Sovereign AI Means Infrastructure

Sovereign AI means infrastructure.

I spend quite a bit of time on X these days. Some of that comes with being a founder. You want to know what people are building, what investors are thinking about, what engineers have discovered and, occasionally, what everyone has collectively decided to be angry about that morning.

X is a strange place to do this. I sometimes think of it as the closest thing we have to a global brain, except the brain has an algorithm deciding which neurons you are allowed to see. There is also the old problem of the spiral of silence. Some of the most important people aren’t posting, and some of the most important conversations are happening somewhere else entirely. What you see is therefore not the market so much as one noisy representation of it. Every so often, though, you come across a question worth following, and I recently found one about data centres in Africa.

The discussion started with a fairly simple observation: for all the talk about sovereign AI, there is still relatively little of the infrastructure required to run AI at scale across much of the continent. Nigeria was being used as one of the examples, which naturally caught my attention. The original claim was actually too broad. Nigeria does have data centres, and AWS operates a Local Zone in Lagos, although its only full African Region remains in Cape Town. The Lagos Local Zone also offers GPU-backed G4dn instances, so it would be wrong to describe Nigeria as having no AI compute infrastructure at all. The more accurate observation is that the depth and scale of that infrastructure remain very different from what you would find in established cloud regions.

Correcting the claim made the underlying question more interesting. Having data centres is not the same thing as having enough infrastructure to run AI at scale, particularly when you start talking about large inference workloads, training and the amount of compute increasingly required by both. I’ve been thinking about this because the conversation around sovereign AI tends to begin much further up the stack, with models, data and national ambition, when much of what determines whether any of that is possible sits underneath them.

Sovereign AI means infrastructure

I was asked about this in an interview with ThisDay in July, shortly after a strange episode involving Anthropic and the US government. At the time I was thinking mostly about sovereignty from the perspective of the organisations we work with: where their data is stored, where AI models process it, whether customer information leaves their environment and what happens when an AI system starts taking actions against a bank’s or government’s system of record.

The Anthropic episode extended that question beyond data. On 12 June, the US government imposed export controls on Claude Fable 5 and Mythos 5 that restricted access by foreign nationals, including foreign nationals physically inside the United States. Anthropic said it had no reliable way to verify nationality in real time, so it suspended access to both models for all users. The restrictions were lifted on 30 June, and access began returning the following day. The disruption lasted less than three weeks, but it demonstrated a dependency that is easy to ignore when everything is working normally: an organisation can build an important workflow around a frontier model and still have access to that model determined by a government thousands of miles away.

The obvious response is to say that countries should build their own AI, but that doesn’t get us very far. Building a sovereign model while renting all of the infrastructure required to run it doesn’t make you particularly sovereign. Keeping data inside your borders while having no meaningful control over the systems processing it solves only part of the problem. Even owning the data centre doesn’t remove your dependence on chips, networking equipment, software and supply chains that may originate elsewhere.

Sovereign AI means infrastructure. Simple.

Before a country can seriously discuss sovereign models, it needs compute. Compute needs data centres, and data centres need energy, cooling, fibre, land, engineers and enormous amounts of capital. The GPUs themselves sit inside a global supply chain that very few countries control. Once you follow the stack far enough down, sovereign AI starts looking as much like industrial policy as AI policy.

None of this means sovereignty requires a country to produce every component itself. Nigeria doesn’t need to manufacture every GPU it uses, any more than Britain needs to manufacture every server inside a London data centre. Brazil doesn’t need to build its own equivalent of AWS before it can exercise meaningful control over critical AI systems. The more useful question is how much control a country needs at each layer and which dependencies become strategically important when AI moves from generating marketing copy to processing payments, healthcare records, national identity data or government services.

Policy creates the market

Nigeria is useful here because some of these questions are already moving from discussion into policy. The Central Bank of Nigeria has directed financial institutions and payment participants to store and manage payment transaction data generated in Nigeria locally from January 2027. In my ThisDay interview, I argued that AI complicates what data residency means. It is no longer enough for a bank to know where its database sits. If an AI agent is reading a customer’s transaction history, reasoning over it, sending some of that context to a model and then taking an action inside the bank’s systems, the bank also needs to understand where that processing is happening.

A policy like this creates a market because the institutions affected by it still need somewhere to run their workloads. Existing local providers can expand, new companies can enter, banks can build more of their own infrastructure, and international cloud companies have a stronger economic reason to increase their local presence. Infrastructure doesn’t get built simply because it would be strategically useful to have it. Somebody has to be able to make money from building it, and one of the things policy can do is make previously uncertain demand much easier to see.

