Governance by history

From Technology Governance Wiki
Jump to navigation Jump to search


contributions

Civilian AI at the Front - Robert Florijn

links things that popped up during class


Workplace surveillance

Information Control and Workplace Surveillance - by Jihae Park and Mariona Fortuny Jané

The Current Governance Issue

There are numerous current governance problems arising from technological innovations, yet here we will specifically focus on the information control and workplace surveillance. Whether intended or unintended, there were certain historical incidents which resulted in social and political consequences. As noticed from the Control Revolution, the dramatic evolution in mass media and telecommunications technologies has led to the reinforcement of the economic and political control during the Industrial Revolution. (Beniger, 1986) By applying this logic to a modern example, this is clearly how many corporations are getting in control of their employees through surveillance technologies. Along with the emergence of remote and hybrid work after Covid-19, employers are now relying on “bossware” to analyze the emails, chats, and messages of their employees on platforms such as Teams or Slack. (Ollier-Malaterre et al., 2026, p.237) Some companies have strict restrictions on transferring data both internally and externally in protection of confidential information and consistently monitor their workers to the certain extent which could slow down their efficiency. This is problematic as it’s not only violating the privacy of their employees but also depriving employees of rights they should be naturally be able to exercise.

Historical Lineage

In relation to the current issue of the information control and workplace surveillance, Max Weber’s argument of bureaucracy and rationalization in the middle and late nineteenth century are good historical examples which are still frequently visible nowadays. Weber stated how the range of control could extend not only by expanding the capability to process information but also by limiting the amount of information to be processed. Here, the former is directly related to the control implemented by bureaucratization (in the mid 19th century) and computerization (in the late 20th century), whereas the latter addresses rationalization. An example from bureaucracy is the expansion of paperwork in response to the growth in information. We can still note these traces of bureaucracy from numerous procedures that we have to go through for a decision-making to take place in corporations. He brings up the example of standardized time zones for rationalization which is simply an example for an omission or a breakdown of information to facilitate its processing. (Beniger, 1986, p.15-16) This is also commonly seen from nowadays as some companies may use internal terminologies for the sake of both confidentiality and efficiency.

==== Mechanism of Path Dependence ====

Following along the line with the historical lineage, there are certain mechanisms at work of path dependence for information control and workplace surveillance. Bureaucracy is closely related with organizational routines. Many companies have already settled down with a fixed workflow and decision-making process, which makes it more costly for one to change from its conventional way of working. This means switching costs from an old system to a new one is clearly high considering its time and capital, not necessarily will guarantee increasing returns. Hence, many firms end up maintaining its old strategy, reluctant to make any modifications or changes.

Politics and Distribution

As businesses grew too large for managers to oversee everything directly, companies replaced human judgment with formal rule-systems, as we see explained by Beniger's 'intellectual technology' (Beniger, 1986, pp. 17, 21). Algorithmic management continues this: a call once made by a supervisor's judgment is now a rule built into software that runs without one, extending a logic Beniger traces as far back as to Taylorism and the so called 'assembly line'. Winner (1980) asks whether such systems must concentrate power this way or were driven there, when other designs were equally possible; nothing required shutting workers out of the decision, so leaving them out was a political choice, and not necessarily a technical necessity. Barbrook and Cameron (1996) explain the reasoning why some can defend this 'domination': they show how 1990s tech culture sold new technologies as freedom even as it tightened control, so the rhetoric could become one of closer oversight as flexibility and autonomy. Together, authority accrues to employers, who in some systems can set and revise these oversight rules unilaterally and face less resistance because the arrangement is framed as freedom; workers bear the cost, losing the informal give-and-take they had with a human supervisor for oversight that is harder to see, question, or contest, while being told it grants them more freedom, not less.

Contingency and Alternatives

A plausible alternative happened on U.S. post WWII: in some workplaces, labor and its pacing was demanded to be set through collective bargaining and enforced jointly, by both management and union stewards monitoring compliance on the line (Breznitz et al., 2025, p. 211). So, in this context, workplace surveillance could have developed as a bargained, dual-administered practice rather than a management monopoly. The 1950 Reuther's Treaty of Detroit cut that possibility off: unions won guaranteed wages and benefits in exchange for surrendering any general claim over how work was supervised (Breznitz et al., 2025, pp. 210–211). Workers depended on wages to survive, and employers controlled the wage relationship. That asymmetry meant employers, given their position of power, were able to arrange things so that, at the very least, workers would have far less capacity to sustain a fight over control itself. Thus, the "trade" wasn't a fair exchange between equals. This reflects which side could afford to hold out on which issue. Hypothetically speaking, had that joint administration model prevailed instead, this could have greatly impacted norms internationally and good practices with worker-union agreed surveillance would have become a widely accepted standard practice.

