--- title: "AI Infrastructure" date: 2026-07-12 description: AI infrastructure now runs from the power station to the protocol, and energy access decides what an organisation can do with AI. author: Mario Thomas canonical: https://mariothomas.com/briefings/ai-infrastructure/ --- ## Start here Two short reads before the deeper work: what AI infrastructure covers now, and the order to take this briefing in. ### What Your AI Strategy Is Standing On For most of the organisations I work with, infrastructure has meant something a technology team looks after: servers, networks, a cloud contract. AI has quietly broken that definition. The infrastructure beneath enterprise AI now runs from power stations and grid connections, through the choice of model, down to the protocols that connect agents to the systems where your competitive advantage lives. Goldman Sachs put the headline conclusion plainly in 2025: the most pressing bottleneck for AI progress is not capital but the power needed to fuel it. Most of the Boards I meet get this wrong in two ways. The first is altitude: treating infrastructure as an operational matter that sits beneath Board attention, when energy access has become a determinant of AI capability and UK organisations pay industrial electricity prices roughly four times those of their US competitors (Social Market Foundation 2025). The second is direction: reading every efficiency announcement as a signal to spend less, when a century and a half of evidence, from Jevons' steam engines to cloud economics, says cheaper compute expands demand rather than reducing it. The position these articles take is that infrastructure is a capability question before it is a cost question. Hyperscalers building their own generation are becoming grid actors, not just customers. Model choice is an infrastructure decision, because the wrong model class turns a solvable problem into an expensive one. And the plumbing that connects agents to proprietary data, now converging on a single standard, is being decided in technology teams today, often by default. Read this briefing and you should come away able to make one judgement well: when the next infrastructure headline lands, a capacity deal, an efficiency breakthrough, a nuclear partnership, you can say what it changes for your organisation. What is newly possible, which dependencies deepen, and what governance needs to exist before the decision gets made beneath the Board. ### From the Grid to the Plumbing The core sequence in [the articles](#core-reading) is ordered deliberately. Start with A New Grid Actor, which makes the central claim: AI infrastructure operators are becoming energy infrastructure operators, and the governance implications reach every Board with AI ambitions. Then The Headroom Argument, which corrects the most common misreading on this subject, that efficiency progress means the compute will not be needed. It will. The third and fourth pieces supply the evidence and the stakes. The Goldman Sachs analysis shows the institutional money reaching the same conclusion, that power rather than capital is the binding constraint on AI progress. The UK sovereignty piece is where this thinking started: the fourfold energy cost disadvantage, the sovereignty paradox it creates, and the questions UK Boards should be asking about strategic autonomy. Finish the core with the MCP explainer, which takes infrastructure below the data centre to the protocol layer, where agents meet your proprietary systems. The further pieces go deeper. The LLM selection piece and The Return of Traditional AI form a pair on model choice, and on the discipline of asking whether generative AI is the right tool at all. The Inference Migration connects the two ends of the briefing, showing how always-on agents multiply both the inference demand and the energy stakes. The Trilemma piece, When a Frontier Model Vanishes, takes the dependency up the stack to the model itself: a capability withdrawn overnight by a directive nobody in the organisation could appeal to, which is an infrastructure risk the Board now carries. Behind the Trilemma piece sit the original Trilemma article, which makes the argument at the level of national regimes, trust against speed against control, and The Year AI Grew Up, which records the year energy reached the Board and those regimes hardened. After the reading, [Remake](#remake-assets) holds the mechanisms that turn the thinking into apparatus you can use in the boardroom, and [the questions](#faqs) collect the ones I am asked most often, each answered in a paragraph. If you have 20 minutes, read the first two articles and the questions. The rest will be here when the subject reaches your agenda, and on current evidence it will. ## Core reading The core sequence runs from the grid to the protocol layer; the further pieces cover model choice, the shift to always-on agents, and what happens when a frontier model is withdrawn. 