AI at the Frontier of the New Industrial Policy Turn: How Should Africa Respond?

PART 1

The past few years have seen an expansive return of industrial policy (IP) — a set of government interventions aimed at economic growth or restructuring through the provision of targeted assistance to specific industries or firms. In particular, COVID-19-induced supply chain disruptions have revealed several countries’ declining productive capacities, for which concerted government interventions have been necessary. But IP’s renewed popularity does not entirely rest on justifications of market failure or externalities such as Covid19. National security concerns and growing geopolitical competition (especially the rise of China) are creating an impetus for government support of strategic sectors such as the green transition and emerging technologies, especially in advanced economies.

AI and associated industries are at the centre of this industrial policy boom. State investment and investment incentives in AI-enabling infrastructures have increased significantly in the last few years, and so has the use of regulations to stifle competition. In 2023, the semiconductor industry — a critical input for AI GPUs — and the earth minerals that power it constituted the most active sectors of IP use. While popular narratives around AI industrial policy have centred EU, US and China tool boxes — such as the US’ subsidisation of green data centre construction, China’s extensive institutionalisation of R&D and tax grants for AI firms, and strategic regulatory deployment, such as the US Chips and Science Act, the AI Diffusion Rule (now repealed), and the EU Chips Act — state-led AI development is equally gaining popularity among emerging powers. For example, India, through its India AI Mission, has committed to investing over $1.2 billion in AI over the next five years in order to expand its compute capability, support local tech startups and upscale semiconductor manufacturing. The UAE has positioned itself as a rising AI innovation hub, creating an enabling environment for foreign direct investment, investing in the development of alternatives to Nvidia's AI chips, and establishing a university solely dedicated to AI research and training.

Increasing levels of interventionism in an industry that traditionally thrives on market competition prompts two forms of scepticism: first, that, historically, government attempts to create national champions or winners often engender state capture and monopoly that distort innovation, which becomes even more pronounced when we look at the centrality of large firms (big techs) to the recent state AI investments in many economies. The second critique is whether the socioeconomic value derived from AI so far justifies the public policy attention it is gaining, given the fiscal cost of these interventions.

This piece interrogates whether an industrial approach to AI should be pursued in the African context, given the continent’s current institutional and economic constraints. It argues that arguments around state capacity, fiscal feasibility, and the potential of a bubble burst should not limit African states from taking a big bet on AI industrialisation.

In defence of AI industrial policy in Africa

The issue of implementation capacity is a key question for African and other low-income countries, given that less distortive instruments (such as trade finance, public procurement reforms, and loan guarantees) require greater administrative and fiscal capacity, which is often lacking in these countries. There is also the point as to whether developing extensive local AI capabilities should be a priority for a continent grappling with other, more pressing socioeconomic challenges — including, notably, digital inequality.

To begin, the notion of shelving state-led AI development because it requires large administrative and fiscal resources, which could be used for sectors requiring more urgent attention, does not always hold. In reality, government resources — even in the most advanced economies — are finite, and policymakers constantly have to balance trade-offs between near-term exigencies and long-term strategic capabilities. In fact, the experience of the East Asian late industrialisers shows that industrialisation can happen amidst severe constraints in state capacity, human capital, and finance. For example, South Korea was able to, amidst the political challenges it faced after the war, institute educational reforms which staffed public sector bureaucracy for policy implementation, instituted land reforms to expand its agricultural tax base, and created a world-class integrated state-owned steel producer (POSCO) despite no domestic precedent in the industry and opposition from major external financiers like the US and the World Bank. Bi-directional causality is also true: that industrialisation is both a driver and a product of institutional, fiscal and social development outcomes. In short, the real question for African states is not whether they should commit already scarce resources to a new sector, but how to achieve significant economic gains from AI that make large investments in the sector defensible.