Governance Implication

An interesting proposition for present-day governance that wants to fill a gap in workplace surveillance is what Kayas, Ong and Belal (2025) are proposing: They argue that EU law, through the GDPR and its required data protection impact assessments (DPIAs), already offers meaningful protection for workers in terms of lawful high-risk data collection and data processing. Nevertheless, these DPIAs are built around data and privacy, so they leave out several dimensions of monitoring itself: its purpose, how intrusive it is, how visible it is to the people being watched, and its effects on workers' autonomy, trust, and standing relative to management. These authors propose a "surveillance impact assessment" scoring a system against these characteristics directly, before any surveillance system is put in place. Although this mechanism "is constrained by its reliance on existing literature and the lack of empirical validation," it feels like a genuine alternative for two reasons: 1. it targets the actual mechanism at issue, and 2. at the same time it builds upon a mechanism of active participatory approach where employees have a direct say in how surveillance is designed (Kayas et al., 2025, p. 11), which, combined with the actual assessment, would put workers' own judgment at the centre of the one decision that affects them. It's important to note that this alternative is proposed as worth piloting, not as a proven fix, and it is fair to ask whether another mandatory assessment stacked on top of DPIAs would function as intended or simply read to workers as more compliance paperwork.

References


Barbrook, R., & Cameron, A. (1996). The Californian Ideology. Science as Culture, 6(1), 44–72. https://doi.org/10.1080/09505439609526455
Beniger, J. R. (1986). The Control Revolution: Technological and Economic Origins of the Information Society. Harvard University Press. https://archive.org/details/beniger-james-the-control-revolution 
Breznitz, D., Levy, K., Lipartito, K., & Zehavi, A. (2025). An equity-focused research agenda for workplace surveillance. Industrial and Corporate Change, 35(1), 206–221. https://doi.org/10.1093/icc/dtaf038
Kayas, O. G., Ong, C. E., & Belal, H. M. (2025). From humans to algorithms: A sociotechnical framework of workplace surveillance. Digital Business, 5, 100120. https://doi.org/10.1016/j.digbus.2025.100120
Ollier-Malaterre, A., Parent-Rocheleau, X., Benkrid, K., & Provost Savard, Y. (2026). Making Sense of Workplace Surveillance: How Workers and Union Representatives Experience, Imagine, and Resist Surveillance Exposure. Surveillance & Society, 24(3), 236–254. https://ojs.library.queensu.ca/index.php/surveillance-and-society/article/view/19100/13263
Winner, L. (1980). Do Artifacts Have Politics? Daedalus, 109(1), 121–136. https://www.jstor.org/stable/20024652


Algorithmic management and workplace surveillance - by Sebastián Muller and Kate Davey''' '''

1. The current issue.


Algorithmic management and workplace surveillance can be defined as the use of computer-programmed procedures to coordinate labor input in an organization (1). Simply put, it is a way to control workers, and workers generally have no say in how the system that tracks them is created. As Winner says "different people are differently situated and possess unequal degrees of power as well as unequal levels of awareness" (2).

The population affected is workers; both “white collar” and “blue collar” workers are impacted. Whether it’s Amazon tracking warehouse workers’ “Time Off Task” with handheld scanners (3) or Meta employees having their every click tracked (4) to train AI, employers have the power to watch and control their employees in an unprecedented way.

It matters legally because workers have very few rights to fight back against these surveillance programs; even the The Electronic Communications Privacy Act of 1986 allows exceptions to surveillance of employees (5).

===
2. Historical lineage. ===


2.1 Panopticon: it is an architectural apparatus developed by Jeremy Bentham in 1791. The design of this building effectively generated the sense of constant monitoring, which in turn motivated the prisoners to behave as if they were being watched at all times. As human labor became commodified, capitalists began implementing measures to ensure the maximization of productivity, including the replication of this design (6). While unused today, the panopticon is widely considered as both a theoretical and practical inspiration of other forms of surveillance in the workplace.


2.2 Scientific management: this theory attempts to analyze and standardize workflows to improve labor productivity. Frederick Winslow Taylor, its main pioneer, used early forms of data analysis during the XX century to decrease walking distances, eliminating unnecessary movements and preventing deliberate slow-walking among workers. (7). There are many modern forms of data-driven decision-making processes in the workforce. For instance, Amazon warehouses track individual workers' scan rates in real time through handheld scanners, generating data feeds to track and evaluate performance (8).


