The author wishes to thank the anonymous reviewers as well as the editors of the Environmental Materialities of the Digital issue of RESET for their thoughtful and constructive feedback, which greatly contributed to improving the clarity and depth of this article. He is also grateful to Jérôme Baudry, Nicolas Nova, and Emma Tsuji-Harrison for their insightful comments, which helped deepen and refine the text. This article is dedicated to the memory of Nicolas Nova, who passed away suddenly in December 2024, in tribute to his scientific and human contributions.
1Our algorithmic media landscape is driven by a specific incentive: the miniaturisation of smartness. Smartness is a concept that has been closely examined by academia. It is described as the core element to produce an environment that enables automation of « many human processes and respond in real time to human choices » (Halpern, Mitchell and Geoghegan, 2017: 108). Tech corporations materialise this in their products by trying to further embed computational power inside exponentially efficient and compact objects of our daily lives. Behind smartness, one hardware is critical: the Graphics Processing Unit (GPU). The GPU (fig. 1) is a specialised electronic circuit designed to accelerate the processing of images, videos, and other complex computational processes. Originally built by the US-based corporation Nvidia to process graphics, this hardware became of core importance due to a type of computational architecture it embodies: parallelisation. In comparison to our computers’ central processing units (CPUs) handling commands and information in a linear way, GPUs enable the process of large sets of data in simultaneous – parallel – ways. Today, this hardware architecture has become critical in Artificial Intelligence (AI) and is used in a wide range of applications beyond video gaming practices. Whether integrated into our laptop motherboards or operating as a standalone in desktops (e.g, dedicated GPUs), it is key in materialising the agenda of large tech corporations. Largely powering the data centres of social media corporations — Meta alone is estimated to have 1.3 million GPUs by the end of 2025 (Wiggers, 2025) — as well as ‘smart’ sensors of our Internet of Things (IoT), GPUs and their miniaturisation are the key success factor to corporate desires of smartness.
Figure 1. GPUs.
On the left, a dedicated GPU card designed for desktops or server racks, featuring a chip wired to a heatsink with attached fans for effective heat dissipation. On the right, an « onboard » GPU chip integrated directly into a laptop’s motherboard.
Source and credit: Nvidia.
Drawing from the corporate objective of smartness miniaturisation, my contribution brings together two communities situated on both ends of the GPU’s value. The first community, operating from Taipei (Taiwan), is known as liquid-nitrogen overclockers. Engaging with heat as a by-product of computation, these gamers and hobbyists focus on competitive computer cooling. Pushing this need to the extreme, they use liquid nitrogen (ln2), as a cooling agent and pour on their GPU computer chips. The second community I place overclockers in dialogue with is located in and around Agbogbloshie, Accra (Ghana). It is composed of ‘urban miners’ (Grant and Oteng-Ababio, 2016) who are dealing with the recycling and repurposing of discarded GPUs and other computational devices. Operating in markets, workshops, or larger facilities, they engage in a wide range of activities. These include vending, maintaining, dismantling, scrap-dealing, and casting second-hand GPUs and its materials. In order to disentangle dominant narratives occurring within miniaturisation and its sociotechnical processes, I draw three main sections and lines of inquiry. These sections place in dialogue inquiries from the analysis of overclockers (Taiwan) and urban miners (Ghana).
From Silicon Valley figures such as Nicholas Negroponte prophesizing in early 1995 a digital revolution in which bits would replace atoms (Negroponte, 1995), computational power has long been imagined through a rhetoric of dematerialisation. Today, this continues in the rhetorics of smartness, where intelligence is approached as an algorithmic, autonomous process that decouples agency from human actors (Halpern and Mitchell, 2023). Taking the counterpoint to this dominant narrative through the practices of overclockers and urban miners, I first argue that tracing the geological and thermal entanglements of miniaturisation reveals how modern technology, in fact, relies on the role of communities and their elemental practices.2In addition to tropes of computational immateriality, algorithms are often perceived as universal: capable of operating seamlessly across diverse sociocultural contexts and solving problems autonomously. This notion, rooted in the early ambitions of AI pioneers like Marvin Minsky and John McCarthy to develop Artificial General Intelligence (AGI), continues to shape how corporations frame smart technologies, emphasising adaptability and applicability while downplaying situated constraints. By bridging into dialogue practices of these two communities, I then argue that studying learning techniques brings computation and its hardware to a body of local know-how and « situated knowledge » (Haraway, 1988). Tracing these embodied gestures and ways of sense making reveals, in other words, how computation is deeply bound to the specific sociomaterial context in which communities operate.
The third narrative I tackle through my article has emerged as a techno-utopian approach to technology, grounded in inclusivity and oriented toward collective empowerment. From John Perry Barlow’s Declaration of the Independence of Cyberspace (Barlow, 1996) to the United Nations’s AI for Good initiative, such discourses depict the digital realm as a site of liberation where technological advancement aligns seamlessly with humanitarian ideals. By tracing these elemental and situated practices of GPU optimising and recycling, I argue, finally, that exploitative dynamics arise and structure this incentive. From the privatisation of overclocker’s knowledge to the production of chips deemed exponentially more opaque by urban miners, this brings into focus how the miniaturisation of smartness is not simply about computation and efficiency but is embedded within a broader system of economic and sociotechnical stratification.
