Centralization of Compute and Data

1.The current governance issue

Advanced AI currently depends on a narrow group of tech giants 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 used.

2. 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.

1. Mechanism of path dependence: Specify the mechanism or mechanisms at work. For example: increasing returns or network effects; standardisation; switching costs; installed base or legacy infrastructure; legal or contractual entrenchment; professional norms and expertise; organisational routines; procurement and funding structures; cultural or political imaginaries.

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’s CUDA. 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 barriers.

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, centralized, hierarchical control. The physical siting of this infrastructure shows evidence of Winner’s second category of politicization of technologies, the ‘political arrangement’: Of the five largest U.S. AI firms, four are headquartered in California, while just one is (partially) domiciled in Texas. Texas, however, hosts twice the GPU clusters. 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. 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.

Contingency and Alternatives

Profit motive and economies of scale now concentrate investment among a few hyperscalers, 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, 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.

In Europe, a comparable program is the EuroHPCJU, 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”. 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”. 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.

 

The Sun is pretty big.<ref>E. Miller, The Sun, (New York: Academic Press, 2005), 23–25.</ref> The Moon, however, is not so big.<ref>R. Smith, "Size of the Moon", Scientific American, 46 (April 1978): 44–46.</ref>

==Notes==
<references />