2.3 PC in the workplace: John Blankenbaker designed the first commercially available personal computer in 1971 (9), however, it wasn’t until the mid-1980s that the use of personal computers in the workplace became an industry standard. This achievement gave rise to the digital monitoring of work, which in turn allowed employers to monitor employees in real-time, effectively replacing the human eye with a more efficient surveillance mechanism. This represents one of the key aspects of modern-day workplace monitoring.


3. Mechanism of path dependence.


When analyzing the historic lineage of work surveillance, it is possible to point out two main path dependencies mechanisms in place.

Firstly, it is clear that many of the technological developments that arose through these events have one major goal, which is to increase wealth within a hierarchical ecosystem. Employers resort to several mechanisms to maximize profits and cut inefficiencies at the expense of workers. 

In addition, standardization can also be singled out as a key mechanism at play. Once one corporation implements any of these developments (such as the digital monitoring of the workplace), the market incentives would force the rest to do the same in order to be profitable and remain competitive.

4. Politics and distribution.


One simple answer to this question would be that capitalists implement and leverage these mechanisms to increase their wealth while the working class remains powerless and unable to withdraw itself from the system in force. 

However, if we resort to Winner's framework, we see that the technical elements of these technologies are not intrinsically capitalist, since workplace surveillance can also be observed in alternative regimes (including chattel slavery and the Soviet regime). It seems that, rather than being exclusively tied to capitalistic values, the methods in question respond to strong hierarchical structures.

In addition, even under a libertarian framework, structural inequalities are still present, allowing a privileged class to enjoy individual freedoms while segregated groups keep performing low-income jobs (10), arguably still subject to surveillance. Thus, not only are these technologies compatible with hierarchical top-down forms of government, but also with structures that allow such hierarchies to form without state intervention.

That said, this form of management seems less compatible with democratic societies, which tend not only to value individual liberties but to turn them into enforceable rights (including workers' privacy, health and integrity), thus establishing boundaries for workplace surveillance. This could explain why studies of American business leaders have found them unhappy with democratic scruples (11).


=== 5. Contingency and alternatives. ===


Bush writes about the memex, a machine that could extend a person’s memory and that the memory would belong to the person who owned it. As Bush expands, “For mature thought there is no mechanical substitute" (12). His idea was that a machine could take over the repetitive nature of someone’s job, while the person could record their own thoughts for their own use. Bush’s memex was never created, but the idea inspired products like Microsoft’s Productivity Score that were made for employers (the buyers) not for the person whose memory it used (the worker).

This immediately brings to mind Meta's Model Capability Initiative, which was done to train AI agents the company hopes to sell to other employers as well as watch their employees (13). If this was Bush’s memex, it would belong to the employee using it, not Meta. 

6. Governance implication.

Democratic oversight of the use of algorithms as a form of surveillance is essential to ensure that workers’ rights are not overstepped in the name of efficiency. As such we propose: all workers should have a legal right to be consulted and approve a surveillance program before a company installs and uses it. A tracking system should be a group decision, not an inevitable, unilateral one. While the United States has very little protections for workers in terms of surveillance, the EU leads the way with the GDPR, specifically Article 22 (14), and the AI Act Article 5 (15). It’s also worth noting that in December 2026, the Platform Work Directive (16) will come into effect which will ban automated termination.

Platform Dependence in Public Communication (News) Ceren Isiklar & Rodrigue Kayembe

Centralization of AI Compute and Data

Centralization of AI Compute and Data - by Sisi Hammoudi & Juliette Chamot

Governance issue

Today, several path-dependence phenomena are emerging across advanced computing technologies. This assignment addresses one of them: the centralisation of AI compute and data. Training and running frontier AI systems now depends on a handful of hyperscale providers, namely Amazon, Microsoft, Google, and Nvidia, which dominates the GPU market. In other words, near monopolies exist throughout the supply chain. This concentration creates a governance problem affecting public and private actors alike, both legally and institutionally. On the institutional side, it prevents public administrations, and states other than the US or China, from reaching sovereign AI capacity. It also raises competition-law concerns, since dominant market positions can prevent the market from delivering the best products at the best prices to consumers. Finally, AI developers, like states, are often forced into contracts with hyperscale providers, creating a structural dependence on them.

Historical lineage

1945

Wartime scientific instruments should have been redirected to extend human command over knowledge, helping establish the postwar model of sustained federal funding for computing research (Vannevar Bush's, "As We May Think », 1945).

1990s

The privatisation of the NSFNET established the idea that critical digital infrastructure, even though publicly subsidised, could legitimately be privately owned. The EU Commission followed the US's lead, recommending privatisation and deregulation of telecommunications and media as the best way to build the "infobahn" though EU members could instead have chosen nationalisation of the infrastructure (Barbrook & Cameron, "The Californian Ideology," 1996).