3On both sides of our computational culture, this article centres on the GPU and its sociotechnical entanglements. From this perspective, I draw from existing academic contributions undertaken in the field of Science and Technology Studies (STS). These inquiries, focusing on technological hardware – « stuff you can kick », as argued by Lisa Parks (2005) – all proceed with the same aim: zooming in at the level of one key object of our sociotechnical cultures in order to reveal its broader social, cultural and environmental implications. These inquiries range from second-hand computers in Ghanaian internet cafes (Burrell, 2012) to undersea cables (Starosielski, 2015), server farms (Hu, 2016) and their water-cooling systems (Hogan, 2015), semiconductors (Lenny and Markoff, 1985; Gabrys, 2011; Rella, 2023) or antenna trees (Parks, 2009). They also include consumer devices like AirPods (Taffel, 2022), energy storage technologies such as lithium batteries (Friedman-Rudovsky, 2009), as well as specialised hardware required for Bitcoin mining like ASICs devices (Cooper, 2023).
4
In exploring the material side of our technological cultures, my work is also inspired by academic research tackling our computational structures’ « earth layer » (Bratton, 2016). From plastic (Davis, 2015), tin (Ingulstad, M, Andrew, P. & Espen, S, 2014), air (Horn, 2018) or rare metals (Vikström, 2017), these inquiries all tie into the theoretical framework guiding my analysis: elemental media studies (Peters, 2015). In this approach, Peters proposes here to challenge our conventional perception of our sociotechnical systems as purely technological or symbolic.
Moreover, he argues that technology should be understood in the light of its connection with the core elements it is embedded in: earth, water, air or fire. Inside this elemental understanding of technology, two analyses further connect to my fieldwork on overclockers and recyclers communities. The first one is developed by Jussi Parikka. Under the impulsion of his « A geology of media » (2015), this analysis invites us to decenter the temporal and spatial understanding of media away from the digital. By doing so, it proposes to ground technology - and computational power - within the temporal and material reality of rocks, substrates and other minerals from which technology is built upon. Connecting the fieldwork of overclockers and urban miners, I draw on Parikka’s geological view of media as a starting point to explore the infrastructures behind GPU metals: how they're optimised, maintained, and recycled in specific contexts. The second framework I build upon emerges from Nicole Starosielski’s « thermocultures of media » (Starosielski, 2016; 2019; 2022). Further developed in her seminal book « Media: Hot and Cold » (2022), the author invites us to understand computational heat – and thermal management – as sociotechnical: playing a key role in stabilising the day-to-day computational processes of our digital infrastructures. Starting from the fact that computation is, at first, « the work of managing heat » (Brunton, 2015: 159), Starosielski’s concept is essential in order to further reveal how, through tracing « hot and cold » temperatures at play within the lifecycle of the GPU, economic practices and uneven labour dynamics arise.
Bridging both Taiwan’s overclockers and Ghana’s recyclers, I structure my contribution’s claims using three frames of analysis. These frames mirror each section of the contribution in order to shift away tropes of smartness as immaterial, universal and neutral. The first framework is the Chaîne opératoire (Mauss, 1935; Leroi-Gourhan, 1993; Lemonnier, 1992; Coupaye, 2022). Initially developed by the French Anthropologist Marcel Mauss and later defined by his student and French Archeologist André Leroi-Gourhan, the concept argues that:
« A technique is made of both gesture and tool, organized in a chain by a true syntax, which gives operational sequences both their fixity and flexibility » (Leroi-Gourhan, 1993: 114).
5Instead of focusing on finite objects, the framework proposes then to disentangle the circulation and transformation of materials in « organic, temporal, relational and unfolding » (Coupaye, 2022: 4) ways. Similarly to Jenna Burrell or Martin Oteng-Ababio, whose works reveal transformations of materials occurring within the Ghanaian economy of second-hand computers, I use this frame to trace and disentangle labour implications and workers operating within the lifecycle’s transformation of the device. Bridging overclockers in Taiwan and Ghanaian recyclers, tracing the chaîne opératoire of the GPU and its components helps me reveal how our algorithmic culture relies on a set of elemental (Peters, 2015) processes and techniques undertaken by these hidden yet critical communities. Building from these entangled geological and thermocultural human practices, I then focus on the devices’ local knowledge and know-how. In this context, I draw from the work of Marcel Mauss and his seminal « techniques of the body » (Mauss, 1935), focusing on the embodiment of knowledge and
« the ways in which from society to society men know how to use their bodies. » (Mauss, 1973: 70)
6Aiming to challenge cognition and its universal characteristics, this approach helps me contextualise how, both within the context of overclockers and urban miners, techniques of optimisation and recycling are « physio-psycho-sociological » (Mauss, 1973: 85). In other words, they are specific to communities and bound to the cultural contexts where these actors evolve. Third and finally, zooming in at the level of these learning processes enables me to shed light on corporate mechanisms and strategies at play for control and the privatisation of knowledge. From this perspective, this third concluding framework of my contribution expands from Fred Turner’s understanding on prototypes (Turner, 2016). Embedded through design and engineering, the author proposes to approach these unstable objects not only from their technical realm but as apparatuses that materialise specific sociotechnical futures driven by tech corporations while weakening others. With these material embodiments,
« particular socio-technical configurations become available as potential visions of a larger and presumably better way of organizing society as a whole. » (Turner, 2016: 260)
7Through prototypes produced by GPU corporations using freely-shared knowledge acquired from overclockers, specific sociotechnical ideologies of privatisation and close sourcing are materialised. These practices, in turn, affect labour and recycling communities in Ghana.
Table 1. COMPUTEX2023’s organisers, overclockers and GPU enthusiasts.