There is no simple answer to this. Arguments dissuading states from large investments in AI are backed by several empirical studies on AI’s economic impact, including those of notable economist Daron Acemoglu. Acemoglu argues that not only will productivity gains from AI be minimal, but their effect will also increase inequality, albeit modestly. Alternative estimates from Anthropic, the creators of Claude AI, show that the widespread adoption of AI could lead to annual US productivity growth of approximately 1.8%, potentially doubling economic growth rates in the long term. However suspect one may be of this estimate due to the limitedness of the data, the narrow designation of LLMs as AI, and concerns around authorship credibility (given Anthropic’s position as an AI company), the fact is that AI’s economic trajectory currently follows a pattern seen in other technological innovations: the productivity J-curve, which shows that in the short term, these technologies require massive investment whose benefits take time to materialise as the technologies diffuse and usage expand. As a result, there is often an underestimation of such technologies’ economic effects during the early years. As an example, accurate estimation of the productivity impact of digital innovations that began as far back as the early 90s remains a challenge in many national accounts.

History also teaches that government support is critical to derisking and providing strategic direction to private sector investment in new technologies. The internet as we know it today began as a US Defense Department-led endeavour (the ARPANET) to create a decentralised communication system following the USSR’s launch of its Sputnik satellite. If anything, the lesson from the massive initial capital investment in the internet is that a limited period of profitability preceded an eventual ‘boom’. Even then, narratives of the internet’s bubble burst were already being floated in economic policy circles as of the early 2000s — an irony for a sector that will later become the fastest growing sector of the global economy ten years later. While this was going on, concerted state investments in economies like the US, China, South Korea, Japan, and some Western countries gave them a head start in what today constitutes a digital hegemony. At the same time, early narratives of ICT development in Africa were centred on technology adoption, mobile telecomms usage, and technology assistance. By the time the innovations began to diffuse globally, African countries lacked their own sustained technical capacity to build ICT infrastructures; rather, success was defined by frugal internet innovation and shoots of success in the digital payment sector. The failure to take an early bet on strategic investment in internet infrastructures shows up today in the continent’s growing digital inequality and difficulty in building sustained ICT development. There is a temptation to explain African states’ inability to build capacity using the very argument this piece tries to refute: that Africa’s ICT sector’s reliance on aid and tech transfer assistance gave little room to build self-reliance. This is, in fact, untrue; as the South Korean example shows, technical assistance speeds up domestic capacitation, and is no excuse for inaction.

The lesson from the history of state intervention in the internet industry is that popular projections of a bubble burst should not limit African states from throwing their weight behind a technology that is still up for grabs. Much of the world is behind two leading AI superpowers (the US and China), yet these two take AI industrial policy ever more seriously than the rest.

Like the telecoms boom era, there is an inclination to encourage African countries to focus on AI adoption and use cases. While use cases are necessary to determine AI’s actual socioeconomic impact, domestic adoption cannot be guaranteed if infrastructure vulnerabilities are not resolved, and no large-scale impact can occur without creating the incentives for AI infrastructure development. Even if the goal for African states is not geoeconomic competitiveness, meaningful domestic adoption still requires articulating some industrial blueprint.

Lastly, African countries must do away with the notion that the development of large-scale general-purpose models should not be pursued due to these models’ increasing environmental footprint. Similar to the adoption narrative, this school of thought promotes preference for lean, small language models (SLMs) tailored to sector-specific usage. This narrative, at best, creates a partial equilibrium: while SLMs address immediate, context-specific sectoral and language needs within African markets, helping bypass the limitations of accessing US and Chinese large models, it risks creating a fragmented, non-interoperable system. Besides, many of the so-called African SLMs are built on foreign models anyway. Arguments about the potential environmental footprint of large-scale AI development in Africa must engage existing debates on the hypocrisy of rich countries’ climate policies: Africa has barely had its chance to industrialise, yet disproportionately bears the cost of increasing industrialisation in the Global North and elsewhere. The point here is not that African countries must pursue massive AI infrastructure for data centralisation and processing, but that they must build strategic advantages that enable the pursuit of large-scale development, should they want it. The pursuit of geocompetitiveness in the value chain aside, industrial policy is necessary to chart a path toward technological catch-up.

To rephrase the words of renowned economist, Dani Rodrik: the key question for African governments is not whether they should do industrial policy for AI, but how to do it — a question that is explored in detail in Part 2 of this piece.