Sovereign AI: Why infrastructure is the starting point

In April 2025, the International Energy Agency released what it called the most comprehensive, data-driven global analysis to date on the relationship between artificial intelligence and energy. According to the IEA’s 2025 Energy and AI report, the world’s data centres are projected to consume 945 terawatt-hours (TWh) of electricity annually by 2030, which is more than double the 415 TWh recorded in 2024, and roughly equivalent to Japan’s entire current electricity demand. For financial institutions, that figure is not an environmental statistic, it is a signal about AI infrastructure decisions being made right now, which will shape operational costs, regulatory exposure, and competitive positioning for a decade.

Why has sovereign AI infrastructure become a strategic question?

Global investment in data centres has nearly doubled since 2022 and reached half a trillion dollars in 2024. IEA figures show a typical AI-focused facility already consumes as much electricity as 100,000 households. When a technology’s compute requirements begin to reshape national electricity grids, the infrastructure decisions of individual organisations aggregate into consequences that extend well beyond their own operations.

For instance, in the United States data centres are on course to account for nearly half of all electricity demand growth between now and 2030. For the financial sector, the question of where AI runs has always had operational implications. It now has regulatory, strategic, and systemic ones as well.

What does the cloud actually cost to run?

The phrase “cloud computing” was always a metaphor, and it has done a subtle disservice to strategic thinking. As researchers at MIT noted in January 2025, data centres are present in the physical world, with direct implications for the land, water, and energy systems of the regions where they are located. Cloud is not weightless. It is a specific collection of hardware, in a specific place, drawing from a specific electricity grid.

That specificity matters because grids are not uniform. Coal currently accounts for approximately 30% of the electricity used to power global data centres, with sharp regional variation: the coal share approaches 70% in China, while natural gas accounts for over 40% of data centre electricity supply in the United States. Moving to cloud is not, in itself, a move towards lower-carbon compute. The energy profile of that choice depends entirely on where the infrastructure sits.

How should the sovereign AI infrastructure decision actually be made?

The argument for on-premises AI is not ideological. It is conditional. There are specific operating circumstances under which local infrastructure is the more operationally efficient choice, and when efficiency improves, the environmental outcome follows from that efficiency, not the other way around.

Those conditions are:

Condition	Why it matters for operational efficiency
Constant high utilisation	On-premises hardware running GPUs at sustained capacity avoids the cooling overhead and transmission losses associated with shared cloud infrastructure. The efficiency argument holds when utilisation is consistently high, not intermittent.
Localised renewable energy	If an institution’s local grid draws predominantly from wind, solar or hydro, the carbon intensity of running on-premises compute can be significantly lower than a cloud data centre powered by coal or gas in a different region.
Avoiding data transmission	Sending large datasets to a remote cloud facility requires energy at every network hop. Local processing eliminates transmission losses entirely, a consideration that compounds at the scale of continuous, high-volume financial workloads.
Hardware longevity	Keeping and maintaining existing on-premises hardware is often preferable to frequent replacement. Manufacturing new hardware carries substantial environmental costs typically the largest share of its total lifetime impact

Research published in Nature Sustainability in November 2025, led by Cornell University‘s Process-Energy-Environmental Systems Engineering lab, found that smart localisation, grid decarbonisation, and operational efficiency, applied together, could reduce the carbon footprint of AI infrastructure by approximately 73% and its water use by 86% compared with worst-case scenarios. There is no single lever where location, energy source, and operational discipline must work in combination. None of the four conditions above makes a categorical claim for on-premises over cloud. Each one describes a context in which the efficiency calculation shifts.

Sovereign AI infrastructure: What are the risks in getting this wrong?

The infrastructure question carries two categories of risk. The first is operational: data centres clustering in regions with strained electricity grids face potential service disruptions. The IEA estimates that around 20% of planned data centre projects globally could face delays due to grid connection constraints where wait times in some advanced economies have already extended to four to eight years. For institutions running AI workloads that depend on continuous availability, infrastructure location is a resilience question, not only an efficiency one.

Regulatory exposure compounds this. Data sovereignty, the principle that data is subject to the laws of the jurisdiction where it is processed and stored, is hardening across multiple regulatory environments. For financial institutions, this is not a future consideration. In several jurisdictions it is already a compliance requirement. An institution that defaults to a global cloud provider without understanding where that provider’s infrastructure physically sits may find itself with a data sovereignty and sovereign AI compliance problem it did not anticipate.

What does the hardware picture look like in five years?

Neuromorphic computing, an approach to hardware design that mimics the structure of the human brain’s neural and synaptic architecture, offers a fundamentally different energy model. In neuromorphic systems, only the active computational elements consume power; the rest of the network remains idle, a structural departure from how GPUs operate continuously regardless of load. IBM’s research documentation notes that neuromorphic computing is progressing quickly but has not yet reached the maturity required for mainstream adoption. Swiss company FinalSpark has gone further still, building a neuroplatform powered by lab-grown human brain organoids, rented over the internet by researchers at 34 universities, with a stated goal of AI running on a tiny fraction of current energy requirements.

These architectures do not resolve the infrastructure question today. However, the organisations that have begun asking where their AI runs and what powers it will be better positioned to adapt as the hardware landscape shifts. The infrastructure commitments made in 2025 and 2026 will be difficult to reverse by 2030.

For more expert content on industry outlooks and innovation, subscribe to our newsletter or visit our Insights page.

Q&A: Key Questions on Sovereign AI