2006

Hyperscale cloud computing turned massive fixed capital investment in data centresinto a rentable service, favouring firms able to invest at scale.
Same period: Nvidia's proprietary CUDA software, released alongside its GPUs, became the de facto standard for machine learning, locking in a decade of tooling and expertise around one firm's hardware.

From 2022

Semiconductor export controls and the CHIPS Act turned compute into an explicit instrument of geopolitical strategy.

Mechanisms of path dependance

Several mechanisms drive this centralisation, but path-dependence ones dominate. Because AI compute is a capital-intensive industry, new entrants struggle to compete with hyperscalers that have already sunk huge sums into infrastructure, a barrier that produces monopoly-like situations. This dynamic, known as increasing returns, generates system-scale economies that harden into a near-irreversibility of investment (David, "Clio and the Economics of QWERTY," 1985). It overlaps with a second mechanism, standardisation, itself a form of technical interrelatedness: researchers' investment in Nvidia's CUDA-based tooling locks the field into that ecosystem regardless of whether better alternatives exist. A further mechanism is legal and contractual entrenchment. For example, Microsoft's stake in OpenAI binds developers to a specific provider by agreement rather than technical necessity.=== Politics and distribution ===

Technology does not have a history of distributing power equally. Rather, its development has strengthened the interests and authority of powerful groups that already possessed economic and political resources, and it is yet to be determined whether AI compute and data will follow suit.

Author Langdon Winner illustrates the historical constraints of technological development with the case of New York from the 1920s to the 1970s. He argued that the city’s roads, bridges, and parks’ physical arrangement contained “explicit and implicit political purpose.” These infrastructures privileged wealthy and white communities while restricting poor and Black populations’ choice of mobility. Similarly, the development of industrial machinery increased productivity, simultaneously weakening union and non-union workers’ bargaining power.

In the twentieth century, following the Industrial Revolution, a Control Revolution emerged from the increasing need to manage expanding production, transportation, and information flows requiring strengthened organizational bureaucratic control. This was heavily shaped by the political context of post-World War II warfare.

Therefore, the bearer of the initial cost of technology is the person putting in the capital to erect it, but ultimately, the person bearing the long-term cost of technology, whose choices diminish, is not the rich and privileged, but the everyday person subject to technologies. Forced to adapt, to reshape their professional skills to adapt to an environment trying to make them obsolete.

AI compute and data are at risk of creating the same effect. In their critique of dotcom neoliberalism, authors Barbrook and Cameron exposed a similar socio-economic tension. The Californian Ideology claims digital technology empowers individuals; however, perpetual unequal access to it creates an apartheid between “information-rich” and “information-poor” populations.

Finally, the development of AI compute and data depends on expensive and scarce computing infrastructure and a highly concentrated semiconductor supply chain. Today, this is one of the most concentrated industries, with single companies having near-total monopolies over the production of chips necessary for the functioning of AI. This monopoly model threatens to reinforce the existing economic hierarchy.

=== Contingency and alternatives ===

If a different technological trajectory was plausible, the most apt time for it to have occurred would be during the transition from the Industrial Revolution to the “Control Revolution”. When industrial production accelerated, our society experienced a “crisis of control”, creating opportunities to organize information and communication differently.

Instead of allowing control technologies to be monopolized within businesses and states, our societies could have pursued a more democratic and decentralized system. Barbrook and Cameron provide such an alternative path through their idea of the “electronic agora,” where digital networks could encourage community participation and direct democracy, where government could actively engage with the people to shape digital infrastructure and expand access. These alternatives did not actually prevail because corporate and state market participants possessed greater resources and authority to shape technological development at the end of the twentieth century.

Inherently, technological development is not predetermined; it can support either authoritarian or democratic arrangements. Today, monopolies control chip production needed for AI compute and data, but other technologies like blockchain permit democratic participation.

=== Governance implication ===

History always threatens to repeat itself. As Barbrook discussed, digital societies are shaped by the political context in which they develop. It is a valid concern, then, that the development of information technology, through AI computing, would follow the same trajectory- that is, be constrained to benefit the wealthy who funded its development, and ultimately harm the middle and lower classes enabling its functioning.

However, AI’s place in our society has not yet crystallized. Technology governance does not have to adhere to this system. Such a seemingly sustainable approach to AI compute and data governance has already been broached. In 2022, the Notre Dame Journal on Emerging Technologies published an article in its issue on Artificial Intelligence where the authors proposed an AI compute system based on a “dual investment strategy”. Firstly, the model would be launched by “subsidizing and negotiating” cloud computing for human AI researchers through “existing” cloud vendors, which would expand on existing AI compute initiatives. Secondly, the compute system would be specifically invested in as a publicly owned infrastructure that would be built and operate similarly to national libraries which own “academic supercomputing facilities”.

