I think the word cloud has enjoyed one of the better marketing decisions in technology. It sounds light, soft and pleasantly detached from anything expensive. Then you start looking at what keeps an AI system running and find buildings full of equipment, electrical connections, cooling systems and people who know how to keep the whole thing working. Somewhere behind that neat little chat window is an extremely physical operation. The name cloud leaves out a remarkable amount of plumbing.
That physical side is what interests me about the global AI divide. Access to powerful computing infrastructure is uneven, and its location and ownership can affect who gets to develop technology on their own terms. But the old headline about only 32 countries holding the future made a limited infrastructure count carry far too much meaning. This October 2026 update follows the research behind that number, then looks at the more useful questions of access, ownership, energy and who gets a meaningful chance to build.
What the 32-Country Claim Actually Measured
The International Telecommunication Union's 2025 AI governance report, citing Oxford researchers, reported that 32 countries hosted data centres supplying compute for advanced AI development. That is a dated infrastructure finding, rather than a count of every country that can use AI. A person can use a service hosted elsewhere, and research infrastructure can exist outside the particular facilities covered by a survey. The number needs its definition beside it, especially when it starts appearing in headlines.
The working paper by Zoe Jay Hawkins, Vili Lehdonvirta and Boxi Wu describes a census of nine major public-cloud providers using November 2024 availability data. The sample included US providers AWS, Google and Microsoft, Chinese providers Alibaba, Huawei and Tencent, and three European providers. The accessible paper version reports 33 host jurisdictions with AI accelerators, including 24 with hardware it classifies as training-relevant. Those counts differ from the ITU summary, and I haven't treated them as interchangeable or as a live October 2026 total. The paper also excludes private in-house and public-sector compute, which makes its coverage important to the interpretation.
Training a Model and Using One Are Different Jobs
Training adjusts a model's parameters using data, while inference means running a trained model to generate an output or make a prediction. IBM Research explains the distinction as learning patterns during training and applying the resulting model during use. Both require computation, but the scale and equipment depend on the model and the task. Training a leading general-purpose model, adapting an existing model and running a smaller system locally are different undertakings.
This is where the word building starts doing an impressive amount of work. A team might be creating an application, preparing a dataset, testing a model or training one from the beginning, and all of those can get swept into the same conversation about building AI. I want to know which activity somebody means before deciding what resources they need. Otherwise, we end up judging every useful project against the requirements of the largest training run anybody can afford.
Location, Ownership and Chips Give Different Answers
Hawkins and her colleagues separate compute sovereignty into three levels: where the infrastructure is located, the nationality of the cloud-provider firms and the nationality of accelerator vendors. A facility inside a country's borders can therefore coexist with dependence on an overseas provider or chip supplier. Their framework helps explain why attracting a data centre and gaining control over the entire supply chain are different policy goals.
Naturally, my wine brain immediately wants to complicate this. Knowing where a wine comes from gives me one part of its story. Who owns the winery, who makes the wine and how it reaches the people buying it tell me other parts. I'd ask different questions at each point, even while following the same bottle. I find it useful to bring that curiosity back to computing infrastructure. A location gives us somewhere to begin, and then there is a whole set of relationships to understand.
A Dot on a Map Has Limits
The OECD's October 2025 work on measuring public-cloud AI compute uses cloud regions and the availability of relevant hardware to describe the geographic distribution. A cloud region is an organisational grouping of infrastructure, rather than a standard-sized unit of computing power. Counting regions or identifying which hardware is offered can reveal something important about availability. It doesn't automatically tell us how many chips every facility contains or how much capacity an individual researcher can actually secure.
This is the kind of distinction I wish travelled with every dramatic infrastructure map. A dot looks wonderfully definite, especially when the surrounding area is blank. But I would want to know what the map includes, when it was collected and what an empty space means. I also want to know whether the resource is affordable and available to the people being discussed. A map can begin that conversation very effectively, provided we keep the questions that sit behind it.
The Electricity Connection Is Part of the AI Story
The International Energy Agency's October 2026 report on integrating data centres into electricity systems describes constraints involving network capacity, connection queues and permitting. It also explains that the effect on other customers' bills depends on how infrastructure costs are allocated and how existing capacity is used. A large computing facility has to fit into an actual electricity system. That makes the location consequential in ways a photograph of a server aisle won't show. It also makes me curious about the equipment on my own desk. I’d try a plug-in electricity monitor to see what those devices draw while I’m using them, because I like having an actual measurement to attach to something that’s usually invisible.
Water adds another layer. The IEA notes trade-offs between energy and water efficiency in cooling, as well as water use upstream in power generation. The details depend on the equipment and its setting, so I would be wary of attaching one universal water figure to every AI interaction. What interests me is whether a project explains its local demands and how they will be managed. The benefits of a new facility and the costs around it deserve to be considered together.
