To curb AI-driven inequality, we must unlock global access.

frankly, the UNDP report’s headline gets the story half-right and then walks off stage.

“AI risks sparking a new era of divergence as development gaps between countries widen” is directionally accurate. Digital change will not arrive as a benevolent global tide; it will hit where infrastructure, capital and institutional capacity already sit. The report lands the warning shot, but it largely stops at mood. It flags the risk without really tracing the mechanics of how divergence hardens—or how it might still be reversed.

Start with where the report is strongest: AI as a rich-country booster. The basic mechanism it sketches is plausible enough. Advanced economies host the firms building and owning AI platforms; those platforms then export services that undercut or absorb local providers elsewhere. Rents concentrate where the IP, data centers and legal firepower live. That’s not a new pattern—just a new technology riding an old rail line.

But that story is missing layers.

The first blind spot is inside the nation-state. Development gaps are not just a North–South story; they’re capital-city vs. everywhere-else. When AI lands on top of existing advantages—dense talent pools, deep financial markets, friendly regulators—you get acceleration in a few metros and stagnation in everyone else’s zip code. Rural regions, informal sectors and small cities will experience the automation pressure without the buffers: weak safety nets, thin capital markets, limited retraining options. National averages will look fine long after politics turns ugly.

Second, the report underplays corporate strategy and market structure. If foundational models and data flows are effectively controlled by a small set of multinational platforms, countries with limited bargaining power will find themselves locked into someone else’s stack on someone else’s terms—paying for access, trading data for functionality, or conceding regulatory space for “partnership.” That’s not just inequality; that’s dependency.

There’s a historical echo here. Think about the early days of mobile telecoms: vendors like Nokia and Ericsson sold the infrastructure, and many emerging markets adopted whatever stack arrived first. Years later, switching costs made it punishing to renegotiate. AI platforms risk replaying that script at software speed and with far higher stakes for data, governance and industrial strategy.

Policy, then, isn’t a “nice to have”; it’s the gearbox. The report treats “risk” as a macro headline when what we really need are instructions. Industrial policy that lowers barriers to compute, support for open models that can be localized, cross-border data arrangements that keep some value anchored domestically, migration and visa channels for AI talent so people can circulate without permanent brain drain—these are not theoretical. They’re choices.

Back when I was sitting through slide decks at Goldman, the math didn’t lie: countries that shaped their own capital formation and institutional rules turned new technologies into multipliers. Those that outsourced design decisions to external vendors got efficiency, then dependence. With AI, the equivalents will be sovereign or shared compute capacity, education systems skewed toward complementary skills (data stewardship, domain expertise, critical reasoning), and legal regimes that keep a single vendor from becoming the only gatekeeper.

The report also leans a bit too hard on the doom. There is a credible counter-argument: lower-cost AI tools, accessible via cloud platforms or open-source projects, can help micro and small firms raise productivity, especially in services. A small logistics operator that automates routing, or a health clinic that uses basic diagnostic support, does not need a national data center to see gains.

But that upside is tightly conditional. It assumes stable electricity, affordable connectivity, functioning institutions and some baseline digital literacy. Where those fundamentals are missing, cheap AI tools don’t magically create them. And even when small firms benefit at the margin, that doesn’t negate the macro-level capture if a few global players are skimming fees, data and attention from the entire system.

There’s also a time dimension the article barely touches. In the short run, countries that import AI services might see faster “catch-up” on specific applications—chatbots for government services, translation, simple back-office automation. Over the long run, if they don’t build any of the underlying capability, they cement themselves as permanent subscribers rather than co-creators. Short-term convergence can coexist with long-term lock-in.

On blind spots, I’d extend the list.

Yes, geopolitics of standards will matter: fragmented regulatory regimes raise compliance and engineering costs, especially for late adopters who can’t dictate terms. Yes, financial channels matter: venture money will continue to flow to jurisdictions where exits are clear and IP protections predictable, starving others of scale capital. But the social politics are more than background noise. Rising inequality inside countries—fueled by AI-augmented high earners and displaced low earners—can generate protectionist or nationalist policies that cut off exactly the international cooperation the report wants.

Policy responses have to live in that mess. Think tax incentives that are contingent on local hiring and skill transfer, data-sharing compacts that preserve sovereignty while enabling interoperability, and training programs that emphasize using and governing AI in context, not just churning out generic coders for a labor market that already looks saturated at the bottom of the stack.

Private tech firms, meanwhile, will quietly choose whether their architectures act as fences or bridges. Open licensing models, transparent documentation, and contractual terms that share upside with data-providing countries can tilt value back into developing markets. Lock-in pricing, opaque model behavior and aggressive IP tactics will do the opposite. The report could have pressed harder on that corporate fork in the road.

So yes, the headline is a warning. But unless public policy grows teeth and countries negotiate harder on standards, data and compute, that warning is going to read less like a risk and more like a quarterly update.