Infrastructure for All: Bridging the AI Divide

The New York Times recently reported that Sam Altman toured a construction site in Texas that, when finished, will be larger than New York’s Central Park. The site buzzed with the grind of earthmovers, the clang of steel, and the low rumble of generators—a discordant symphony of ambition. It’s not a theme park or a housing development. It’s a data centre—a $60 billion temple to computing power, complete with its own natural gas plant, built to serve the demands of artificial intelligence.

At the same time, in a quiet university building in Argentina, computer science professor Nicolás Wolovick keeps his country’s most advanced AI infrastructure running from a converted classroom. Cables snake between ageing GPUs and second-hand servers. "We are losing," he told The New York Times, describing a sense of global exclusion.

Adam Satariano argues that we’re witnessing what might be the emergence of a new fault line in the global economy: the divide between those with AI compute, and those without it.

The Compute Divide: A Strategic Frontier

The report cited, published in June 2025, outlines the yawning global disparity in AI infrastructure. Out of nearly 200 countries, only 32 host advanced AI data centres. Over 150 nations are essentially shut out. Unsurprisingly, the US and China dominate, with Europe trailing and regions like Africa and Latin America largely absent. What’s at stake? Scientific progress, economic development, national security, and technological sovereignty.

Just as the oil-rich nations of the 20th century shaped geopolitics, countries with access to large-scale compute may now hold disproportionate influence over the development of AI, drug discovery, climate modelling, and advanced defence systems. Compute is becoming a strategic resource—and, to namecheck a concept on the rise, a matter of cloud sovereignty. As control over infrastructure and platforms concentrates, questions emerge about who sets the terms of access, who benefits from innovation, and who is left dependent.

Indeed, for most countries, building their own AI infrastructure is unrealistic. The capital costs are immense, the hardware scarce, and the operational demands extreme. Even if a government managed to fund construction, access to state-of-the-art chips like Nvidia’s H100s remains a geopolitical bottleneck.

However, it’s a false dichotomy to assume that nations must either build or be left behind. There is a third option: to borrow compute.

Enter the Cloud

Cloud platforms like AWS, Google Cloud, and Microsoft Azure offer pay-as-you-go access to the world’s most powerful machine learning infrastructure. For a researcher in Nairobi or a startup in Lima, the cloud transforms what was once a hardware problem into a software one. Instead of laying fibre and pouring concrete, you can spin up an EC2 instance.

While we shouldn't be under the illusion that this model erases the compute divide, it does blur the edges to a certain extent. For one, it enables access without ownership—no need to raise venture capital or build a data centre to train a model. It also allows experimentation without buckets of capital, letting startups prototype solutions or researchers run simulations that would otherwise be out of reach. And finally, it decouples innovation from geography by allowing anyone with an internet connection to build, train, and deploy advanced models—regardless of whether they live near a hyperscale data centre or a rural satellite dish.

AI in the Cloud: Democratising by Design?

But what does that actually look like in practice? Among the hyperscalers, AWS has arguably taken some of the most deliberate steps toward democratising access:

• Advanced Hardware Access: Through services like Amazon EC2 P5 instances, users anywhere can run models on Nvidia H100 GPUs. AWS also offers its own custom silicon—Trainium and Inferentia—optimised for AI training and inference.
• AI Without the AI Team: Amazon SageMaker simplifies the process of building, training, and deploying machine learning models. Amazon Bedrock allows developers to access foundational models via API without managing infrastructure.
• Global Reach: AWS continues to expand its infrastructure into underserved regions, with new Local Zones and data centres in Africa, South America, and Southeast Asia.
• Ecosystem Support: Programs like AWS Activate for startups and Cloud Credits for Research offer funding, mentorship, and compute grants to those who need them most.

Of course, AWS isn’t alone in this space. Both Google Cloud and Microsoft Azure also offer extensive AI tooling, global infrastructure, and academic support programmes aimed at widening access. Each provider brings its own flavour—whether it’s Google’s integration with open-source ML frameworks or Microsoft’s alignment with enterprise ecosystems—but the broader trend is certain: cloud platforms are becoming the gatekeepers of opportunity in the age of AI.

Caveats and Challenges

Whilst we’re considering the broader landscape, let’s also be clear: renting compute is not the same as sovereignty. Many countries still grapple with issues around data localisation, digital autonomy, and cloud affordability. Even within the cloud model, English-language dominance and biased training data create uneven terrain for non-Western users.

Moreover, cost remains a barrier. The promise of “AI for everyone” only holds if the pricing models work for nonprofits, universities, and small enterprises—not just well-funded startups.

A Call to Shared Stewardship

Still, the opportunity provided by the cloud is a real thing. If hyperscalers invest wisely—in regional infrastructure, transparent pricing, educational partnerships, and representative data sets—cloud platforms could be one of the keys to unlocking a more inclusive AI.

As Smart Africa’s Lacina Koné put it: "It’s not merely a hardware problem. It’s the sovereignty of our digital future."

If we want AI to benefit everyone, compute can’t remain a privilege. It must become a shared resource—a commons for the digital age. The cloud won’t solve the compute divide on its own. But it might just give us a fighting chance.