Renting Compute Can Open a Door
A cloud service can give a team access to equipment it could never sensibly purchase and maintain itself. The OECD's October 2026 policy brief describes domestic facilities, public computing resources and cloud services as parts of the wider landscape, with different approaches fitting different national needs. It also discusses international and regional access arrangements. There is room for several approaches to coexist, including buying capacity abroad, sharing resources and investing in appropriate infrastructure at home.
I think renting deserves a more useful conversation than simply treating it as defeat. Access can make a project possible, which matters enormously to the people trying to do the work. At the same time, I would want to understand the price, allocation process, technical support and options if the arrangement changes. Those are fairly ordinary questions about depending on a service. They become especially important when that service supports research or an institution's longer-term plans.
Downloadable Models Help, With Their Own Requirements
The availability of model weights can let other teams study, adapt or deploy a model under its licence, rather than only interacting with a provider's hosted interface. The terms and the available materials differ between projects, so open weights and a completely open development process shouldn't be assumed to mean the same thing. BigScience's BLOOM, released in 2022, is one example with a published model card, training information and a RAIL licence. It's a historical example of collaborative access, rather than my recommendation for a particular project today.
There is still work between downloading files and having something useful. Hardware, memory, software and maintenance all matter, and the appropriate setup depends on the task. IBM Research describes approaches such as reducing model precision to make inference more efficient, while also explaining the importance of the software and hardware working together. I like the possibilities here, especially for carefully chosen tasks. A model that someone can actually run and maintain may be much more useful to them than an impressive model they can only admire from a distance.
Language Deserves Attention of Its Own
BLOOM's model card lists training data in 46 natural languages and 13 programming languages, which is a useful counterexample to the idea that every accessible model is built around the same narrow set of languages. The same card warns about errors, stereotypes and uneven representation of viewpoints. Including a language in training data is a beginning; it doesn't establish equal performance across every language, dialect, subject or setting where somebody might want to use the model.
This is one of the parts of the discussion I care about most. I would want the people who know the language and the intended setting involved in deciding what good performance means. Can the system handle the actual wording, references and tasks that matter to them? Can they report a problem and influence what changes? A domestic address for the hardware won't answer those questions by itself. Useful participation also includes deciding what gets tested and whose assessment carries weight.
Shared Access Is Already Taking Several Forms
India's Press Information Bureau reported that 237 projects had accessed subsidised AI computing by August 2026 under the IndiaAI effort. The government's account describes shared compute intended to support researchers, start-ups and other innovators. That is evidence of an access programme and reported use, rather than proof that every applicant faces the same conditions or that all infrastructure barriers have disappeared. Still, it gives us a concrete example of an approach beyond expecting every research team to purchase its own machines.
Europe's EuroHPC AI Factories also provide access routes. Its Playground mode is aimed at smaller allocations and entry-level users, including eligible European start-ups and smaller businesses, with onboarding support. Applications still undergo eligibility and technical assessment. I like that support appears alongside access in the description, because getting permission to use unfamiliar equipment is only part of getting useful work done. Programmes like these make the conversation more interesting than a simple split between countries with everything and countries with nothing.
The Cold War Comparison Only Takes Us So Far
I understand the appeal of calling this an AI Cold War. The phrase immediately suggests competition, strategic resources and concern about dependence. It also encourages us to search for two sides and a scoreboard, which can make the rest of the story harder to see. Researchers, companies, public institutions and the people who use their work can have very different priorities within the same country. Those differences are part of the story, and each deserves a place in the conversation.
The OECD's 2026 policy brief makes a practical point that I keep returning to: compute plans should begin with local needs, and approaches can include public provision, grants, procurement and shared access. That leaves room to ask what somebody is actually trying to achieve. A useful research resource, a dependable public service and a bid to train a frontier model are different ambitions. Their success deserves to be assessed against the work they are meant to support.
What I Want to See Behind the Announcement
When I read about another ambitious AI infrastructure project, I want the ordinary details alongside the ambition. Is the equipment available now or still planned? Who is eligible to use it, how much access can they obtain, and is there help for people who are new to the system? I'd also want to know how running costs and upgrades will be supported. These questions make an announcement easier to assess without requiring everybody reading it to become a semiconductor specialist first.
I occassionally wish the practical side received the same attention as the enormous headline numbers. There's a difference between a promise that looks spectacular and a resource someone can reliably use for the next stage of their work. The people maintaining systems, preparing data, teaching new users and checking results belong in this story too. I find that wider picture more interesting than another ranking in which the only thing we learn is who spent the most.
A Wider Chance to Do Useful Work
The concentration of powerful computing resources is a serious part of the AI story, and I think it deserves attention without turning a dated country count into a verdict on everybody else's future. Following access, ownership and the physical infrastructure gives us much better questions to ask. Following actual projects and the people using shared resources gives us ways to see whether the answers are improving.
What I would like to see is more room for useful work to come from people with different priorities, languages and resources. That might involve shared facilities, a carefully chosen model, a research collaboration or better support for the equipment already available. The part that keeps me interested is what somebody can do once the resources become usable. Behind all the chips and policy documents, that's the future I want to hear more about.
