AI Exposure Is Highest Where the Desks Are

by Kent O. Bhupathi

AI;DR: AI exposure is currently highest in advanced economies because their employment structures are concentrated in clerical, professional, managerial and digital tasks that generative AI can automate or augment. Developing economies face less immediate displacement, but this partly reflects weaker digital infrastructure, institutional capacity and managerial capability, which also limit their ability to capture AI-driven productivity gains. Using Google’s proposed $15 billion data-centre hub in Visakhapatnam as a case study, the article argues that hosting large-scale compute infrastructure does not automatically constitute development strategy or technological sovereignty. Such projects should be judged by their long-term employment, local value capture, environmental costs, infrastructure demands and contribution to domestic capability. For most developing economies, the more credible sequence is to adopt existing AI, adapt it to local needs and advance towards deeper technological capacity only where demand, institutions and economic conditions justify the investment.

I had barely finished the World Bank’s World Development Report 2026 when my mind began drifting out to sea… well, at least back to Vizag, that is.

The report works to overturn a lazy assumption about AI. Poorer countries are often expected to suffer the first wave of job losses because they have weaker safety nets and less fiscal room. Yet current exposure is far higher in rich economies, where clerical, professional, managerial, and digital work is concentrated.

The World Bank estimates that 14.2% of jobs in high-income countries face meaningful automation risk, compared with 4.5% in developing economies. These are not unemployment forecasts, but the three-to-one gap matters. Rich countries have more accountants, analysts, administrators, coders, and customer-service workers in AI’s path. A language model can draft a contract or summarize a report. It cannot harvest crops, repair a roof, or care for a child from inside a server rack.

For developing economies, the better bet is usually adoption, rather than imitation. Most stand to gain more from adapting existing AI than from trying to recreate the frontier-model, semiconductor, and hyperscale-compute systems of the United States or China.

That is what brought me back to the proposed $15 billion (1.4 lakh crore INR) Google data center hub in Visakhapatnam (a.k.a, Vizag). This coastal city already faces water stress, and the project raises serious questions about environmental assessment, infrastructure demands, and its proximity to the Kambalakonda Wildlife Sanctuary.

India should participate in the AI economy. Nobody will ever say otherwise. But not every giant data center qualifies as development strategy.

The Geography of Disruption Has Flipped

The World Bank’s findings are striking to many, but they are not mysterious. AI exposure follows the structure of employment.

In high-income economies, millions of workers spend much of the day producing, processing, or communicating information. They write reports, review applications, reconcile accounts, answer routine questions, prepare legal documents, schedule appointments, analyze data, and maintain software. These are precisely the tasks generative AI handles best.

The International Monetary Fund reaches the same broad conclusion using a different method. It estimates that AI could affect roughly 60% of employment in advanced economies, 40% in emerging-market economies and 26% in low-income countries. When the IMF narrows the category to highly exposed work with low potential for human-machine complementarity, the shares fall to 33, 24 and 18% respectively.

The IMF and World Bank percentages should not be treated as competing forecasts. The IMF asks whether AI could materially affect the tasks within an occupation, then distinguishes between work likely to be complemented and work more vulnerable to substitution. The World Bank applies further adjustments for connectivity, local task content, and practical deployability. Different definitions shift the percentages, but advanced economies remain consistently more exposed.

Exposure, however, is not the same thing as elimination. AI will often change the composition of a job before it removes the job itself, automating some tasks while making others faster or more valuable. Whether that leads to displacement or growth will depend on how firms use the productivity gains, how much new demand those gains create, and whether workers have the power and support to share in the benefits.

Early evidence from labor demand nonetheless supports the rich-country gradient. A World Bank study examined about 555 million online job postings across 84 countries between early 2021 and mid-2025. After ChatGPT’s release, postings in highly substitutable occupations fell by an estimated 5.8% relative to less-substitutable occupations in high-income countries. The gap widened over time, reaching 8.9% by the third year. Outside high-income countries, the estimated decline was only 1.5% and was not statistically significant.

That does not settle the employment debate. Online vacancies skew toward formal, urban, and skilled work, but they still reveal a basic constraint. AI cannot disrupt jobs at scale until the economy is digital enough to use it.

The Exposure Paradox

Developing economies may be less exposed to immediate automation, but they are also less equipped to capture AI’s productivity gains.

The World Bank estimates that 16.2% of developing-country jobs have meaningful potential for augmentation, compared with 4.5% at risk of automation. On paper, that is an encouraging balance. AI could help scarce doctors review records, assist teachers with lesson preparation, help agricultural officers translate advice, and allow public administrators to process cases more quickly.