The pilot program would have a “base level of compute access” which would keep costs of processing down and enable it to meet the compute needs of a majority of academic researchers. Absent any strengthened legislation on the use of AI compute and data, the hardware would be accompanied by an AI ethics framework to enable sustainable technological governance. In addition to an ethical staff, the human resources needed to keep an AI compute model running would include data officers, scientists, data center and software engineers, research developers, and system administrators. These positions require specialized training, allowing new types of professionals, from any background, to have access to professional prosperity in the field of AI.

By implementing these frameworks, never neglecting the human element, and the economic and intellectual impact AI has on our society, we move from a historical determinist perspective to one acknowledging historical constraints and learning from them.

Bibliography

[1] Langdon Winner, Do Artifacts Have Politics, Daedalus, Winter, 1980, Vol. 109, No. 1, Modern Technology: Problem or Opportunity? pp. 121-136, 124, (1980)

[2] Langdon Winner, Do Artifacts Have Politics, Daedalus, Winter, 1980, Vol. 109, No. 1, Modern Technology: Problem or Opportunity? pp. 121-136, 124, (1980)

[3] James R. Beniger, The Control Revolution: Technological and Economic Origins of the Information Society, Harvard University Press, p. 443, 450 (1986)

[4] Richard Barbrook, Andy Cameron, The Californian Ideology, Science as Culture, 6(1), 44–72, 61 (1996)

[5] Anton Korinek, Jai Vipra, Concentrating Intelligence: Scaling and Market Structure in Artificial Intelligence, National Bureau of Economic Research, 10.3386/w33139, p. 11, 13-15 (2024)

[6] James R. Beniger, The Control Revolution: Technological and Economic Origins of the Information Society, Harvard University Press, p. 443, 442 (1986)

[7] Richard Barbrook, Andy Cameron, The Californian Ideology, Science as Culture, 6(1), 44–72 (1996)

[8] Langdon Winner, Do Artifacts Have Politics, Daedalus, Winter, 1980, Vol. 109, No. 1, Modern Technology: Problem or Opportunity? pp. 121-136, 130 (1980)

[9] Daniel E. Ho, Jennifer King, Russell C. Wald & Christopher Wan, The Centrality of Data and Compute for AI Innovation: A Blueprint for the National Research Cloud, 3 Notre Dame J. on Emerging Tech. 71, 81(2022)

[10] Daniel E. Ho, Jennifer King, Russell C. Wald & Christopher Wan, The Centrality of Data and Compute for AI Innovation: A Blueprint for the National Research Cloud, 3 Notre Dame J. on Emerging Tech. 71, 82 (2022)

Centralization of AI compute and Data - by Atmika Dechamma & Pablo Latorre

Governance Issue

AI computing capacity and data presents the governance problem of an extreme concentration of technological infrastructure and information in a restricted number of companies that privately and globally dominate the market. This affects public and private actors - such as small and medium enterprises, potential new competitors, and Global South countries - that increasingly depend on AI systems outside their control and/or are deterred from competing with established companies because of impenetrable entry barriers - most notably initial investments and operational costs (1). The institutional and legal relevance of this is that it consolidates, reproduces and amplifies asymmetries of power in favor of the “Big Tech”, reverting the technological trend from digital decentralization to physical infrastructure dependency and curbing innovation and market efficiency.

Historical Lineage

==== Limitation of Machine Learning ====

The AI revolution was built on the limitation of human capabilities to process large information and derive efficient results - where before there existed only facilitative tools to allow for human-enabled processes, AI now offered derivations and cognitive reasoning in moments and at large scales.


==== Revert to Physical Infrastructure Dependency in the rush to be the best ====

The present AI model heavily relies on hardware in order to perform complex data processing for deep learning. The data centre models followed require larger model sizes (which require higher resource, information and capital investment) to be able to yield better results. Due to the limitation of resources and a rush to establish the best working model of an AI agent, data centres are built on a rudimentary model heavily dependent on massive amounts of electricity and fresh water (2).


==== Regulation and Decentralization of Information ====

Regulation and influence of sovereigns and institutions has resulted in data sharing and information exchange structurally through open source and cloud computing. But, while AI models claims their selling point to be accessibility for all, their information and derivations remain in a closed server, stored for the purpose of training their AI agents to be the best (3). Big AI contenders attract higher consumer and information flows overall resulting in a better developed AI model, proving disadvantageous in any market where early contenders have an advantage and accumulation of power.

