"We see a future where intelligence is a utility, like electricity or water, and people buy it from us on a meter." — Sam Altman, BlackRock U.S. Infrastructure Summit, March 2026

For the world's low- and middle-income countries, making the most of metered intelligence looks a lot like electrification: building the infrastructure to take delivery of it, and reorganizing their societies to put it to use.

It's important to have two expectations. The first is that models will keep getting better. The second is that serving any given level of capability will keep getting cheaper, something like an order of magnitude a year, as efficiency lets a small new model do what a large old one did and open-weight competition pushes the same way.

An LMIC cannot organize around the first. Pushing the raw capability frontier takes resources only the United States and China can assemble. But an LMIC can organize around the second: intelligence that gets cheaper every year. Today's models are severely underused, and I expect the gap to hold or widen as capabilities improve faster than people's use of them. Plenty of hard problems sit between a model and its best use, mostly around building the structures that give people and models the right context and agency. These countries can get a long way by best structuring themselves to absorb capabilities already on offer, and the better and cheaper ones still to come. Reaching the absorption frontier requires a secure supply of this utility and mass exposure to it.

Before putting metered intelligence to use, a country has to be able to take delivery, and for the typical LMIC there is a problem of infrastructure and cost. Some 2.6 billion people are still offline and hundreds of millions have no power, most of them in Africa. Where power and a signal are available, they can be unreliable, and for many businesses that is serious enough to stall growth. In the low-income world, a one percent rise in outages is estimated to cut output by about three percent. But coverage is not the only barrier. Several billion people live under a mobile signal and still do not use it, more of them in South Asia than anywhere, and in these cases the concern is usually cost. To use that signal takes an entry-level smartphone and a data plan, and both cost too much. The phone alone runs about a sixth of monthly income, more than half for the poorest fifth, and a gigabyte of data still runs about three times the affordability line, which the poorest countries have yet to clear. And even where none of that is prohibitive, the cost of the models themselves still sits on top. The American labs charge about twenty dollars a month for an entry-level subscription, and short of a subsidy or some way to sell access in far smaller units, that is more than most can pay.

There may be an instinct to fix some of this by building at home. But compute is a very competitive business line, and for an LMIC sovereign compute is mostly a vanity, the kind of thing none of the countries I have in mind could build soon. A frontier training campus can draw on the order of a gigawatt, the low end now as leaders build toward several (Kenya's entire grid is about three). Even a smaller build is hard. The billion-dollar Microsoft and G42 data center at Olkaria was an ordinary cloud region, not a training site, and it still stalled when the government could not guarantee the power and would not commit to the capacity payments, for a project meant to grow into a third of national supply. The campus is the cheap part anyway. A gigawatt of AI compute runs something like forty billion dollars, most of it chips, bought in hard currency, that have to be kept busy to pay for themselves and draw more power each generation. The firms at the frontier are investing at a scale beyond any LMIC. Most of the world is out of the frontier training race for good, and largely out of at-scale inference too, though smaller inference stays within reach. For the typical LMIC, the resources that would go toward a data center go much further spent on what lets people take delivery of metered intelligence: power, connectivity, devices, and the training and subsidies that support their use of it. A few do have the infrastructure, like Egypt, Morocco, and South Africa, and they should probably host some compute. For everyone else, power, connectivity, and supply are the ballgame, and self-reliance on compute can come later, if at all.

There are two broad categories of firms, American and Chinese, with trade-offs between the categories and among the firms in each. The United States sells the absolute frontier, the best models and chips, in approved contexts with export controls and end-use conditions I expect to grow more restrictive as the offerings harden into an American stack. You pay a small premium over the Chinese for that frontier performance, even as I expect the price to keep falling. China sells near the frontier, and at it in some contexts, mostly through commercial channels for now. But I see the Chinese offer as more of a bundle, where a preference for its ecosystem can more straightforwardly come with state financing and help building the physical infrastructure I described earlier, the power, the networks, and the data centers Huawei lays across the continent, some of it on Chinese state loans. And I expect that to couple more tightly soon, and not only in China. For now you can move between firms and categories fairly freely. In complex cases it takes more than an API key swap, but shared interfaces and routers keep even those manageable, though that may change. For the typical LMIC, the decider should be which firm or partner will more reliably guarantee your supply over the long run, and do more to help you put it to use.