In addition to a twenty-five-page fieldwork diary, the produced data consists of nine semi-structured interviews that vary between thirteen and seventy-eight minutes. The profiles of the interviewees, referred to by the nicknames they use in the ln2 overclocking community, are the following:
Nickname
|
Profile, location
|
Interview length (approx)
|
BisoBiso
|
Overclocker, South-Korea
|
Thirty-four minutes
|
IvanCuppa
|
Overclocker, Indonesia
|
Thirteen minutes
|
Lucky_n00b
|
Overclocker, Indonesia
|
Twenty-four minutes
|
Mark
|
Field-engineer, Taiwan
|
Twenty minutes
|
Micka
|
Overclocker, China
|
Thirteen minutes
|
MP Customized
|
Case modder and Scratch-builder, Belgium
|
Thirty-five minutes
|
Selby9123
|
Overclocker, The U.S.A.
|
Sixteen minutes
|
ShaggySVK
|
Overclocker and brand ambassador, Slovakia
|
Seventeen minutes
|
Trouffman
|
Ex-overclocker and event organiser, Canada.
|
Seventy-eight minutes
|
Table 2. Agbogbloshie’s second-hand GPU recyclers.
In addition to a twenty-page fieldwork diary, the produced data consists of six semi-structured interviews that vary between eleven and forty-five minutes. The profiles of the interviewees, referred to with letters in order to grant anonymity, are the following
Nickname
|
Profile, location
|
Interview length (approx)
|
A
|
Dismantler, Ghana
|
Twenty-four minutes
|
C
|
Computer and GPU vendor and repairer, Ghana
|
Twenty-one minutes
|
D
|
Scrap-dealer, Ghana
|
Eleven minutes
|
F
|
Computer importer and vendor, Ghana
|
Seven minutes
|
H
|
Computer importer and vendor, Ghana
|
Ten minutes
|
L
|
Computer and GPU vendor, Ghana
|
Twenty-one minutes
|
M
|
Computer and GPU repairer, Ghana
|
Twenty-nine minutes
|
O
|
Smelting factory technician, Ghana
|
Six minutes
|
S
|
GPU (specific) repairer
|
Forty-seven minutes
|
X
|
Computer vendor and repairer, Ghana
|
Twenty-three minutes
|
Z
|
Computer vendor and repairer, Ghana
|
Twenty-nine minutes
|
- 1 A Small Form Factor (SFF) PC is a compact computer designed to deliver full desktop performance. It (...)
8
The elemental infrastructure of miniaturisation becomes visible when one attends major global computer fairs like Computex2023 in Taiwan. Walking across the venue makes it tangible how technical issues and challenges of thermal regulation and control are critical to the miniaturisation of chips and exponentially smaller devices. From computer fans adapted to Small Form Factor (SFF) PC1 to immersion tanks or water-cooling tubes (fig. 2), vendors do not simply sell computers but showcase devices that optimise and manage heat in efficient ways. At many booths, cooling components are presented ornamentally (fig. 3), integrated into the aesthetic design of gaming computers — often enhanced by the extensive use of red, green, and blue (RGB) LED lighting.
Figure 2. Water cooling device used to dissipate heat in a medium server rack.
Source and credit: the author.
Figure 3. Triangle-shaped ornamental structure created using RGB cooling fans.
Source and credit: the author.
At Computex, this thermoculture of media reaches its peak during live demonstrations and competitions of liquid-nitrogen overclockers (fig. 4). As Trouffman (former overclocker and event organiser, Canada) explains, the event serves as a unique annual gathering point for the community. At the event, these gamers are working hand-by-hand with presenters to attract a broader audience composed of tech-hobbyists. With dedicated space in the presenters’ booths, they publicly engage in demos, workshops or tournaments, both hosted on-site and streamed on specialised platforms. Overclockers are often neighboured by another community of tech-enthusiasts: case modders (fig. 5). Although different in their techniques and outcomes, both communities expand from the same observation: that computers « prefer to run cold » (Selby9123, overclocker, the U.S.A). However, while case modders typically rely on air or water cooling to modify gaming machines with both aesthetic and precise engineering, LN2 overclockers push boundaries further by harnessing liquid nitrogen to achieve ultra-low temperatures.Figure 4. liquid-nitrogen overclockers, Computex2023.
Source and credit: the author.
Figure 5. Case mods with integrated water-cooling system.
Source and credit: the author.
9Studying ln2 overclocking and the situated practices of its members reveals how, as argued by Starosielski, computational power
« only operate[s] within specific temperature ranges, failing in conditions that are too hot or too cold » (Starosielski, 2016: 294).
10Defined by the community, the practice consists in creatively ‘tinkering with hardware’ (Lucky_n00b, overclocker, Indonesia), software and liquid nitrogen in order to produce the most efficient computer cooling setup. Benchmarked with the help of energy extensive programs, these builds are then compared. Competing individually or in a team, those who produce the best apparatus wins various computer hardware equipment as well as cash prizes. As a way to optimise their practices, members use custom computer-setups that are case-less and enable them to quickly interact with parts of their machines that need to be cooled down: GPUs, CPUs or memory sticks (fig. 6). This open-hardware design comes with risks of frequently short-circuiting chips and other components when exposing their devices to extreme temperatures. However, it also enables them to quickly intervene on their boards – whether by isolating integrated circuits (ICs) with tape or Vaseline, or by monitoring condensation using torches and hair dryers.
Figure 6. Case-less computer setup with liquid-nitrogen cooling rig.
Source and credit: the author.
