NeelKuthi to Data Center: A Persisting Colonial Reality

In this Blog, I use dependency theory to explore the colonial reality that persists in today’s world, reading the AI race alongside the history of indigo cultivation in Bengal to show that while the means and strategies of colonisation have changed, the cycle of extortion remains unchanged.

In contemporary times, many believe that the concept of colonialism is no longer relevant, given that the world is in such a globalised setting. It is almost inspirational how many developing economies have achieved convergence in manufacturing with a strong balance of trade. However, a substantial change in technology might have turned things south for the developing economies. 

If we observe the ongoing race of Artificial Intelligence (AI), it appears that we are far from having overcome colonialism; instead, we have entered an era of digital colonialism complemented by a subtle systematic racism. 

Even Stiglitz has flagged in his NBER Working Paper that AI might be the next big thing that will disrupt the so-far convergence between developing and developed states. The divide may resemble that which existed before the decolonisation spree. Thus, it compels us to throw in another question: will AI be the reason behind, what Frank refers to, as “underdevelopment”? To answer this question, we can draw on the framework of “Dependency Theory,” arguing that the rising AI divide between the Global North and Global South is not merely a difference in technological capabilities but rather a conscious policy outcome. 

History repeats itself? 

The central thesis of the dependency theory is that the global capitalist system is fundamentally structured into two interconnected parts: a dominant, wealthy “metropolis” and a dependent, impoverished “satellite”. The wealthy metropolis systematically exploits the dependent satellite by extracting its economic surplus, concentrating wealth at the centre, and cutting the satellite off from potential investment funds. On the other hand, Samir Amin points out that, apart from international trade, the core actively distorts the internal economic and social structures of the satellite to serve the needs of capitalist accumulation in the core, which is the very essence of the underdevelopment thesis. 

First, for clarity, let us put the first "Industrial Revolution” on trial. By the normative view, the British institutional setting provided the right incentive mechanisms for technological capabilities, which led to the first industrial revolution. Then again, it was these very institutions that helped Great Britain become the global metropolis by exploiting its colonial satellites. For example, in the case of textile production, the system was designed in such a way that it prevented the colonies from producing textiles and forced them to produce raw materials (with forced specialisation of the goods the colonies themselves did not need) for the textile industry, while at the same time forcing the colonies, particularly the Indian subcontinent, to buy the textiles produced in British factories. The history of Indigo Plantation in the Indian Subcontinent provides some strong evidence. The British indigo producers, called "Neelkar", first incentivised and then forced local farmers in Bengal and Bihar to produce indigo instead of food crops so that the indigo could be used for textile production. 

Now, at this point, a question might arise: what does AI have to do with the story of Indigo farming or colonial intervention? Well, if we put on the dependency lens, the analogy can be shocking. 

First, some AI statistics. 

PwC projects that AI will add about $15.7 trillion to global GDP by 2030, with China and North America alone capturing almost 70%, while Latin America, Africa, and most of Asia share only 3–8% each. So, what is the role of the remaining regions in this economy? We like to argue that it will be like the indigo farmers, producing the raw material for a good they will rarely own. How so? With Data centres. 

Image credit: Geoffrey Moffett via Unsplash.

NeelKuthis to Data Centers

Data is the fuel for operating a strong AI model. But this data is not usable instantly; like indigo, it needs to be processed. In technical terms, this process is known as "data labelling". So, what is the problem with "data labelling"? 

When we scrutinise AI mechanics, we can see that data labelling hardly requires advanced technological capabilities; rather, it is very labour-intensive and is often regarded as psychologically draining. 

So when the AI hyperscaler's such as Meta, Google, and Amazon, create and recruit Business Process Outsourcing (BPO) data centres in the name of AI inclusion, these centres actually collect labelled data at a very low cost, which is then used in their product development models. This situation reminds us of the Neelkuthis, indigo processing and storage facilities that the East India Company's white officers supervised.

Paramanu Sarkar, Mangalgunj Neelkuthi, Parmadan. CC BY-SA 4.0, via Wikimedia Commons.