Task-level experiments show why the promise is credible. In one study involving 453 college-educated professionals, access to ChatGPT reduced the time required for writing assignments by 40% and improved evaluated output quality by 18%. Lower-performing participants gained the most.

Yet a tightly defined writing task bears little resemblance to running a ministry, clinic, or small business. Useful AI depends on the quality of the institution around it. A system can produce an answer in seconds, but someone must still know when it is wrong, how to use it when it is right, and whether the organization can act on it.

A randomized experiment involving 640 Kenyan entrepreneurs points in this direction. Participants received access to a GPT-4-powered business adviser through WhatsApp. On average, the tool produced no statistically clear improvement in revenues compared with a conventional business guide. And these averages concealed a sharp divide. Stronger firms gained roughly 15% or more, while initially weaker firms experienced declines of about 10%.

Ultimately, Access did not equalize managerial capability. The tool’s value depended on whether firms could assess its advice and translate it into action. Rather than substitute for management, AI increased the returns to it.

This is the risk the ILO and World Bank describe as “disruption without dividend.” Automation can spread through a thin layer of connected firms long before the wider economy has the reliable power, connectivity, usable devices, trained workers, and secure data systems needed to convert it into broad productivity growth. “Disruption” can travel down a narrow digital corridor… but the dividend cannot.

Low exposure is therefore an ambiguous advantage. It can mean that AI has fewer opportunities to replace workers. It can also mean that too few workers have access to the tools that could make them more productive.

Vizag’s Compute Mirage

The proposed Visakhapatnam data-centre hub should be evaluated through that lens.

A project of this scale may bring real benefits. It can create construction activity, improve digital infrastructure, generate tax revenue, and strengthen access to cloud services. After all, India has a large domestic market and a deep technology sector, both of which are strategic reasons to expand its computing capacity.

But size is not an economic argument. A server farm does not become development policy simply because it is enormous and has legacy branding.

The relevant questions concern value capture and opportunity cost. How much reliable employment will remain after construction? Which local firms will enter the supply chain? Who owns the models, patents, platforms and customer relationships built on the infrastructure? What public concessions, grid upgrades, water allocations or land commitments are required? Who absorbs the environmental costs if the projections prove optimistic?

Remember, hosting infrastructure is not the same as controlling the AI stack. A country can supply the physical base while strategic authority and economic value remain elsewhere.

Frontier AI is concentrated at nearly every layer. Stanford’s 2026 AI Index reports that the United States produced 59 notable AI models in 2025, while China produced 35. The United States hosts roughly 5,427 data centres, more than ten times the number in any other country. TSMC fabricates almost all leading AI chips. Advanced lithography, chip-design software, specialized materials, cloud platforms, and model talent are similarly concentrated.

Even the European Union, with its large integrated market and wealthy member states, is pooling resources on a continental scale. Its AI gigafactory plan envisages facilities containing more than 100,000 advanced processors, with up to €10 billion in public financing intended to mobilize at least €20 billion in private investment.

That is the scale required to play near the frontier. One subsidized facility does not dissolve dependence on foreign goods and services.

Energy and water constraints make location especially important (just a gentle reminder!). The International Energy Agency has warned that data centres impose large, concentrated electricity loads that can grow faster than generation and transmission capacity. Where grids are already strained, the costs may surface through higher tariffs, public subsidies or delayed connections for other users. Cooling systems can also intensify local water pressures.

For Vizag, the test should be practical. Its new, large data centre may very well pass above scrutiny. Time will tell. But it should not be granted a passing grade merely because “AI” appears in the project description.

Sovereignty Without Capability

The alternative to full-stack technological sovereignty is not helpless dependence.

Developing economies should build institutions capable of adapting imported AI, judging its performance, protecting sensitive data, and bargaining with suppliers. Demand and comparative advantage should determine where domestic infrastructure makes economic sense. Hosting compute alone creates little strategic capacity.

But strategic capacity ultimately rests on optionality. Governments should structure procurement so they can change suppliers without losing control of data or disrupting essential services. Regional facilities and compute vouchers can widen access at lower cost, while public infrastructure should serve only clearly defined needs such as security, latency, data residency, or research, and only where a credible user base can sustain it.

The World Bank’s sequence of adopt, adapt, then advance is quite useful here.

Adoption means giving firms, schools, clinics, and agencies practical access to existing tools. Adaptation means improving those tools for local languages, laws, and administrative systems. Advancement comes later, where domestic capabilities, utilization and economic conditions support deeper investment.

For many developing economies, the scarce input is not another frontier model. It is dependable electricity, affordable connectivity, usable public data, managerial competence and institutions capable of redesigning workflows. In parts of Sub-Saharan Africa, nearly one-third of rural schools lack reliable electricity and more than two-thirds lack dependable internet. Under those conditions, basic infrastructure may enable more productive AI use than an expensive national-model program.