11To be eligible for competitions and prizes, overclockers must follow a set of rules. For example, they are required to use GPUs and hardware produced by recognised manufacturers. Another key requirement is the use of specific benchmarking software (fig. 7) with which they can assess their build’s performance. These benchmarks are carefully selected by brand manufacturers and competition organisers. This, in turn, highlights what South Korean overclocker BisoBiso emphasises: competitive LN2 overclocking is never an isolated activity. Instead, it is embedded within a broader infrastructure of testing, tweaking and optimising results (BisoBiso, Overclocker, South Korea). This broader ecosystem of monitoring, research, and development is structured by field engineers (FEs): experts hired by brand manufacturers playing a key role in the industry. Supervising overclockers and their experiments, they act as bridging points between these hobbyists and the corporations they work for. As overclocker Lucky_n00b from Indonesia, therefore, describes, while overclockers « mess around with the hot and cold » of these GPUs, FEs systematically document the thermal and material characteristics and performances emerging from the builds. This monitoring includes, for example, how silicon semiconductors react to specific temperatures or how efficiently GPU fans dissipate heat. Citing Starosielski,
« understanding amalgams of natural and cultural objects through their chemical connections and relations » (Starosielski, 2019: 2)
12make moreover visible how overclockers do not operate in silos but act as key actors providing crucial knowledge to corporations. This knowledge is in turn incorporated into the design of the next generation of GPUs, enabling them to be exponentially smaller and more compact.
Figure 7. GPU benchmarking results from 3DMARK, generated after running a 3D video sequence using the NVIDIA GeForce RTX 3090.
Credits: Techpowerup.
- 2 The term « Burgers » originates from Hamburg, Germany, which hosts a large Ghanaian diaspora. It wa (...)
13Behind design choices of general-purpose computers marketed worldwide or overclocked devices favoured by gaming communities, a sociotechnical phenomenon is at work: planned obsolescence. This business incentive refers here to the deliberate design and production of hardware with a limited lifespan to encourage frequent upgrades. With planned obsolescence, a continuous cycle of consumption is ensured, where users are pushed to replace or upgrade their devices on a regular basis. In countries of the Global North, institutions such as schools, hospitals, and corporations periodically renew their computer systems, whether driven by declining performance, software compatibility, or other desires motivated by newer technological standards. It is here that the business of Ghanaian bi-national workers known as burgers2 starts. Often active in countries such as the U.S.A, UK, U.A.E, Germany, or Canada, they supply Ghana’s markets with second-hand computers and other electronic devices. Such items are bulk-bought and then shipped or carried by planes. The role of these individuals is critical (Burell, 2012). As the ones behind the importation of computers in Ghana, they help nuancing reductive discourses depicting computers as simply ending in Ghana’s graveyards (Yeebo, 2014; Akbar, 2015; Hardy, 2018; Hugo, 2019). They make, in other words, visible the
« broader ecosystem of distribution, repair, and disposal (...) necessary for the material support of Accra's internet cafés. » [as well as the local market] (Burrell, 2012: 161).
- 3 Hotguns is a term used by Ghanaian recyclers to describe heat guns — tools commonly used to melt, l (...)
14The circulation of these devices imported by burgers shifts our understanding of smartness and digital growth away from its tropes of immateriality. In other words, they reveal the infrastructures of maintenance and recycling that are materially situated and specific to metals, substrates or temperatures. When bought by burgers, computers first arrive in the capital’s vendors’ markets and shops such as La Paz or Circle. Often bought untested, importers then collaborate with local technicians (fig. 8) who are able to perform a variety of repairing tasks before putting devices back on the market. As mentioned by M, these fixes are often due to problems which occurred during transportation. With glue, the computer ‘necks’ connecting the laptop’s screens to the motherboard and embedded GPU are restored. With toothbrushes, fans – enabling the dissipation of heat – are cleaned and optimised. When devices arrive « faulty » (M, Computer and GPU repairer, Ghana), hotguns3 are, when possible, used to detach and swap chips, resistors, capacitors, and other integrated circuits (ICs) embedded on boards.
Figure 8. GPU and computer repairs in shops and workshops around La Paz and Circle.
Source and credit: the author.
Rather than buying new parts on the internet, repairers often keep some of the discarded devices they worked on in order to swap pieces. For more specific components, computer chips’ maintenance and recycling are also dependent on another category of workers operating from the neighbouring district and innovation hub (Osseo-Asare and Abbas, 2015; Oteng-Ababio, M., & van der Velden, 2019) of Agbogbloshie. These workers, known as dismantlers, tour across the city and its computer shops to buy motherboards and embedded GPUs from vendors. Due to space restrictions, these « condemned » parts are acquired at a very low price - ranging from 20 to 100 cedis4 (A, Dismantler, Ghana). Following the words of L, repairers and dismantlers also proceed with exchanges. On standalone desktop GPUs as well as laptops in which soldered parts can be extracted, these newly acquired devices are then used for replacement purposes. They serve, for example, to replace missing cooling devices and fans as well as capacitors and other chip elements whose compatibility is confirmed (fig. 9).
Figure 9. Printed circuit board (PCB) of a GPU showing visible integrated circuits (ICs).