=== Mechanisms of Path Dependence ===

Early on in the industry there is an expectation of human and capital investment from all stakeholders involved, this high cost for entry into the market and expected capital to sustain and develop a model by both consumer and enterprise makes for a high switching cost. For the AI model to run and scale there is a presumption that the rate of growth is dependent on the size of data fed, resulting in a better model and hence, more consumer usage. A closed centralised data loop is created but the requirement for storage and processing far surpasses current infrastructures capabilities - heavily relying on early market entry advantage and sustained capital investments taking away from smaller aspirants and hoarding of power within closed circles. Despite its physical dependency, AI data processing has unique requirements beyond legacy infrastructures (4) - larger demand and production of GPUs and designed to handle immense thermal heat generated while consuming electricity beyond current capabilities. 


The AI space largely remains unregulated due to its undefined form and fluid nature, however, this has given way for unethical and morality issues regarding data privacy, exploitation of labour and environmental harms.

Politics and Distribution

Concentration strengthens the values, interests, capacities, and authority of the “Big Tech” over other actors of public and private nature. In this context, Barbrook and Cameron (5) illuminate the ethos behind AI technology development while Winner (6) makes observable the power relations surrounding it: concentration is the consequence of an AI development anchored in the belief that it must be free from all institutional and normative restriction since it is inherently liberating and non-political which, in turn, provides a political/moral justification for the concentration (if technology is inherently liberating and non-political then there is no power relations to observe or question) that translates itself into power dynamics where the AI developer not only owns the AI technology but also decides how it is trained and deployed and which models the user can or cannot access.

Contingency and Alternatives

Concentration is due, first and foremost, to the fact that AI computing capacity requires massive and cutting-edge physical infrastructure - such as data centers and microprocessors - which, in turn, requires upfront investments that are unaffordable to most companies. In David’s (7) language, companies that could finance the infrastructure and position themselves first obtained self-reinforcing advantages and a “lock-in” effect in the market. In other words, concentration is the consequence of the development of AI physical infrastructure privately with no mixed or public alternatives as a public service. Therefore, there was a space for an alternative future at the end of the 1980s or the beginning of the 1990s when the privatization model of the technological infrastructure was consolidated in line with the “Californian Ideology”. In that moment governments had the chance to treat technological infrastructure as a mixed or public good rendering it universally accessible as it was the case of the Internet; this would have permitted companies of all sizes to profit from economies of scale, reducing entry barriers and fostering competition. However, this alternative path did not prevail since the Californian Ideology provided the moral/political justification to exclude governments from the development of technological infrastructure in the exact moment in which such a decision would produce structural consequences. If that path had been followed it could have contributed to collective access to economies of scale, lower entry barriers, and a more competitive AI market.

Governance Implication

Considering the above, public or mixed investment on AI physical infrastructure is a necessary governance measure for market decentralization and collective access to AI computing capacity and data. From an historical perspective, the proposed measure is not merely a corrective action for a market failure but for a contingent and ideological decision made between the late 1980s and early 1990s that produced a structural “path dependence” that now is necessary to change.

References

(1) The Quiet Re-Centralization of Tech: Why AI Is Reversing 20 Years of Platform Decentralization. (2026, January 10). Retrieved on 8 September 2026, from Medium website : https://medium.com/@sumitbhardwaj1357/the-quiet-re-centralization-of-tech-why-ai-is-reversing-20-years-of-platform-decentralization-ce15962b6a13

(2) Ginzburg-Ganz, E., Lifshits, P., Machlev, R., Belikov, J., Krieger, Z., & Levron, Y. (2025). Technical challenges of AI data center integration into power grids—A survey. Energies, 19(1), 137.

(3) Singh, A., Lu, C., Gupta, G., Chopra, A., Blanc, J., Klinghoffer, T., ... & Raskar, R. (2024). A perspective on decentralizing AI.

(4) The Rise of AI Data Center Models and the Decline of the General-Purpose Data Center. (Circa 2026). Retrieved on 8 September 2026, from SubZero Engineering website : https://www.subzeroeng.com/the-decline-of-the-general-purpose-data-center/

(5) Barbrook, R., & Cameron, A. (1996). The Californian Ideology. Science as Culture, 6(1), 44-72.

(6) Winner, L. (2017). Do artifacts have politics?. In Computer ethics(pp. 177-192). Routledge.

(7) David, P. A. (1985). Clio and the Economics of QWERTY. The American economic review, 75(2), 332-337.

