There's an assumption underneath most of the AI conversation in America: the biggest model, trained on the most computing power, will be the most valuable. We call that assumption the frontier model fallacy. It's so common here that it usually goes unsaid.
But when we spent four days at the AI for Good Global Summit in Geneva, we discovered the assumption doesn’t travel well. Hundreds of people came through our booth. And not one of them asked us for a bigger model. What they asked for was AI they could actually run where they live and solve their real-world problems.
A summit built for the people AI is supposed to serve
AI for Good is the United Nations' flagship event on artificial intelligence, hosted by the International Telecommunication Union and co-convened with the Government of Switzerland. It draws government leaders, agency officials, researchers, and technologists from well over a hundred countries, many of them from the developing world.
A lot of AI events gather the people building the technology. This one gathers the people who have to put it to use, often in places where the version of AI being sold in Silicon Valley simply won't run. webAI joined as an exhibitor in the United States pavilion, alongside a small number of other American companies. Representing the country abroad wasn't something we took lightly.
People need AI that’s both accessible and sovereign
A teacher told us she wants to show her students how to use AI and can't get reliable access to it herself. A government official described trying to coordinate disaster relief in places where the internet goes down exactly when it's needed most. A delegate from a West African country said his options for adopting AI are limited because his country can’t build a data center.
None of them were asking for more raw capability. They were asking for intelligence they could reach.
Staff from UN agencies and international organizations raised the same problem from another angle. They handle confidential information that can't be shared across organizational lines, which rules out any tool that has to send their work somewhere else to be useful. A model that only works by routing sensitive data to an outside company's servers doesn't clear that bar.
The gap is real, and it's measurable
A June 2026 report from the United Nations University found that only 32 countries host AI-specialized data centers, that over 90 percent of that capacity sits in just two countries, the United States and China, and that more than 150 countries have little or no access to sovereign AI compute.
In other words, for most of the world the thing holding AI back isn't that the models aren't smart enough. It's that the models can't be run at all: not enough data centers, not enough power, not enough control over where the data goes. A more powerful cloud model does nothing for you if you can't connect to it, or can't legally send your data to it in the first place.
A different kind of model for a different set of problems
webAI came to the event to share a different vision: a model doesn't have to know everything. It just has to do the job at hand.
Most AI today is built the opposite way. One enormous model is trained on the whole internet to answer any question anyone might ask, which is why it takes a data center to run. But a model that can do your taxes and write a sonnet and diagnose an engine is carrying a lot of weight it doesn't need for any single job. Train a smaller model on one specific task instead, and it can match or beat the giant one on that task, while being small enough to run on hardware you already own. No data center. No constant connection. And because it runs on your own device, your intelligence never leaves it.
For a lot of the people we talked to, that was a new thought. Most had assumed AI meant one giant model in the cloud, because that's the only kind they'd been shown. Hearing that they could have a set of specialists, each one deep in the work they actually do, running on the hardware in front of them, changed the conversation. It described something they could actually use. When they heard our pitch and tried our demos, the feedback was overwhelmingly positive.
The frontier model fallacy is easy to believe when you have unlimited connectivity, cheap compute, and no reason to ask where your data goes. Take those things away, as they are for most of the world, and the assumption falls apart. A bigger model isn't better if you can't run it in your environment.