Artificial intelligence is already reshaping industries, jobs, and the way organisations deliver value. But there is a hard truth we must face: without deliberate policy and design choices, AI risks amplifying existing inequalities. The technology’s promise is enormous — automating routine tasks, improving medical diagnostics, boosting productivity — yet the economics of building and running advanced AI systems create structural forces that can widen the divide between the well-resourced and the underserved.

Why cost and concentration matter At the core of the problem is cost. Building, training, and operating state-of-the-art large AI models requires tremendous compute resources, specialised hardware, and highly skilled teams. Estimates for training leading models run into the millions — in some cases tens or even hundreds of millions of dollars — which means only a small set of large firms, deep-pocketed startups backed by big investors, and nation-scale labs can realistically develop them on their own. That concentration of capability naturally leads to concentration of economic value.

On top of training costs, inference (the day-to-day cost of answering user queries) also matters. While inference costs per token have fallen over time, high-quality, low-latency deployments at scale remain expensive — and many businesses offset those costs by charging users or by embedding services behind paywalls or paid APIs. Where AI is priced out of reach, people and businesses with limited means are excluded from both the productivity gains and the new services AI enables.

The connectivity and access bottleneck The problem doesn’t stop at compute. Even if models were free, billions of people still lack reliable internet access, affordable devices, or the digital skills needed to use modern AI services. Recent global connectivity statistics show roughly 2.6 billion people remain offline — a reality that maps closely to income and geographic divides. In low-income countries only one in four people may be online compared with nearly universal access in high-income countries. Without addressing this “usage gap,” AI benefits will cascade mostly to regions and communities already well connected.

Evidence on labour and wage impacts Policy and economics researchers are already studying how automation and AI affect workers and wages. A recent OECD analysis highlights that AI exposure differs across occupations and can exert downward pressure on wages in certain tasks or roles — particularly where tasks are routine or easily automated. That means lower-skilled workers or those in occupations exposed to AI-driven automation could face wage stagnation or job displacement unless supported with retraining, social protections, and access to complementary technologies that raise productivity.

Two plausible futures Concentrated AI — concentrated benefits. If the status quo persists (high costs + concentrated compute + limited access), large companies capture most AI value. Services are paid or subscription-based, leading to faster adoption and richer functionality for organisations and individuals who can pay. The rest face reduced relative opportunities and an expanding digital divide. Distributed AI — democratized benefits. With intentional public policy, open-source alternatives, infrastructure investments, and creative business models, AI can be broadly useful. Smaller, efficient models (and improved infrastructure) can power local services affordably; public-interest models and subsidised APIs can serve education, health, and public administration in low-income communities.

Which future we get is largely a matter of choices now.

Practical steps to make AI inclusive Below are pragmatic policy, industry and civic actions that can help tilt outcomes toward the second — more equitable — future:

1. Invest in connectivity and devices. Closing the offline gap is basic: more affordable internet, public Wi-Fi, and cheaper smartphones give people the means to access AI services. Coalitions between telecoms, governments and multilaterals can reduce hardware and data cost barriers.

2. Support open-source and efficient models. Encouraging the development and deployment of small, efficient open models reduces barriers. Smaller models can run on cheaper hardware or in edge environments, lowering inference costs for local apps and startups. Funding research into efficient inference and distillation directly reduces pricing pressure for end users.

3. Public-interest AI and subsidised access. Governments and foundations can fund “public models” for education, agriculture, health and legal-aid that are free at the point of use. Subsidised APIs for non-profits and schools — or capped free tiers — would ensure critical services reach low-income users.

4. Re-skill at scale, with targeted programs. Labour-market interventions — from modular micro-credentials to work-based retraining and wage insurance — can protect workers at risk of automation while helping them move into higher-value roles that complement AI.

5. Tax and redistribute AI rents wisely. If major productivity gains concentrate in a few firms, fair taxation and smart use of revenues (for social programs, connectivity, and public AI infrastructure) can offset inequality and fund inclusion efforts.

6. Encourage local data centres and regional cloud capacity. Localising compute and data storage reduces latency and cost for regionally-focused services, helps comply with data sovereignty needs, and grows local tech ecosystems that can build AI products tuned to local languages and contexts.

Why business leaders should care Inequality isn’t just a social problem — it’s a business risk. Markets with deep structural inequality have lower aggregate demand, fragile institutions, and higher political risk. Companies that invest in shared prosperity — by offering affordable tiers, localised product versions, and partnering on public-interest deployments — build more durable markets and stronger brands.

A concrete ask If you’re building AI today, ask three questions about every product:

Who cannot afford or access this product today? What is a low-cost, low-friction variant we can offer? How can we partner with public institutions or NGOs to distribute value more broadly?

Conclusion — build inclusive AI or watch inequality widen AI can accelerate human progress, or it can concentrate gains among the few. The difference will be made by policy choices, business models, technical innovation (efficient models and edge AI), and sustained investment in connectivity and skills. We have a narrow but powerful window to steer AI toward inclusion. If we miss it, the tech that promised to lift lives could instead widen the gap between rich and poor.