15As smart technology becomes more miniaturised, motherboards and GPUs shrink in size, making it impossible to unsolder, repair, or reuse components that are tightly integrated into PCBs. This consequence brings, in turn, the elemental process of miniaturisation to even more geological ground, disentangling the
« chemical, rocky, and metallic reality, which feeds into metal metaphysics and digital devices. » (Parikka, 2015: 24)
16When these objects cannot be used in a computational context anymore, a fifth category of workers comes into play: scrap-dealers. Acquiring parts of computers and GPUs deemed unusable from maintainers, scrap-dealers aim to extract metals they can resell. In this context, another classification of objects and their value applies. While burgers, vendors, and clients are looking for thinner and more compact devices, scrap-dealers are searching for large electronic debris. These debris, including older-generation GPUs with higher amounts of aluminium, steel, copper, or gold in their heatsinks (fig. 10), frames, or PCBs, are classified by weight. In this process, heavier components are considered more desirable.
Figure 10. GPU heatsink made of aluminium and copper.
Source and credit: the author.
17Separating metals from a variety of electronic objects ranging from computer parts to fridges, cars and washing machines, such pieces are then scaled and sold to the last category of workers: casters and smelters. Circling back to the GPU, it is here by tracing the operational sequence of its metals that material and design choices enforcing privatisation strategies unravel. Disentangling in geological and thermal ways, the processing of the card’s metals activates two economic circuits of recycling and transformation, enabling, in turn, specific labour infrastructures and exchanges. One characteristic that differs from these two circuits is the temperature at which their processed metals liquefy. The first is specific to aluminium and copper, melted respectively at 660.3°C or 1,085.0°C. Sold in local markets, it emerges from local independent craftsmen and metal casters. Situated in recycling hubs like Agbogbloshie, these workers run outdoor workshops equipped with improvised furnaces connected to fans that supply air and oxygen for combustion. In such furnaces, a steel rim and a can are used to process aluminium or copper scraps and extract the molten material. Often family-owned and operating within communities of individuals and neighbours, such businesses recycle scraps into various everyday items, such as pots, frying pans, or charcoal ovens, using sand-casting techniques (fig. 11).
Figure 11. Homemade furnace for aluminium casting. Pots made with sand-casting, alongside a charcoal BBQ.
Source and credit: the author.
18Other metals with higher melting points require more advanced infrastructures, tools, and techniques. Among these is steel, which is found in computer and GPU cases. Steel scraps burn at temperatures ranging from 1,085°C to 1,370°C. To recycle steel, scrap dealers rely on larger smelting factories (fig. 12), located about a 30-minute drive away from Agbogbloshie. Such factories, positioned near the industrial harbour of Tema and along Ghana's border with Togo, are owned by Indian corporations. They oversee both on-site operations in Tema and remotely manage the high-level logistics of the corporation. From steel scraps, these corporations produce standardised construction rods (fig. 13) used in the construction of buildings, roads, and more. As O, a smelting factory technician in Ghana, explains, their location is strategically chosen: it provides easy access for scrap dealers by land while also streamlining the processing and export of steel products through sea shipments.
Figure 12. Steel smelting factory located in Tema. Left: sign written in English and Hindi with the facility in the background. Right: interior of the facility where steel is melted.
Source and credit: the author.
Figure 13: Rods produced from materials purchased from scrap dealers (left). Right: machine used to test the quality of the produced rods.
Source and credit: the author.
To conclude this part, I posit that the practices of overclocking and urban mining challenge the dominant narratives of immateriality and autonomy commonly found in the discourse of tech corporations. In other words, I argue that tracing the environmentally-situated practices of these two communities help reveal how communities are critical in optimising, maintaining and recycling the network’s hardware infrastructure, while corporate discourses on technology obfuscate the role of human actors. At the level of overclockers, this is materialised in their core implication in stabilising computational efficiency through their trials and errors. These experiments are then critical to produce the next generation of miniaturised cards whose performances are optimised in exponentially smaller ways. Countering the prevailing tropes of immateriality, examining Ghanaian maintenance and urban mining communities at the other end of the GPU value chain is essential for emphasising the physical entanglements and ecological embeddedness of modern technology and what powers it.
19Shifting from the previous focus on tools and gestures, I now turn to examine how local and situated knowledge, developed within communities, is shaped and often disrupted by the global drive toward miniaturisation, propelled by corporate agendas. This knowledge transfer disentangles when we examine the variety of learning paths and motivations of overclockers, and how these paths all lead towards the standardisation of computational power and GPUs. Overclockers such as Selby9123, ShaggySVK, and BisoBiso for example, based in developed countries like the U.S.A, Slovakia, and South Korea benefit from more traditional learning routes and access to specialised events. For instance, Selby9123, based in California, was introduced to overclocking through electrical engineering studies, supported by a robust network of cutting-edge products, specialised events, talks, and demos organised by tech corporations. In contrast, overclockers from developing countries like Indonesia or China navigate different learning environments to hone their skills. Indonesian overclockers like Lucky_n00b and IvanCuppa, for example, engage with online gaming communities and informal gatherings where hobbyists and amateurs collaborate and experiment (Lucky_n00b, Overclocker, Indonesia). These grassroots overclocking communities operate independently of major corporations and their standards. Instead, they flourish through the enthusiasm of tech-savvy individuals who come together to optimise and hack budget-friendly computers. Micka, an overclocker based in China and selected for the Computex event, shares with me insights that highlight the parallels between the overclocking community and the optimisation of affordable computers. Growing up in rural China, Micka reflects on how his path to overclocking unfolded:
« At that time, I was still a student and had no money to get good gaming machines. So I wanted to buy lower [value] hardware and then, through air cooling, improve the system’s performance. Then, through the years, I got a bit obsessed and started to explore other types of cooling systems. I first started with watercooling, and finally ln2. It was a progression of wanting to always have higher performances through the hardware and its capabilities » (Micka, Overclocker, The Republic of China).
