Gates puts up $1 billion so AI doesn't leave poor countries behind, but OpenAI, Google and Anthropic set the rules

By Observer · October 7, 2026.

The Gates Foundation, with a $92 billion endowment as of the end of June, has committed at least $1 billion over the next two years to artificial intelligence projects: 40% to health, 40% to education and the rest split between agriculture and digital infrastructure, including local-language datasets. The announcement came in September alongside the Goalkeepers Report, the foundation's annual report on progress toward the UN's sustainable development goals. The line that sums up the bet is Bill Gates's own, written in that report: "AI could be a great equalizer, or widen the gap."

Gates's money does not fund AI research: it buys access and distribution, and the technology partners provide the rest. In May, Anthropic signed a four-year, $200 million alliance combining grants, Claude credits and technical support for health, education, life sciences and economic mobility; the company describes it as extending AI into areas where, in its own words, "the market alone does not reach." In January, OpenAI launched Horizon1000 together with the foundation, $50 million for a primary care pilot in Rwanda with the goal of reaching 1,000 clinics across Africa by 2028; the company puts the health workforce shortfall in sub-Saharan Africa at 5.6 million. In September, Google expanded — with $100 million in joint funding — an agricultural program that aims to reach 200 million smallholder farmers in sub-Saharan Africa and South Asia, with satellite plot mapping at sub-meter resolution, more accurate weather alerts and farming advice in farmers' own languages, drawing on organizations such as India's Wadhwani AI and Digital Green.

The underlying problem driving all this is one of data, not only of money. According to the Goalkeepers Report, more than 90% of the data used to train the first large language models came from English-language sources; Gates cites speech recognition systems with error rates below 6% in English against more than 60% in Yoruba, a language widely spoken in Nigeria and other parts of West Africa. An agricultural assistant needs to know local crops, pests and prices; a health one, the treatments actually available in each country. That requires homegrown data, and it raises the question of who controls it. The timing is no accident: the Goalkeepers organizers note that 26 of the world's 34 richest countries cut their development aid last year, while the UN describes progress toward its goals as "uneven and insufficient."

In that same package, the foundation last month joined a coalition of 60 organizations — among them Google, Microsoft, Nvidia, Anthropic and the OpenAI Foundation — to build, over five years, openly licensed language datasets aimed at the 3.4 billion people whose languages are barely represented in current models. Its governance, according to the article itself, has yet to be defined, which is no small detail: deciding who sets the rules for that data is, in practice, deciding who controls the next generation of AI useful to the global south.

Our reading is that two stories coexist here, and it is worth not settling for the more comfortable one. The first is genuinely good: Gates contributes money and decades of relationships with governments and NGOs that no AI company has, and in return gets the technology into clinics in Rwanda or farms in Ethiopia that the market, left to its own devices, would ignore as unprofitable. It is exactly the kind of transfer the abundance thesis needs: if AI is going to bring us closer to eradicating diseases and freeing up human time, it also has to reach those who cannot pay for it today, not only those who already can.

The second story is more uncomfortable: it also suits OpenAI, Anthropic and Google to be there first. Whoever supplies the health infrastructure, the language datasets and the standards with which a poor country starts using AI gains a distribution advantage that is not measured in benchmarks but weighs just as much; it is the same pattern of technological dependence we already saw with the cloud, now transferred to health and education. And it is happening just as public development aid is shrinking, turning a handful of private foundations and companies into de facto substitutes for state funding that is withdrawing, without the same democratic scrutiny. Gates acknowledges this in his own way when he writes that "the goal is not for AI to be designed far away and handed to communities to fend for themselves"; the test of whether that principle is real will be whether governance of those language datasets — still unresolved — ends up in the hands of the countries that generate them or of those who already dominate the model market.