== Centralization of AI compute and Data - by Sergey Abramov and Paul Stein ==

The current governance issue

Advanced AI currently depends on a narrow group of tech giants1 and semiconductor oligopolies controlling microchips and software, hyperscale cloud infrastructure, large datasets, frontier models, and distribution platforms. This presents a governance problem where this narrow group exercises control over foundational digital infrastructure. This directly affects consumers, businesses, and academia out of AI. Legally and institutionally, this matters because it creates a profound democratic deficit: this neutralizes antitrust frameworks designed for traditional markets, often bypassing conventional copyright and data privacy protections through unchecked data harvesting thereby complicating enforcement.This ultimately might affect national sovereignty as the countries that host the buildout will shape the future of AI: who controls it, what values it contains, how it is used2.


Historical lineage

In 1987, Taiwan Semiconductor Manufacturing Company (TSMC) was founded, establishing the world’s first semiconductor fab. Instead of designing and making its own chips, TSMC only manufactured chips designed by other companies. Today, nearly 100% of the world's most advanced AI processors are manufactured by TSMC, leaving the entire global AI compute supply chain centralized under one company.


In 2006, Nvidia launched CUDA (Compute Unified Device Architecture), a free software layer that allowed developers to use Nvidia's graphics chips (GPUs) which later became perfect for training deep neural networks. Today, this creates a near-monopoly on AI compute hardware.


Soon after Amazon launched AWS, Microsoft and Google launched their own cloud platforms marking the structural pivot to the cloud service provider model, shifting computing from a distributed architecture (where companies bought and managed their own physical servers) to a centralized infrastructure. As training modern AI models requires tens of thousands of GPUs, today’s frontier AI companies are entirely dependent on the extremely expensive physical infrastructure of a handful of cloud hyper-scalers (Microsoft, Google, AWS, and Meta).


In 2009, Fei-Fei Li launched ImageNet, a massive database of over 14 million images, which proved that AI model performance scales directly with data volume. This created an immediate institutional advantage for consumer internet platforms (like Google and Meta), because they were already architected to collect, centralize, and store masses of user data through search engines and social networks, they possessed the unique data stack required to train high-performing AI.


Mechanism of path dependence

The centralization of AI compute and data is a classic case of technological lock-in. When structural choices were made decades ago, they triggered powerful, self-reinforcing loops. One such loop is the switching costs and software standardisation. As developers have spent more than a decade working with CUDA, moving to a rival hardware provider (or a decentralized compute network) requires an enterprise to completely rewrite its foundational software stack. The cost of code migration, developer retraining, and potential hardware downtime force companies to stay tied to dominant hardware monopolies such as Nvidia. Commitments, licensing conditions, and exclusivity provisions add contractual costs. Fewer than 1% of cloud customers switch provider annually, identifying technical differentiation and data-transfer charges as important barriers4.


Network effects also matter for computing power. More customers generate more revenue, usage information, and opportunities to improve infrastructure and models. Better services attract additional users and developers, whose applications make the platform more valuable. In cloud–AI partnerships, providers may additionally obtain information about model performance and future infrastructure demand. These feedback loops strengthen firms already operating across cloud, models, software, and distribution.


Politics and distribution

According to Langdon Winner's definition, modern AI computing infrastructure may qualify as an "inherently political technology" because its large scale and significant financial requirements not only lend themselves to, but necessitate,5 centralized, hierarchical control. The physical siting of this infrastructure shows evidence of Winner’s second category of politicization of technologies, the ‘political arrangement’:6 Of the five largest U.S. AI firms, four are headquartered in California,7 while just one is (partially) domiciled in Texas. Texas, however, hosts twice the GPU clusters.8 This placement geographically insulates beneficiaries of the data infrastructure boom from its negative externalities. The specific choice of location may itself be a deliberate preference on account of Texas’ unique approach to water usage rights.9 Juxtaposition of this unequal and even exploitative balance of interests with the pontificating rhetoric of the urban, information rich, beneficiaries is evocative of the internal contradictions of the Californian Ideology - the virtual artisans existence and privileged position is still based on the exploitation of a vulnerable out-group, the "information-rich" consuming the output, the "information-poor" hosting the extraction.10


Contingency and Alternatives

Profit motive and economies of scale now concentrate investment among a few hyperscalers,11 but this reflects policy choice, not government incapacity: Compute infrastructure may be inherently political, but the technology is agnostic as to where the organizing authority originates, and states have historically funded infrastructure of comparable or greater scale, while facing less pressure to show short-term returns. The clearest turning point was 1995, when the National Science Foundation retired NSFNET, the publicly funded backbone that had connected U.S. research and supercomputing centers since the mid-1980s,12 handing internet infrastructure to commercial providers with no conditions attached. This was a deliberate policy decision made amid the deregulatory push of the Clinton era. The privatization that followed ushered in the private industrial dominance that persists today in the form of hyperscalers. Had a publicly governed backbone instead been scaled up, today's compute infrastructure might rest on a different foundation entirely.