20This contrast between localized ingenuity and corporately supported expertise illustrates how overclocking communities negotiate, adapt to, and sometimes resist the pressures of technological standardisation and miniaturisation. Ultimately, these diverse learning trajectories and techniques reveal how the corporate agenda of GPU computational power and cooling relies upon a body of situated, embodied knowledge initiated by hardware enthusiasts and hobbyists.
21From Selby9123 to ShaggySVK, BisoBiso, Lucky_n00b, IvanCuppa or Micka, relationships with brand companies start at the level of the benchmarking website HWBOT5 (fig. 14). When they upload their results on this platform, it is of core importance in the process of datafication. As explained by Trouffman, HWBOT is a central structuring element of the community (Trouffman, ex-overclocker and event organiser, Canada). In essence, it takes the form of a ranking system that allows members of the community to upload their experiments and benchmark scores, compare them to others, and track progress. Each entry on the website is structured around three levels. The first level contains general data about the experiment and its author. This includes a name or pseudonym, country, and date of submission. Additionally, it documents the type of cooling used, along with a series of hardware and software specifications. These specifications include the model of the card used, the computer’s Operating System (OS), the type of benchmark used, and so on. The second layer — and the important "proof" overclockers need to submit — is a screenshot of their benchmark’s output. The screenshot contains more technical details, including the computer’s clock rate and the computing power gained through the cooling setup. It is within this data that the success of the build is revealed. Finally, submissions must include a picture of the build used to conduct the cooling performance (fig. 15).
Figure 14: Frontpage of HWBOT.
Source and credit: the author.
Figure 15. HWBOT’s required picture uploaded by a user.
Author: AKM. Source: https://hwbot.org/benchmarks/3dmark_-_fire_strike_extreme/submissions/5595009.
22Selected from this platform, where tech corporations monitor the evolution of chips and identify the best overclockers, Computex then serves as an open lab to further test specific devices provided by these corporations. Structured by the work of field engineers (FEs), gaming mechanisms are also introduced to facilitate the production — and acquisition — of data and knowledge. Either individually or in teams, overclockers race against the clock. In this process of knowledge transfer, the entertainment aesthetics specific to the Computex fair also play a crucial role (Auray, 2001). These include, for example, hostesses parading with top-tier devices and engaging with hobbyists, a female crooner cheerfully commenting on demos, as well as smoke machines, free meals, and giveaways. Last but not least, this open-lab infrastructure of datafication and knowledge privatisation is further reinforced by financial rewards that overclockers can obtain, with a maximum prize of $10,000 (fig. 16). Winners are then photographed alongside influential figures in the tech community, such as CEOs and other industry evangelists.
Figure 16: Selby9123 winning the first cash prize of 10,000 USD at COMPUTEX 2023.
Source and credit: the author.
23Whether in Taiwan or Ghana, communities are inevitably affected by global influences. One characteristic structuring the knowledge of urban miners is moreover disrupted: the way they learn from other devices. In this context, Z recalls:
« My brothers were into TVs… So, I learned from observing them. In the end, TVs, computers, and other PCBs [such as GPU] are pretty much the same. They contain the same type of ICs, capacitors, transistors, and so on. They told me the foundations, and I picked up from there » (Z, Computer vendor and repairer, Ghana)
24From there, Z shared with me the importance of repairing cathode-ray tubes (CRTs) and other devices like liquid-crystal display (LCD) televisions to better understand the foundations of electronics, television technology, and other hardware focused on graphics. This modularity amongst devices Z works with on a daily basis is tangible when we visit the repairer’s warehouse, where both computer’s motherboards, screens and TVs (fig. 17) are stored.
Figure 17: Warehouse containing both liquid-crystal display (LCD) televisions and computers.
Source and credit: the author.
When discussing newer GPUs as well as more compact devices Z and his colleagues exponentially find on the market, an observation is shared: smaller cards are harder to maintain. This trend reflects a broader shift in device manufacturing toward increased miniaturisation and integration, which often sacrifices repairability for sleekness and performance. As a result, these devices present significant challenges for workers who rely on precise manual skills to keep hardware functional. While TVs or larger desktop computers are more modular and fixable in their design, the design of newer devices impacts these workers' ability to desolder components, whether transistors, capacitors, ICs, and so on. As shared by Z, this leads, in turn, to damage and the production of more « condemned » devices, which are then sold to scrap-dealers and dismantlers.
25This disruption, driven by exponentially smaller boards, also becomes evident when examining scrap-dealing and dismantling communities' ways of understanding and dealing with these changes. Here, a blatant example of this impact is tangible when we investigate the structure of apprenticeships (fig. 18). Typically carried out by individuals from rural regions of Northern Ghana, apprenticeships span several years. Under the guidance of seasoned professionals, young men acquire essential skills for the job, such as gathering and evaluating scraps and disassembling electronics to extract valuable metals. Along the process, knowledge-transfer between older and newer generations occurs. According to D, who acquired the trade by learning from his former mentor, there is a shift occurring between traditional knowledge and the newer kinds of objects. This shift is driven by the rise of a new type of PCB design, commonly known as 'green' boards. D argues:
« I learned in my younger years several techniques to quickly find and remove copper from PCBs [such as GPUs] or other hardware parts. These boards, we call them original boards. One of these [techniques] consists of using fire to remove the paint, which then shows all the cards’ copper lines. To see how fresh the copper is, we can scratch it with spikes. If the lines turn red, then the quality is proven. Now, new types of boards arrive. We call them green boards. They are less valuable and have less metals. » (D, scrap-dealers, Accra).