The size of the Nigerian corporate market makes this more than a theoretical demand signal. The Nigerian Exchange represents roughly $106–110 billion in listed equity value at recent exchange rates, spanning banks, telecommunications companies, industrial groups, insurers, consumer businesses and other companies that increasingly depend on software and data infrastructure. These organisations aren’t all going to become AI companies, nor do they need to, but many of the systems they rely on will increasingly contain AI somewhere inside them. Payments, fraud detection, customer service, collections, underwriting, identity verification, internal operations and software development are all becoming more compute-intensive as AI is introduced into the workflow.

You can see a more mature version of the same relationship in countries that already have deep cloud markets. Governments establish policy around sensitive data, public-sector procurement and critical infrastructure, while large pools of enterprise and government demand make it economically attractive for infrastructure providers to build locally. The resulting infrastructure makes those policies easier to implement because organisations actually have somewhere suitable to put their workloads.

For countries that don’t already have that infrastructure, the sequence is more difficult because policy can create demand before supply is ready. Nigeria’s recently introduced National Digital Cloud Policy appears to recognise this problem. It doesn’t deal only with where data should reside. It explicitly identifies investment in data centres, cloud infrastructure, connectivity and AI compute capacity as part of the country’s strategy, while attempting to position Nigeria as a hosting and processing hub for West Africa. Telling companies that certain workloads should remain locally is one part of the job. Creating the conditions under which somebody can provide that capacity competitively is another.

The energy and economics underneath sovereignty

Energy is one of the constraints that quickly appears when you move from talking about sovereign AI to actually building it. Conventional data centres already require significant energy, and GPU-dense AI facilities push those requirements considerably further. This isn’t peculiar to Nigeria. AI infrastructure is forcing countries with far more developed electricity systems to think again about generation, grid capacity, transmission and how quickly large new facilities can be connected. Countries where reliable energy was already expensive or constrained enter the same infrastructure race with an additional cost embedded in the economics from the beginning.

Capital creates another constraint. Building AI infrastructure at meaningful scale isn’t a normal startup capex decision. The current infrastructure build-out in the United States, Europe and the Middle East involves investments measured in billions of dollars and energy requirements measured in hundreds of megawatts or gigawatts. There is a legitimate debate about whether all of that capital will eventually produce sensible returns, but the scale gives some perspective to what we mean when we casually say a country should build its own AI infrastructure.

For an emerging market, these two constraints interact with policy in ways that can become uncomfortable. If a bank is required to run certain workloads domestically but comparable compute costs materially more locally than it does in a mature cloud region, the difference becomes a sovereignty premium whether anyone calls it that or not. A bank, government or critical-infrastructure provider may decide that the premium is worth paying. A startup operating with thin margins may reach a very different conclusion. Enough of that price difference across an economy and a policy intended to encourage domestic capability can inadvertently make local companies less competitive.

We encounter a version of this at BimpeAI. Enterprise customers increasingly want to understand where their data will sit and whether systems can be deployed on-premise or inside approved infrastructure. Those are reasonable requirements, but the engineering conversation eventually gets down to GPUs, capacity, concurrency and utilisation. If an organisation wants dedicated infrastructure rather than elastic compute from a global cloud, somebody has to pay for enough capacity to support the peaks as well as the periods when those GPUs are doing very little.

The infrastructure environment therefore becomes part of AI policy whether policymakers intend it to or not. Public-private partnerships, tax incentives, access to land and energy, government procurement, development finance and other incentives can change the economics for companies considering whether to build locally. Government doesn’t necessarily have to own the resulting infrastructure, and there are good reasons why much of it should remain private, but the market has to become attractive enough for somebody to make the investment.

This is also why I find the current American infrastructure build-out interesting. It is easy to look at the enormous sums being spent by technology companies and conclude that this is simply private capital chasing an AI opportunity. That is only part of the story. Those companies are building inside an environment shaped by energy policy, tax treatment, permitting, government procurement, capital markets and an enormous existing technology industry. The private investment is real, but so is the system around it.

What does it mean to control the stack?

Foreign investment complicates the sovereignty question because physical location and technological control aren’t the same thing. An American or Chinese company can build a large AI data centre in Nigeria, keep Nigerian data physically inside the country, employ Nigerian engineers, connect to Nigerian networks and operate under Nigerian law while important parts of the underlying technology remain dependent on software, hardware and supply chains controlled elsewhere.