India is not a small or technologically marginal economy. It has stronger reasons than most countries to invest in compute and domestic AI capability. That makes disciplined project selection more important, not less.

The World Bank’s deeper point is that limited exposure offers developing economies little comfort. It partly reflects the digital and institutional capacity they have yet to build. Governments now face the harder task of widening productive access and ensuring that infrastructure investment translates into real economic capability.

Vizag should face that test. Hosting a hyperscale data centre proves little by itself. The project must leave behind durable capability and broad economic value sufficient to justify the scarce resources it consumes. Development demands that harder standard.

Sources:

‍Barasa, Hilda, PeiChin Tay, Keegan McBride, Alexander Iosad, and Jakob Mökander. Sovereignty in the Age of AI: Strategic Choices, Structural Dependencies and the Long Game Ahead. Tony Blair Institute for Global Change, January 19, 2026. https://institute.global/insights/tech-and-digitalisation/sovereignty-in-the-age-of-ai-strategic-choices-structural-dependencies.

Cazzaniga, Mauro, Florence Jaumotte, Longji Li, Giovanni Melina, Augustus J. Panton, Carlo Pizzinelli, Emma Rockall, and Marina M. Tavares. Gen-AI: Artificial Intelligence and the Future of Work. IMF Staff Discussion Note SDN/2024/001. Washington, DC: International Monetary Fund, January 2024. https://www.imf.org/-/media/files/publications/sdn/2024/english/sdnea2024001.pdf.

Gmyrek, Paweł, Mariana Viollaz, and Hernan Winkler. Disruption without Dividend? How the Digital Divide and Task Differences Split GenAI’s Global Impact. ILO Working Paper 166. Geneva: International Labour Organization, March 2026. https://www.ilo.org/sites/default/files/2026-03/wp166_web.pdf.

Huang, Jingyun, Yan Liu, He Wang, and Shu Yu. Who Takes the Hit? The Uneven Impacts of Generative AI on Labor Demand Across Countries. World Bank, July 27, 2026. https://thedocs.worldbank.org/en/doc/07f7fe676b8331172af31de2c3fc11e1-0050062026/original/Cross-country-Labor-Demand-in-the-age-of-GenAI-0727.pdf.

International Energy Agency. “Data Centre Electricity Use Surged in 2025, Even with Tightening Bottlenecks Driving a Scramble for Solutions.” Press release, April 16, 2026. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions.

Lundy, Timothy, Manuel Ramos-Maqueda, Shyam Jayanti Patel, Daniel Rogger, Flavia Sacco Capurro, and Josefina Silva Fuentealba. How Is Government Using Artificial Intelligence? World Bank, 2026. https://thedocs.worldbank.org/en/doc/a2d967023f2d5cba345a3a2b9d72f837-0050062026/original/How-Is-Government-Using-AI-final.pdf.

Noy, Shakked, and Whitney Zhang. “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.” Science 381, no. 6654 (July 14, 2023): 187–192. https://doi.org/10.1126/science.adh2586.

Otis, Nicholas G., Rowan Clarke, Solène Delecourt, David Holtz, and Rembrand Koning. The Uneven Impact of Generative AI on Entrepreneurial Performance: Evidence from a Field Experiment in Kenya. Working Paper No. 24-042. Harvard Business School, December 2023. https://www.hbs.edu/ris/Publication%20Files/24-042_567db548-5b24-4182-b703-9d5ef1328043.pdf.

“Research and Development.” In Artificial Intelligence Index Report 2026. Stanford Institute for Human-Centered AI, 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report/research-and-development.

Sautmann, Anja, and Jonas Weinert. Artificial Intelligence Interventions to Support Frontline Healthcare Workers in Low-Resource Settings: A Review. World Bank, July 2026. https://thedocs.worldbank.org/en/doc/ae79099bd924fb298c1922e640f5ac4d-0050062026/original/Artificial-Intelligence-Interventions.pdf.

Tanner, Brooke, Cameron F. Kerry, Andrew W. Wyckoff, Nicoleta Kyosovska, Andrea Renda, and Elham Tabassi. “Is AI Sovereignty Possible? Balancing Autonomy and Interdependence.” Brookings Institution, February 17, 2026. https://www.brookings.edu/articles/is-ai-sovereignty-possible-balancing-autonomy-and-interdependence/.

World Bank. “AI Offers Lifeline to Developing Economies in an Era of Weak Growth.” Press release, August 4, 2026. https://www.worldbank.org/en/news/press-release/2026/08/04/ai-offers-lifeline-to-developing-economies-in-an-era-of-weak-growth.