For years, the assumption around AI infrastructure was easy to accept. Serious compute would be built where hyperscale cloud, developer density, and capital were already concentrated: California, Seattle, London, and a small circle of established technology hubs.
There was a practical reason for that geography. Training and deploying AI at scale requires data centers, compute, networking capacity, energy, and advanced infrastructure to work together. OECD analysis notes that this has pushed AI firms toward services operated by the largest cloud computing suppliers. Over time, that dependence hardened into market concentration. In the third quarter of 2025, Synergy Research Group put Amazon, Microsoft, and Google’s combined share of global enterprise cloud infrastructure spending at 63 percent.
That logic now looks less durable. Compute is becoming more expensive, more power-intensive, and harder to access outside a small group of dominant providers. Builders are starting to confront questions that hyperscale cloud mostly let them ignore. Where will the power come from? Can chips be shipped to this jurisdiction? Whose laws apply to the data once it moves?
Those questions are getting answered in different places now, and most of them are not in Silicon Valley.
What Scarcity Teaches
In established cloud markets, the default answer to rising AI demand is to add more capacity through larger cloud contracts, denser data center buildout, and deeper dependence on the same centralized stack.
That answer is becoming harder to scale. Data centers consumed about 1.5 percent of the world’s electricity in 2024, enough to make energy one of the pressure points in AI infrastructure. The International Energy Agency expects that share to rise to just under 3 percent by 2030, making compute harder to treat as a hidden layer behind AI products.
In much of the developing world, that pressure was already the starting point. Builders there have rarely had the option of treating compute access, power, and distribution as someone else’s problem; they have had to design for it. The result is a quieter pattern that does not get much attention in Silicon Valley coverage: serious AI infrastructure is now being built in places where scarcity is treated as a design problem rather than an afterthought.
What This Looks Like in Practice
The pattern is most visible across four regions.
In India, Yotta Data Services runs Shakti Cloud on more than 16,000 NVIDIA H100 GPUs and is on track to roughly double that by the end of 2025. Over half of the compute behind the IndiaAI Mission — the government’s push to build indigenous foundation models — sits on Yotta’s hardware. In February 2026, the national multilingual platform BHASHINI moved off foreign hyperscalers and onto Shakti Cloud, picking up roughly 40 percent in performance along the way. BHASHINI runs real-time translation across 11 Indian languages at population scale; the people running it had decided that infrastructure they could not govern was the wrong place to put it.
Across Africa, Cassava Technologies, founded by Zimbabwean entrepreneur Strive Masiyiwa, is deploying 12,000 NVIDIA GPUs across data centers in South Africa, Egypt, Kenya, Morocco, and Nigeria. Cassava is the first NVIDIA Cloud Partner on the continent; before this buildout, NVIDIA estimated that roughly 80 of its GPUs were installed across the entire African continent. The constraint was not only compute pricing; it was the basic absence of advanced silicon. Cassava’s response is a pan-African network running on its own fibre backbone, designed so that African startups, researchers, and governments do not have to route through Europe or the United States to train and deploy AI.
In Brazil, the government’s SoberanIA project reserves 500 MW for a sovereign AI factory in Piauí, powered entirely by renewable energy, with Scala Data Centers as lead infrastructure partner. Brazil has committed to attracting up to $370 billion in data center investment over the next decade, tied to the REDATA program’s tax incentives for projects sourcing 100 percent renewable power. Roughly 65 percent of Brazilian data is still stored abroad. The wager is that abundant hydroelectric and solar power gives Brazil a kind of compute the U.S. and China have to work harder to build — clean by default, cheap by geography.
The United Arab Emirates is taking the most expensive route. Core42, part of the G42 group, sells inference capacity on a mix of NVIDIA and Qualcomm chips out of Abu Dhabi, and the country has committed jointly with the United States to a 10-square-mile, 5-gigawatt AI campus that should be partially operational by the end of the decade. The Emirati pitch is straightforward: countries that want sovereign AI but cannot build the underlying stack themselves can rent one from a friendly government. The Middle East Institute describes it as a deliberate strategy of vertical integration — owning the chips, the power, the data centers, and the foreign relationships in one piece.
These projects do not share a politics or an ownership model. What they share is a starting assumption that compute access, power, land, and chip supply are first-order design problems rather than externalities. That assumption produces different infrastructure.
Why Inference Changes the Map
Training large models still rewards dense clusters, large capital budgets, and access to advanced chips. That work is unlikely to leave the largest hyperscale facilities soon.
Inference is a different problem. Models are used continuously, by customers, devices, agents, and enterprise systems. McKinsey expects inference to overtake training in AI data centers by 2030, accounting for more than half of AI compute and roughly 30 to 40 percent of total data center demand.
Inference asks different questions than training does. Rather than where the largest cluster can be built, the questions become where compute should sit, how fast it can respond, how reliably workloads can be routed, and whose laws govern the data while it does so. Those questions have geographic answers that hyperscale concentration does not handle well, especially for the billions of people who do not live within easy latency of a U.S. or European data center.
The compute fabric that inference demand requires is broader than hyperscale cloud alone can provide. Distributed GPU capacity, regional inference clusters, sovereign clouds, and emerging neoclouds in places such as Mumbai, Nairobi, São Paulo, and Abu Dhabi are not substitutes for hyperscale. They are the layer hyperscale cannot serve on its own.
What This Means for the Map
The old map of AI infrastructure was drawn around places where cloud capacity was already concentrated. That map made sense when compute was treated as cheap and abundant.
The next map will look different. It will be drawn around places that learned to build when compute was costly and strategic, and where the question of who controls the stack was never theoretical. The companies and governments doing that work are not catching up with Silicon Valley; they arrived at the problem first, because they had to.
—-
Ilman Shazhaev is founder and CEO of Dizzaract, an AI infrastructure company headquartered in Abu Dhabi. He serves as a UN/UNODC expert panel member advising on AI applications in developing economies and has authored 46 scientific articles and 10 registered invention patents.
