I grew up in Rio de Janeiro. My first startup ran on a shared hosting plan that went down every time it rained hard enough to flood the server room in Botafogo. That was 2010. The infrastructure has improved since then, obviously, but when I watch the AI discourse from San Francisco — all the breathless predictions about how AI will transform every industry — I keep thinking about that server room.

Nearly 4 billion people on this planet lack the basic prerequisites to use AI at all. Not the fancy prerequisites like “access to GPT-4” or “a data science team.” The basics: reliable electricity, internet connectivity, digital literacy in any language the models actually understand.

The Stanford HAI 2024 AI Index put numbers on this. In the Global North, 24.7% of working-age adults use AI tools. In the Global South, it’s 14.1%. And that gap is widening, not closing.

The Infrastructure Nobody Talks About #

Every AI discussion I’ve been in this year — at conferences, in board rooms, on calls with investors — assumes a baseline that doesn’t exist for half the world’s population. Stable power. Broadband. A device capable of running a browser that won’t choke on a web app. English fluency, or at least enough of it to prompt an LLM effectively.

The World Economic Forum flagged this as far back as January 2023, calling it the “AI divide.” The framing was polite. The reality is blunter: AI tools are being built by wealthy English-speaking countries for wealthy English-speaking users. Everyone else gets whatever trickles down, if anything.

Take language. ChatGPT handles English well, manages a few European languages passably, and produces garbage in most of the world’s 7,000+ languages. Try prompting it in Yoruba, Swahili, or Quechua. The outputs range from inaccurate to nonsensical. Training data reflects the internet, and the internet reflects existing power structures. If your language doesn’t have a large Wikipedia or a corpus of digitized books, the models barely know you exist.

Where the Cliff Is #

The adoption gap doesn’t decline gradually. It falls off a cliff. CSIS research from June 2024 mapped AI adoption against GDP per capita, and the correlation is stark: once you drop below roughly $20,000 per capita, AI usage collapses. It’s not a gentle slope; it’s a ledge.

This makes intuitive sense once you stop thinking about AI as software and start thinking about it as infrastructure-dependent software. Running a cloud AI service requires the cloud part to work. That means data centers within reasonable latency, reliable payment systems to buy API access, and developers who know how to integrate these services. Most countries in Sub-Saharan Africa, South Asia, and parts of Latin America are missing at least one of those pieces; many are missing all three.

Open-source models (LLaMA, Mistral) theoretically help — you can run them locally without depending on US cloud providers. But “locally” still means a machine with a capable GPU, stable power, and someone who knows how to set it up. The barriers shift; they don’t disappear.

It’s Not Just Access, It’s Relevance #

Here’s what frustrates me most about the standard “let’s bring AI to developing countries” framing. It assumes the AI we’ve built is useful there. Often, it’s not.

A crop disease detection model trained on images from Iowa cornfields doesn’t help a farmer in Senegal growing millet. A customer service chatbot trained on English support tickets doesn’t serve a WhatsApp-first market in Indonesia where conversations happen in Bahasa mixed with local dialects. A medical AI trained on clinical data from US hospitals misses diseases that are endemic in tropical regions but rare in North America.

The relevance gap is as real as the access gap. And it’s harder to fix because it requires not just deploying existing tools but building new ones — trained on local data, in local languages, for local problems. That takes investment, talent, and institutional support that the Global South largely doesn’t have.

Brazil is an interesting case because it sits in between. There’s real AI talent coming out of USP and Unicamp. The fintech ecosystem (Nubank, PagSeguro) has production ML pipelines. But outside Sao Paulo and a few other metros, the picture looks very different. I’ve talked to founders building for rural Brazil who describe connectivity conditions that would make any cloud-dependent AI architecture laughable.

What Actually Helps #

I’m skeptical of grand plans. The ones that work tend to be specific and pragmatic.

Mobile-first design. In most of the Global South, the primary computing device is a smartphone — often a low-end Android phone with limited storage and intermittent connectivity. AI applications that require a laptop, a fast connection, and a modern browser are irrelevant. The ones that work over SMS, WhatsApp, or lightweight apps on 2G/3G networks have a shot.

Local language investment. Not just translation layers on top of English models (those are terrible), but actual training data collection in underrepresented languages. EleutherAI and a few academic groups are doing this. It needs more funding and more attention.

Training data from local contexts. Agricultural AI for sub-Saharan Africa needs images of sub-Saharan African crops, disease patterns, and soil conditions. Medical AI for Southeast Asia needs clinical data from Southeast Asian hospitals. This is obvious, but the incentive structure pushes the other way: it’s cheaper and faster to build for markets that already pay for AI services.

Infrastructure, not just software. CSIS made this point clearly: you can’t leapfrog to AI if the electricity grid goes down three times a week. The boring investments — power, connectivity, digital literacy programs — are prerequisites. AI vendors don’t want to hear this because it’s not their problem to solve, but it’s the actual bottleneck.

The Part Nobody Wants to Say #

The uncomfortable truth is that the AI boom might make global inequality worse, not better. The countries with the infrastructure, talent, and capital to adopt AI will accelerate. The countries without those things will fall further behind. The productivity gains that AI delivers to the Global North become a competitive advantage against the Global South.

I don’t have a clean answer for this. I’m a CTO at a startup in San Francisco; I’m not in a position to fix the electricity grid in rural Nigeria. But I think anyone building AI products has a responsibility to at least acknowledge the world they’re not building for. The 4 billion people who can’t use what we’re making aren’t a future market waiting to be unlocked. They’re people being left behind right now.

The “AI will democratize everything” narrative is seductive. It’s also, for most of the world, wrong.