26Drawing from the scrap dealer, ‘green’ boards are thus representative of new types of PCBs in which semiconductors are miniaturised. In opposition to older — ‘original’ — boards in which components are more spread out, these boards exponentially arising in and around Agbogbloshie are deemed less valuable, negatively impacting in turn the labour of communities. Circling back to the words of D, this is due to the lack of recycling infrastructures and knowledge, in addition to the fact metals of these objects cannot be extracted easily.
Figure 18: Master (left, wearing a blue soccer jersey) supervising the work of apprentices (right).
Source and credit: the author.
27In this section we have seen that analysing these two communities make tangible how, nuancing tropes of universality and ubiquity, computational power and the process of mininaturisation is crafted by a network of local, embodied and context-driven techniques. At the level of overclockers, this materializes in the various strategies and know-how to optimize, stabilize and extend computer cooling through individual findings as well as the metrics of benchmarks. At the level of urban miners, this is tangible in the way these devices are perceived, fixed and recycled within knowledge bound to the site’s other devices or apprenticeships. Building from this section, I also argue that following these communities reveal how embodied knowledge and tacit know-how is then transferred and disrupted by global trends and strategies put forward by tech corporations. From overclockers, this transfer takes shape with the datafication of performances through field engineers, the monetisation of these practices through cash prizes and the role of HWBOT in helping manufacturers to better assess their devices. Within urban miners, this disruption occurs with the emergence of ‘green’ boards and new types of devices in which the design choices of corporations and the process of miniaturisaiton is revealed.
28The influence of corporations on the overclocking scene is also evident when we examine how dynamics of play between overclockers evolve and how efforts are made to guarantee fairness. On HWBOT, for example, a number of rules are listed. These rules ensure the validity of performances when cash prizes and access to sponsoring is involved. As the guarantee of the ranking system and its integrity, a first set of rules is, for example, connected to benchmarks. On the website, only recognised and listed benchmarks can be used to guarantee consistency and reproducibility of results. For other system information required to assess the validity of the cooling build, a screenshot is mandatory. The screenshot of the benchmark results must display information about the GPU’s name, model, and clock speed, along with key details regarding the BIOS and driver version (fig. 19). Finally, a picture of the setup is required. With the relationship between HWBOT, GPU manufacturers, and the presence of financial rewards, this level of security is, as noted by Trouffman, a crucial measure to prevent cheating on the platform. Here, the overclocker recalls how a technical bug, enabled by a card’s driver, was exploited by members to produce results that were later deemed impossible. Upon discovering the bug, the moderating team decided to nullify all the affected entries.
Figure 19: Details of a GPU’s performance shared on HWBOT (left).
Author: AKM. Source: https://hwbot.org/benchmarks/3dmark_-_fire_strike_extreme/submissions/5595009.
Trouffman also discusses how the integration of overclocking into marketing and capitalist frameworks has led to the circulation of knowledge among overclockers becoming increasingly scarce and difficult. He argues:
« There’s some contexts where overclockers share their techniques, but overall, when there’s money [involved], techniques are kept privately. And increasingly more. They’ll tell you: I spent so much time figuring this out, I’m not sharing…The truth is that it is sad as it really has an impact on the overall community. » (Trouffman, ex-overclocker and event organiser, Canada)
29This highlights how, nuancing the inclusive atmosphere of earlier meetups conducted by the community, the evolution of the practice fitting into more commercialised bubbles changed the way overclockers access and share knowledge. These cooling and thermal control techniques, which were once openly shared within the community, are now increasingly shaped by market forces, making it harder for enthusiasts to exchange information freely.
30While browsing on HWBOT, Trouffman shares with me an insight about WindowsXP:
- 6 The term differs here from how it is used in the industry and further reveals the different aims be (...)
« For a time, we had drivers that worked with Windows XP, delivering excellent results – highly responsive to both the hardware and the LN2. Overclockers liked Windows XP so much that even with newer devices, many continued using it because of its performance6. But for me, at some point, this old OS had to go. Of course, this meant slightly less responsive results, but we had to move on and work with current technologies, not cling to outdated ones ». (Trouffman, ex-overclocker and event organiser, Canada
31This comment indicates that, within the overclocking scene, the independent nature of the practice also reacts to digital progress and planned obsolescence. Beyond this resistance of overclockers to the updating of their devices, the impact of hardware stabilisation and standardisation on enjoyment unravels. Parallel to the case of WindowsXP, the same process is observed by overclockers every time they attend Computex. Here, Lucky_n00b argues:
« The more we come to Taipei (COMPUTEX) and the more we work with brand manufacturers, the more performances are added in the next generation of cards. Which really is a rewarding process. However, this creates generations of cards that are very hard to hack. They manage heat and cooling in great ways, but are less fun to play with. So although it is good for the vast majority of gamers, it becomes less interesting for the overclocking scene and community » (Lucky_n00b, Overclocker, Indonesia).
32From HWBOT, WindowsXP and standardisation, this reveals how elemental failures and successes of overclockers serve the corporate incentive of smartness and miniaturisation. Through the words of Lucky_n00b, what is moreover revealed is that this process occurs at the cost of the overclocking community, making these cards less hackable and less fun to play with.