That doesn’t make the infrastructure useless from a sovereignty perspective. Keeping sensitive workloads within a country’s jurisdiction is materially different from sending everything to another continent. It does mean that sovereignty is better understood in layers. A country can control its data without controlling its compute. It can control its compute without controlling the models running on it. It can host an open model locally while depending entirely on foreign hardware, or own the physical data centre while remaining exposed to export controls affecting replacement GPUs. Each arrangement gives the country control over something without making it technologically independent.

Complete independence isn’t a realistic objective anyway. Modern computing is the product of supply chains spread across countries and companies with different comparative advantages. The practical task is understanding which dependencies matter enough to reduce and which ones are economically sensible to retain. The Fable episode was interesting precisely because it exposed one of those dependencies. Access to a model can feel like a normal software subscription until a geopolitical decision changes the terms.

Nigeria makes these trade-offs unusually visible because the infrastructure is being built while the policy environment is changing. There is a large domestic economy, a substantial banking and payments industry, more than $100 billion of listed equity value and growing enterprise demand for AI, but the depth of hyperscale and AI compute infrastructure is still very different from what exists in London, Frankfurt or Northern Virginia. That creates an opportunity for local infrastructure companies, international providers and investors, but also a coordination problem. Infrastructure is expensive to build before demand is certain, while companies are reluctant to depend heavily on local infrastructure before enough capacity exists.

The shape of that problem changes from country to country. Brazil has the scale to support a much larger domestic infrastructure market. A smaller European country may reasonably decide that regional infrastructure provides sufficient sovereignty rather than trying to reproduce the entire stack domestically. Canada can have substantial data-centre capacity and still ask how much control physical location gives it when important parts of the infrastructure are owned by foreign companies. The answer depends on the size of the economy, its energy system, capital markets, geopolitical relationships and the workloads it considers critical.

Building sovereignty

When I spoke to ThisDay in July, I said that one of the biggest mistakes Nigeria could make in the AI era would be to treat artificial intelligence solely as something the country consumes. I still believe that, but I now think the builder conversation has to extend much further down the stack. Producing Nigerian models and Nigerian AI companies matters, but those companies eventually have to run somewhere, and the infrastructure they run on will determine some of what they can build, how cheaply they can build it and how much control they retain over it.

There is an entrepreneurial opportunity inside that problem. If policy continues to create demand for domestic infrastructure while AI increases the amount of compute required by banks, governments, telecommunications companies and other large organisations, the market won’t be served only by companies training models or building applications. There will be businesses around compute, data centres, orchestration, networking, energy, deployment and the less visible infrastructure required to make local AI economically viable. Some of those companies will be domestic, some will be international, and some of the most useful infrastructure will probably emerge from partnerships between the two.

Government has a role because many of the conditions determining whether those businesses can exist sit outside the control of an individual entrepreneur. This is why I keep coming back to policy before infrastructure. Policy can create a market, reduce uncertainty and make long-term capital easier to justify. Public-private partnerships can distribute some of the cost and risk of building infrastructure before the market fully matures. Government procurement can provide anchor demand. Energy policy can determine whether a data-centre project makes economic sense before the first GPU is installed. None of these replaces private enterprise, but they influence whether private enterprise has a market worth entering.

For entrepreneurs building in countries where the infrastructure is still thin, I think there is value in engaging this conversation much earlier. Waiting until every bank, government agency and large company suddenly needs domestic AI compute is probably the most expensive way to discover that the capacity isn’t there. Builders, infrastructure companies, investors and policymakers need to be talking while the market is still forming, because the decisions being made now about cloud infrastructure, energy, data residency and AI deployment will determine what is possible later.

Nigeria is the case study I know best because I am Nigerian and because these are questions we encounter while building BimpeAI, but I don’t think this is fundamentally a Nigerian question. Any country that expects AI to become part of its banking system, healthcare infrastructure, public services or critical industries will eventually have to decide what it wants to control and what it is comfortable depending on somebody else to provide.

Sovereign AI means infrastructure. The interesting question is not whether every country can own every layer of the stack. Very few can, and trying to do so would probably be economically wasteful. The question is which parts matter enough to control, which dependencies a country is willing to live with, and whether the policy, capital, energy and entrepreneurs required to build the rest are arriving quickly enough.