Governance implication

Despite private hyperscaler dominance, the U.S. does have a nominal public alternative: the National AI Research Resource (NAIRR), established following the National AI Initiative Act of 2020 and piloted from January 2024, which explicitly aims to democratize access to the computational resources essential to AI research.13In Europe, a comparable program is the EuroHPCJU,14 an initiative intended to pool resources to build supercomputing capacity. The undertaking’s legislative basis has since been extended to include networked ‘AI factories', as well as the so-called “Gigafactories”.15 The framing in language of sovereignty indicates intent to at least compete with private infrastructure initiatives. Though government-led measures cannot reverse the centralization of compute, they can democratize access - provided they are intensified. The European Commission has “set a goal of investing more than €1 billion per year in AI research and innovation”.16 While laudable in its own right, it pales in comparison to the investments of industry leaders, which are approaching, if not exceeding, trillions of Dollars.

References & Notes

1.Amazon, Google, Meta, Microsoft, and Oracle collectively hold an estimated 71% of the world’s cumulative AI compute as of Q4 2025, measured in H100-equivalents of computing power. 

Luke Emberson, Josh You and Venkat Somala (2026), "Five hyperscalers now own over two-thirds of global AI compute". Published online at epoch.ai. Retrieved from 'https://epoch.ai/data-insights/hyperscalers-control-most-compute' [online resource]. Accessed 7 Sept 2026.

2. Alasdair Phillips-Robins, Teddy Tawil and Sam Winter-Levy, ‘The Compute Coalition: How to Build the Future of AI in the Free World’ (Carnegie Endowment for International Peace, 8 June 2026) https://carnegieendowment.org/research/2026/06/the-compute-coalition-how-to-build-the-future-of-ai-in-the-free-world accessed 7 September 2026.

3. Amazon Web Services (AWS), a cloud computing platform

4. Competition and Markets Authority, Cloud Infrastructure Services: Final Decision Report (31 July 2025) https://assets.publishing.service.gov.uk/media/688b8891fdde2b8f73469544/final_decision_report.pdf accessed 7 September 2026.

5. Langdon Winner, 'Do Artifacts Have Politics?' (1980) 109(1) Daedalus 121, 128.

6. Ibid. 123.

7. OpenAI, Anthropic, Google DeepMind, Meta AI. xAI is an exception to the rule, domiciled between Palo Alto, CA and San Antonio, TX.

8. Texas’ 17 to California’s 8, cf. Texas Royalty Brokers, 'AI Clusters by State' (Texas Royalty Brokers, 2025) https://texasroyaltybrokers.com/ai-clusters-by-state/ accessed 7 September 2026.

9. The ‘rule of capture’, established in Houston & TC Railway Co v East (Tex 1904), permits landowners to extract groundwater beneath their land without liability for depleting adjacent wells or aquifers, a doctrine now exploited by hyperscale AI data centres to draw large cooling-water volumes with minimal regulatory constraint.

10. Richard Barbrook and Andy Cameron, 'The Californian Ideology' (1996) 6(1) Science as Culture 44.

11. Goldman Sachs, 'Global AI Investment Is Forecast to Exceed $1 Trillion in 2026' (Goldman Sachs Insights, 7 August 2026) https://www.goldmansachs.com/insights/articles/global-investment-is-forecast-to-exceed-1-trillion-in-2026 accessed 7 September 2026.

12.  National Science Foundation, 'Birth of the Commercial Internet' (NSF Impacts) https://www.nsf.gov/impacts/internet accessed 7 September 2026.

13. CASRAI, 'NAIRR: National AI Research Resource, Explained' (CASRAI) https://casrai.org/dictionary/term/national-ai-research-resource-nairr accessed 7 September 2026.

14. Council Regulation (EU) 2021/1173 of 13 July 2021 on establishing the European High Performance Computing Joint Undertaking and repealing Regulation (EU) 2018/1488 [2021] OJ L256/1.

15. Council Regulation (EU) 2026/150 of 16 January 2026 amending Regulation (EU) 2021/1173 on establishing the European High Performance Computing Joint Undertaking [2026] OJ L 2026/150.

16. European Commission, 'Realising Europe's Digital Decade' (European Commission, 11 October 2024) https://commission.europa.eu/strategy-and-policy/priorities-2019-2024/story-von-der-leyen-commission/realising-europes-digital-decade_en accessed 7 September 2026.
Retrieved from "https://wiki.ivir.nl/index.php?title=Governance_by_history&oldid=124007"