The physical bottlenecks of AI scaling in numbers
4.4%
projected share of global electricity demand consumed by data centres by 2035 (IEA).
50GW
contracted UK data centre demand vs. ~45 GW current peak grid capacity.
$600B+
projected cumulative big tech AI capital expenditure in 2026.
The infrastructure behind the intelligence
AI's evolution from early machine learning to today's agentic systems is driving an unprecedented surge in compute demand. While public discourse focuses heavily on foundation models and end-user applications, physical infrastructure has become the binding constraint on how fast AI can scale.
Understanding the AI stack
AI is a five-layer system. Every upper layer is constrained by the reliability and capacity of the foundation below it.
Energy layer
The foundational layer of the entire stack. Power availability and grid connection timelines are now the single biggest bottleneck to scaling compute capacity.
Chips & systems layer
The physical hardware layer. It faces persistent shortages in high-performance GPUs and advanced cooling components alongside surging enterprise demand.
Infrastructure, models & applications
The layers most visible to the public and end users, built on top of physical foundations that are increasingly strained by density requirements.
Where capital and investment are moving
Scandinavia & Spain
With Ireland and the Netherlands hitting capacity ceilings, investment is shifting to these markets, which offer available power and supportive regulation.
The UK
Seeing renewed activity but remains constrained by high energy costs relative to its European neighbours.
Tier 1 cities
Investors are concentrating on primary markets where alternative tenants exist if AI demand fails to materialise as expected. Secondary markets face greater scepticism.
Capital is forming, but constrained
Demand for AI infrastructure capital is outpacing the funding mechanisms built to support it.
The funding gap
Hyperscalers are investing $600 to $800 billion annually, yet the entire global private infrastructure equity market raised only around $200 billion last year. Big tech spent over $400 billion on AI infrastructure in 2025, with cumulative AI capital expenditure projected to exceed $600 billion in 2026.
Institutional caution
Pension funds and insurers are interested but cautious. Many lack mature frameworks for assessing AI-exposed data centre risk, and rating agencies are still developing methodologies.
New structures emerging
Some investors are launching long-duration fund structures, recognising the sector's long-term capital needs. Capital recycling, monetising stabilised assets while retaining operations, will be critical for scaling.
Structural inertia
Organisations are fundamentally structured around human processing speeds. Realising AI's value requires completely redesigning end-to-end processes rather than simply automating existing steps.
Verification at scale
High-stakes decisions, including capital allocation, risk management, and safety-critical systems, require deeper verification and governance than current generative outputs can reliably guarantee.
Workforce disruption
As senior experts retire and junior tasks become automated, organisations face a critical dilemma: how to develop the next generation of expert human judgment when junior entry points disappear.
A Business School perspective on AI infrastructure
The implications of AI are strategic, not just technical. Realising its full value depends on how it is integrated into decision-making, organisational design and management systems. London Business School's Data Science and AI Initiative brings an evidence-based, whole-system lens to this challenge. With 1,300+ executives trained and 44 faculty engaged, it is one of Europe's most active business school AI institutes.
Key questions
How will energy availability shape the geography of AI investment over the next decade?
What needs to change for institutional capital to scale into AI infrastructure at the pace demand requires?
If junior roles are disappearing, how will the next generation of experts develop the judgement AI cannot yet provide?
AI infrastructure: research & perspectives