1. [A New Grid Actor: AI Infrastructure Is Becoming Energy Infrastructure](https://mariothomas.com/blog/ai-energy-infrastructure-grid-actor/) (9 minute read, 14 December 2025): AI infrastructure operators building their own generation become grid actors rather than consumers, and that changes energy economics, nowhere more sharply than the UK. Podcast edition: 14 minute listen. 2. [The Headroom Argument: Why AI Efficiency Means More Compute, Not Less](https://mariothomas.com/blog/headroom-argument-ai-efficiency/) (9 minute read, 10 May 2026): Architectural efficiency expands AI compute demand rather than reducing it. Three forces converge on more inference. Boards should read efficiency news as capability, not cost. Podcast edition: 12 minute listen. 3. [UK AI Energy Constraints: From Niche Concern to Investment Banking Focus](https://mariothomas.com/blog/goldman-sachs-energy-diplomacy/) (10 minute read, 14 September 2025): When Goldman Sachs says power, not capital, is AI's bottleneck, energy stops being a niche concern. UK Boards pay four times their competitors' energy costs. 4. [Beyond Regulatory Uncertainty: Thoughts on the UK's AI Sovereignty Challenge](https://mariothomas.com/blog/uk-ai-sovereignty-energy/) (11 minute read, 22 July 2025): Individual AI training clusters will soon need more electricity than whole nations generate. The UK's AI sovereignty ambition meets its energy reality. 5. [MCP Explained: The Agent Infrastructure Standard Boards Need to Understand](https://mariothomas.com/blog/model-context-protocol-explainer/) (11 minute read, 1 March 2026): An agent that sees only the public internet is an expensive search engine. MCP connects agents to the proprietary systems that constitute advantage. Podcast edition: 16 minute listen. ## Further reading - [Selecting your enterprise LLM: Moving beyond the hype to make the right choice](https://mariothomas.com/blog/choosing-the-right-llm/) (10 minute read, 3 January 2025): With well over a hundred language models available, choosing one is a question of fit, not headlines: match the model to the task. - [The Return of Traditional AI: Organisations Are Rethinking Their LLM-First Strategies](https://mariothomas.com/blog/return-of-traditional-ai/) (8 minute read, 21 December 2025): Forty-two percent of companies abandoned most of their AI initiatives this year, often because generative AI was applied to problems traditional methods solve better. Podcast edition: 13 minute listen. - [The Inference Migration: What Consumer Agents Mean for Enterprise AI's Next Phase](https://mariothomas.com/blog/inference-migration/) (12 minute read, 8 February 2026): Consumers are already paying for always-on AI agents. That consumer-to-enterprise pipeline is the one ChatGPT ran, and it is running again. Podcast edition: 17 minute listen. - [The AI Sovereignty Trilemma: When a Frontier Model Vanishes and Reality Bites](https://mariothomas.com/blog/ai-sovereignty-trilemma-reality/) (10 minute read, 14 June 2026): The visible cost of sovereignty deters Boards. The hidden cost of the convenient alternative was never shown, and that is the cost 12 June presented. Podcast edition: 11 minute listen. - [AI Sovereignty: A Board's Guide to Navigating Conflicting National Agendas](https://mariothomas.com/blog/ai-sovereignty-board-trilemma/) (15 minute read, 7 September 2025): AI governance is fragmenting into incompatible systems: Europe's transparency, America's scale, China's control. Boards can no longer serve all three; they have to choose. - [The Year AI Grew Up: Five Inflections That Changed the Strategic Calculus in 2025](https://mariothomas.com/blog/the-year-ai-grew-up/) (14 minute read, 30 December 2025): In 2025 Boards stopped asking what AI could do and started treating it as strategic infrastructure investment. Five connected inflections drove that shift. Podcast edition: 19 minute listen. ## Remake The mechanisms beneath the thinking: the model, diagnostic, methodology, and principle from the Remake Library that turn this briefing into apparatus a Board can use. - **Model: Eight Disciplines of AI**. The taxonomy of eight AI disciplines: Machine Learning, Deep Learning, Natural Language Processing, Computer Vision, Reinforcement Learning, Robotic Process Automation, Cognitive Computing, and Generative AI. [Remake Library](https://mariothomas.com/remake/library/#eight-disciplines-of-ai) - **Diagnostic: