1 Watts (2013). Read more (about) speculative fabulation in Haraway (2016) and Watts (2023).

2 The research was conducted together with Leman Celik, Laura Kocksch and Stefan Laser. I owe deep thanks to all our interlocutors. I also thank Laura Kocksch, Mace Ojala and Ruth Dorothea Eggel, Stefan Laser and Hollis A. Brown for invaluable comments on an earlier version of this text. The research is part of the Collaborative Research Centre “Virtual Lifeworlds” and funded by the German Research Foundation.

3 UPS: Uninterruptible Power Supply ensures continuous power provision, e.g. through battery packs that hold backup electric power to supply the data centre when the mains power fails.

4 Data centres’ future energy consumption is difficult to predict, as it depends on many factors, such as the development of data use and data processing demands, which are estimated to explode with AI, and on the energy efficiency of new generations of CPUs/GPUs (graphic processing units do parallel processing and increasingly take over tasks from central processing units, which are slower due to their serial processing), cooling systems, etc. In a review article, Hintemann & Hinterholzer (2023) report that estimates of data centre energy consumption in Germany will be between 16 and 34 billion kWh per year by 2030. According to the Ireland Capacity Outlook 2022–2031 (EirGrid & Soni 2022), 28% of the country’s total electricity demand is expected to come from data centres and other new large energy users. Energistyrelsen (2023: 42) forecasts that in 2031 data centres will be responsible for almost 30% of Denmark’s total energy consumption. Pasek (2023) points to estimates of global data centre energy use varying from 8 to 51% in the next decade, and she also emphasises that such numbers probably have more agency for backing up industry’s and governments’ credibility than they have truth value.

5 Bundesregierung (2023).

6 For more on the figure of the distant judging observer, see Sørensen (2009).

7 One of my interlocutors forecasted that in the future, there won’t be many mid-size data centres anymore, but instead many large hyperscale and many small so-called edge data centres, which are closer to the users and end-user devices. The forecast is plausible: When it comes to world-wide social media platforms, e-commerce, cloud computing, digital streaming and other wide-reaching digital services (including AI), hyperscale data centres are most economical, efficient, controllable and secure (from the provider’s point of view). As such, services are likely to increase, as is the number and size of hyperscale data centres. In many cases, it makes sense to bring computation and data storage closer to the sources of data use – through so-called edge computing or edge data centres – to improve latency and save bandwidth.

8 Niewöhner (2014; 2016) coined the notion of co-laboration for STS research and thereby parts from notions such as ‘research subjects’ or ‘informants’. In data centre practices, terms such as ‘customer’ and ‘user’ freeze problematic relations with server operators.

9 For further discussions of server democracy, see Maak (2022).

10 See Haraway (2016).

11 In Sørensen (2009) I developed the difference between witness as a distant mode of knowing and withness as an engaged mode.

12 ‘Thinking with’ is mainly a social act in Puig de la Bellacasa (2018), but it can be extended to servers.

13 For more about this chemist’s data practices, see Sørensen & Kocksch (2020).

14 A key inspiration for thinking through engagement with hardware is Mauro-Flude & Akama (2022).

15 See Puig de la Bellacasa (2017: 81).