The countries that make the things the superpowers need, the chips, the talent, or the machines that make them, can bargain harder, win better terms, and perhaps shape some of how the superpowers act, the way Tom Davidson lays out for the middle powers, the EU's move toward crisis powers over chip supplies one example. Most of the world cannot influence the capability frontier through hard power or soft. For the group this piece is concerned with, it comes down to the effective sourcing and distribution of supply.

The labs may not stay private for long. The Chinese ones already sit close to the state, and attempts to leave its orbit have been blocked, as when Beijing blocked Meta's acquisition of Manus. I expect the American ones to follow, sooner than most assume. Take the labs at their word about what they are building, and it is hard to picture Washington leaving them in private hands. On the open question of nationalization, I think a soft version for at least one lab by 2030 is extremely likely (around 88%, and a hard one around 15%). This is already serious, but for many it will click once it becomes a question of states dealing directly with states, rather than with the private companies of today, as the national securitization of intelligence matures.

For most people in an LMIC, their lives will be first touched by a model secondhand, inside software they already use or through an institution they deal with. Anthropic and OpenAI, for example, both work with the Gates Foundation, Anthropic on the everyday machinery of health ministries and OpenAI on a thousand African clinics, the kind of project I have not seen a Chinese lab take on yet. None of that is bad in itself, and much of the early good arrives this way, the clinics among it. But there is a ceiling to progress a country does not drive itself, and the work is to take this utility and start building with it.

Their first direct contact, though, will be with Meta's models, free inside the WhatsApp they already use. Most people will never download a standalone app to reach ChatGPT or DeepSeek (Meta has barred rival assistants from WhatsApp anyway). Meta is also the lab I find most interesting here. MSL is a relative latecomer and a rebuilding effort, and Alexandr Wang's recent comments point to where a lot of the value may come for these populations, a focus on health and on cheap models. Its newest model lands near the frontier while staying notably token-efficient. Wang talks about an economy of agents working for ordinary users and the firms they buy from, a vision that fits these situations well. The most important corporate relationship the typical LMIC has on AI may well be with this laggard, for reasons of distribution and economics.

Making the most of metered intelligence is hard, and it comes back to the same two things, securing the supply and putting it to best use. For most LMICs, one accountable person should lead both, at least to start.

Securing compute is the better-understood job. Public and private buyers do it constantly, and the quantities that suit different uses can be estimated, though they are easy to overshoot (see Uber). Optimal deployment is murkier still, at least so far. The labs are limited in how much they can serve, and have begun tightening access to their premier models, and I think Anton Leicht is right that this will only continue as capabilities advance. In a compute crunch, the capacity serving lower-priority customers may itself be pulled back to those who rank higher. So priority access is something a country can try to negotiate for, paid for in the geopolitical leverage it holds, and a superpower may guarantee or subsidize that supply as statecraft rather than sell it at market prices. I expect more of this as nationalization arrives. The accountable person's first job, then, is to guarantee a base level of compute for the country.

Making the best use of models is an experimental exercise. The right allocation and uses will differ a great deal by country, and they are things to be discovered, by trial and error and the collective intelligence of these populations. Put cheap models in front of as many people as you can, as early as you can (weak as they are in many local languages), and let them feel the capabilities, feel the problems around them, and find the best solutions they can. Steady contact with frontier models sobers you to the strange time we are in, and the stranger one coming, and it is some of the best preparation for it.

I could not have built Tebeldiya, a direct-delivery malnutrition charity, without these models. We did hard things with them toward running a large-scale delivery operation. Along the way I saw how a few NGOs used models and advised others, and watched it go well and badly, and what I took away was that almost no one was experimenting as often or as widely as they should have. I mean that even here, where the stakes were extremely high and a failure to execute meant unnecessary deaths.

Most of the best use in these contexts will not be centrally foreseen, or decreed. It is, in context, probably the state's job to provide the supply and the environment, encourage innovation, and then be smart about backing the winners that emerge. Reaching the retail edge is a task for private business, and there is room to get creative with the economics of access. A discount for users who let their data be sold for training is one idea. A broader adaptation of the lessons of the Chinese gray market for Anthropic access is another, taking its good, creative pricing and local payment, while stripping out its bad, the silent model-switching and the quiet harvesting of private data.

Most of the world's people live in LMICs, something like 80% of humanity. The terms on which they get metered intelligence will be important for their future flourishing. I would like to see a lot more thinking about getting that right.