33When devices of our Global North countries — stabilised by overclockers — are discarded and bought by burgers, they are inspected by a plethora of repairers and maintainers. With the effects of miniaturisation in which Taipei’s hobbyists play a crucial role, workers of Circle (Accra, Ghana) such as X observe a difference within the laptops and desktops they process:
« Through time, machines become smaller and lighter. That’s what the generation [of customers] is taking us to. Before…you see this thing? This big thing [points towards an old laptop]. If you keep this in your bag, you’re going to sink for the next two days. Now we are moving to something like this flat [notebook computer]. So not big and heavy, but light machines. Lighter and lighter…but also harder to repair. » [X, Computer vendor and repairer, Ghana].
34Drawing from X, it is at the level of lighter machines that struggle to repair in and around Agbogbloshie disentangle. Although preferable for burgers as it allows them to bring more devices during their travels, the miniaturisation of these GPU chips and computers negatively impacts the labour of repairers. While, in other words, newer generations of computers are more in demand with customers, their shrunk design is exponentially harder to repair and often results in damaged chips and PCBs. This is first due to a lack of precision tools used to unsolder and swap items on the board. Another cause of this struggle is the opaque and over-specialized nature of these systems that do not correspond to previous electronics they learned from. Struggles also manifest in health and productivity issues. In this context, X shares with me the recurrence of headaches and ocular migraines caused by the tininess of specification codes written on GPUs and other motherboard chips (fig. 20).
Figure 20. X showing the tiny specification letters written on the GPU and ICs.
Source and credit: the author.
35Due to the lack of repairing devices and the way boards are designed, a number of chips and hardware received by maintainers are labelled as « condemned ». Such objects then enter into another circuit of recycling operated by dismantlers, scrap-dealers and smelters. In these circuits of elemental transformation, miniaturisation sits at the core of new political economies and community struggles. As mentioned by workers operating around Agbogbloshie, such as D, struggles emerge then with the increasing presence of « green boards » : PCBs with less valuable metals. Deemed of less interest than dedicated GPUs of desktops, the arrival of such objects on site is a consequence of the Internet of Things (IoT) and other shrunken devices of our sociotechnical landscape. As mentioned by D, these ‘green boards’ are negatively impacting Ghanaian recycling communities in two overlapping ways. The first way, due to the condensed nature of these chips, connects to the difficulty dismantlers and scrap-dealers face in separating and processing metals. In comparison with the « original boards » in which aluminium, copper or gold components are recognised in easier ways, it is harder for the community to divide metals of green boards. Circling back to the incentive of smartness, this is due to the fact boards are optimised for condensation, with their components soldered and shrunk together in ways that make it harder to be removed. The second struggle experienced by communities emerges from the processing of steel. As shared by D, steel is both in « original » and « green » boards recycling one of the easier metals to remove, as it is used for external framing and therefore does not mix with the PCB’s soldered components. The metal melts however at higher temperatures, and therefore cannot be processed by individual actors operating with homemade furnaces in which aluminium or copper. As a result, dismantlers and scrap-dealers rely on labour exchanges with Indian-based corporations, operating complex industrial equipment near Tema and its harbour. These large-scale factories, detaining the monopoly over the recycling of steel, impact communities and their well-being by controlling prices and the types of accepted scraps.
Placing opposing contexts in dialogue as done in this paper reveals exploitative dynamics that structure miniaturisation and its incentive. These dynamics emerge from the control of tech corporations on smaller, independent actors in which Taipei’s overclockers and Accra’s recyclers belong. From the privatisation of overclocker’s knowledge to the production of chips deemed exponentially more opaque for urban miners, this brings into focus how the miniaturisation of smartness is not simply about computation and efficiency, but is embedded within a broader system of economic and sociotechnical stratification. At the beginning of the chip’s lifecycle, this disruption is observable in overclockers’ platforms such as HWBOT. After the process of knowledge privatisation by tech corporations is completed, asymmetrical dynamics are further made visible, as the blackboxing (Latour, 1999) of these chips impact the hackability and the freedom of exploration.
In the context of Agbogbloshie and its communities, this materializes first at the level of struggles faced by maintainers with chips whose components are exponentially shrunk. The community’s dependence on inviolable market dynamics and actors results in unequal relations in which they have no choice but to adapt.
36The miniaturisation of smartness, framed as an inevitable technological progression, is, in fact, a deeply political and materially entrenched process. By tracing the GPU’s lifecycle from the extreme cooling practices by liquid-nitrogen overclockers in Computex (Taipei) to the labour-intensive maintaining, dismantling, and repurposing of the device in Agbogbloshie (Accra), I show how the promise of ever-smaller, more efficient and seamless AI systems rests on extractive infrastructures, specialised knowledge, and uneven distributions of power. Through the analysis of two communities, miniaturisation emerges, in other words, as a contested terrain where corporate desires for efficiency, automation, and control shape and exploit both material and intellectual resources from smaller and independent actors. Across the three sections of my article, these power dynamics and industrial forces mirror from overclockers to urban miners despite their different positions in the computer chip’s value chain. The first community, through their role in the thermal capabilities and optimization of semiconductor performance, observes how their knowledge is increasingly privatized and commodified by hardware manufacturers. As a result, these devices become progressively harder to hack and overclock. The latter, navigating through exponentially smaller and more compact cards, face disrupting challenges at the level of their repairing and urban mining practices. Ultimately, what appears to be an engineering challenge is revealed as a sociomaterial issue deeply layered and intertwined with environment and labour. In foregrounding the practices of overclockers and urban miners, this article calls for a critical reassessment of smartness and its miniaturisation that accounts for its hidden dependencies and structural inequalities.