The local BPO firm that supplies labelling labour to the model builders act as comprador bourgeoisie: the class that development economists identified as foreign capital's local agents, whose profits flow from brokering their compatriots' labour and resources to the metropolis rather than from building self-sustaining national industry. Now this fact is worth relating to the fact that the Neelkuthi, after all, never ran on British East India Company stuff alone — it relied on its Gomastas, local agents who enforced the planters' cultivation terms on their countrymen. In a nutshell, BPO contractors are the new 'Gomasta' of the hyperscalers. And likewise, the indigo-dyed textile was sold at a higher price back to the satellites and the global market. 

“The AI circuit seems to follow the same method. Here the labelled data leaves the periphery at the cost of low-wage labour, it is fed into AI models, and it returns as AI products where subscriptions are offered at a premium”.

As a result, the model owners receive the surplus from the entire circuit, while the labelling economy leaves behind wages and salaries that are insufficient to fund any home model-building infrastructure. Thus, the data-producing countries are not simply lagging; the circuit deprives them of the investable excess required to catch up. The satellite's given role by the metropolis creates the transfer that keeps it a satellite. 

This “unequal exchange” and structuring of peripheral production processes to suit the economic needs of the core is reminiscent of the classic dependency theory!

Keeping the good stuff locked

Now, the fascinating part is that the reported 11,803 data centres, from 179 countries worldwide, show eagerness to set up new facilities; however, there is relatively less enthusiasm for sharing the actual science or the model with the data-labelling countries. If there were to be true inclusion, there would have been a significant number of AI model capacity-building initiatives in the local population (United Nations, 2024); but instead, the core managerial roles that might build the locality's modelling capacity are taken over by people from the tech giants or Global North countries. 

CSET's mapping of AI R&D at six US tech multinationals — Amazon, Apple, Facebook, Google, IBM, and Microsoft — found a striking mirror image: 68% of their 62 identified AI labs sit outside the United States, yet 68% of their AI staff sit inside it, with the Bay Area and Seattle alone hosting nearly half of the global total. The division of labour across these labs is even more telling: around 85% of labs in North America and Europe conduct basic research, while those in countries such as India focus on applied work and development. Thus, the frontier science stays at the metropolis; the periphery adapts and applies. However, one could argue that this is a classic case of technological "catch-up" or upgrading. In reality, this case lacks every process that actual upgrading requires. 

Data labelling transfers no modelling capabilities to the workers who perform it The peripheral facilities produce only what the metropolis models require, and even that output is enclosed within the parent firms' model rights, preventing any local accumulation. Earlier latecomers, such as the East Asian economies, could upgrade their technological capabilities as capital costs declined. In the AI frontier, the situation is the opposite: training costs increase year on year (Korinek and Stiglitz, 2021), and export controls on advanced chips act as the bottleneck that stifles upgrading in model development. Ultimately, the means to develop a localised model becomes difficult from the very start, creating a dependency that is built in by design.

We produce; I sell

While overall global AI adoption is indeed rising—with 78% of surveyed organisations worldwide reporting AI use in 2024—the rates of integration and the resulting economic benefits are drastically skewed toward countries in the Global North that hold model rights. Despite the absence of localised models in the majority of nations, the surge in AI adoption underscores how a handful of producers have successfully captured both domestic and foreign markets. 

This situation results in strategic gatekeeping by the hyperscalers, who, by refusing to share their models, have created a widening technological disparity. 

However, the problem is not disparity; rather, it is the structural gridlock that prevents disparity from being eradicated. The circuit locks the periphery's labour into the labelling niche and drains the surplus needed for domestic model-building, thereby reproducing underdevelopment in Frank's precise sense — not the absence of development, but the development of a subordinate form of it. Consequently, data-labelling nations remain entrenched in a state of dependency on model developers, finding it increasingly difficult to break free from this vicious cycle of value capture.

So finally, it is safe to state that while colonies may be a thing of the past, racial and colonial bias still prevail in the economic frontier. The AI race validates only that the means and strategies of colonisation have changed, while the cycle of extortion remains unchanged.

Header image credit: Scott Rodgerson via Unsplash.

About the author

Chowdhury Mujaddid Ahmed is Lecturer in economics at the Department of Economics, Metropolitan University Bangladesh, and MSc graduate of Economic Policy from the SOAS University of London, Department of Economics.