a country, their energy consumption would be as large as Japan’s, the world’s fifth-largest economy.
A recent study by the Öko-Institut commissioned by Greenpeace concludes that data centers will account for a significant share of Germany’s national electricity consumption by 2030. The consequences of this enormous energy demand can be seen in the example of the United States. According to META, the Hyperion requires about 5 gigawatts of energy, which corresponds to the output of five nuclear power plants and results in the generation of 6,000 kg of nuclear waste per year. The study by Greenpeace and the Öko-Institut therefore introduces a new indicator: Radioactive Waste Usage Effectiveness. This is intended to highlight that an increasing number of data centers are set to be powered by nuclear energy.
From the cloud to the factory
And what is all this for? Amid the current capitalist polycrisis, AI data centers seem to be a refuge offering one of the last secure guarantees for capital accumulation. In this context, the cloud – which has always been mistakenly portrayed as a dematerialized cloud – is now being redefined as a factory.
This brings us back to Digital Minister Wildberger presenting the National Data Center Strategy on March 18, 2026. “Data centers are, quite simply put, the factories of the 21st century.” As if he himself weren’t entirely convinced, he adds: “They are really factories.” In doing so, Wildberger immediately provides a subtle justification for why the industrial electricity rate should also apply to data centers – a demand that even the IG Metall union has made.
The ‘factory’ rhetoric is justified by the claim that “data centers process data on a massive scale” and that, as a result, AI is “increasingly being used to control societal processes as well.” Interestingly, AI in the medical field is cited as an example. The fact that medicine is not an industry at all is one thing; another is that specialized – and thus very small – AI models, such as those used in the medical field, do not require large AI data centers at all. Wildberger also fails to mention that his ‘factories’ do not incur any business taxes, nor do they create jobs. Yet he is not alone in using this rhetoric; the EU, too, is now speaking of AI factories that are soon to be established at five locations across Europe.
For historian Nathan Ensmenger, too, the cloud is a factory. In his essay “The Cloud is a Factory,” published in 2021, he traces historical continuities between the U.S. retail and mail-order company Sears Roebuck and platform companies like Amazon. For Ensmenger, this reinterpretation of the cloud as a factory – which was still provocative at the time – is productive for several reasons: First, it makes the information economy and its high socio-ecological costs visible. Second, it illustrates that the cloud, like other factories, relies on physical infrastructure such as roads, bridges, land, and cables. Third, it enables an understanding of the interconnections between different infrastructures and technologies. Ensmenger cites blockchain as an example here; today, this could just as easily be applied to AI. Fourth, it allows us to trace the cloud’s supply chains, which include, among other things, chips, rare earth elements, and chemical coolants. Fifth, it raises the question of which promises of the factory – such as job creation – are not being fulfilled in the context of the cloud. Sixth, it positions the cloud as part of a broader information economy, for example, as the foundation for autonomous driving.
The discourse on the cloud as a factory thus serves two opposing functions. While Ensmenger rightly highlights the socio-ecological costs of the cloud and its industrial base, Wildberger, Reiche, and others – quite successfully, if one looks at the approval ratings – aim to position data centers as a necessary industry for Germany’s economy. However, this ‘factory’ rhetoric effectively pushes the questions of what the additional computing capacity is needed for and what societal priorities it is intended to serve into the background. Should we wait until the circular AI bubble bursts to address them?
Socializing the AI factory
For Karl Marx, factories were the physical arena of the capitalist mode of production. Through increasing automation and the one-sided fragmentation of labor in the 19th century, workers came to be subordinated to capital: “In manufacturing and craftsmanship, the worker uses the tool; in the factory, he serves the machine,” Marx writes.
However, workers are now almost nowhere to be found in today’s AI factories. This makes it all the more relevant that activists in the Rhenish Mining Region, Marseille, Switzerland, Aragón, as well as in Ithaca and Andhra Pradesh, are rising up against the extractivism of AI capital. As the central site from which, according to Antonio Gramsci, hegemony arises, factories can also become sites of alternative organization and counter-hegemony. Anyone who thinks the factory metaphor through to its logical conclusion will arrive at the conclusion that the AI factory must be socialized.