Five Pillars of AI Capability**. The five capability domains of an AI capability model that cut across every level of maturity: Governance and Accountability, Technical Infrastructure, Operational Excellence, Value Realisation and Lifecycle Management, and People, Culture and Adoption. [Remake Library](https://mariothomas.com/remake/library/five-pillars/) - **Methodology: AI Business Case**. The integrated decision framework that crystallises across an ADAPT engagement rather than at a single stage: strategic alignment established at Align, cost and readiness evidenced at Diagnose, value shaped at Advise, and execution designed at Plan. [Remake Library](https://mariothomas.com/remake/library/#ai-business-case) - **Principle: Well-Advised**. The framework of five strategic priorities, Innovation, Customer Value, Operational Excellence, Responsible Transformation, and Revenue, used to ensure AI investments create balanced value rather than narrow cost reduction. [Remake Library](https://mariothomas.com/remake/library/well-advised/) ## Questions The questions I am asked most often on this subject, each answered from the work in this briefing. ### Why is energy suddenly a Board matter rather than a procurement line? Because energy access now determines AI capability. UK organisations pay industrial electricity prices roughly four times those of their US competitors (Social Market Foundation 2025), and Goldman Sachs' "Powering the AI Era" (2025) concludes that power, not capital, is the most pressing bottleneck for AI progress. When investment banks dedicate serious research to power constraints, energy sovereignty has moved from policy concern to strategic imperative. I set out the UK position in [the sovereignty piece](/blog/uk-ai-sovereignty-energy/) and the institutional confirmation in [the Goldman Sachs analysis](/blog/goldman-sachs-energy-diplomacy/). ### AI keeps getting more efficient. Should that not reduce what we spend on it? The instinct is understandable and the evidence points the other way. Every prior era of computing converted efficiency gains into more demand, not less, and AI inference is behaving the same way: cheaper inference releases workloads that cost and latency had been holding back. [The Headroom Argument](/blog/headroom-argument-ai-efficiency/) sets out the three forces behind this, and why Boards that read efficiency news as a budget signal risk under-resourcing the opportunity their competitors fund. ### Hyperscalers are building their own power stations. Why does that matter to us? Because operators building behind-the-meter generation are transitioning from energy consumers into grid actors, with surplus capacity to sell and growing influence over national grids. In a system the size of the UK's, that carries both opportunity and concentration risk, including questions of foreign ownership of critical national infrastructure. [A New Grid Actor](/blog/ai-energy-infrastructure-grid-actor/) examines the transition and the governance frameworks Boards should establish before it forces answers upon them. ### Does every AI use case need a large language model? No, and the data points the same way. S&P Global (2025) found that 42% of companies abandoned most of their AI initiatives, up from 17% the year before. On my reading, much of that is generative AI applied to problems better solved by traditional machine learning or deterministic automation. The first question is still whether the use case is a good fit for generative AI at all. [Selecting your enterprise LLM](/blog/choosing-the-right-llm/) and [The Return of Traditional AI](/blog/return-of-traditional-ai/) make the case for matching capability to problem, with hybrid architectures where each discipline does what it does best. ### Why should the Board care about a technical protocol like MCP? Model Context Protocol is the standard that connects AI agents to the proprietary data and systems where competitive advantage lives, and the major enterprise vendors have already converged on it. The decision is being made now, often by default, inside technology teams. What agents can see, and what they do with what they see, is a governance question with a technology dimension. [MCP Explained](/blog/model-context-protocol-explainer/) covers the standard and the questions to ask; [The Inference Migration](/blog/inference-migration/) covers what always-on agents change. ### What happens to our processes if a frontier model is withdrawn overnight? They stop, and nothing degrades first. On 12 June 2026 a national-security directive forced a provider to switch off two deployed frontier models for every customer, and the organisations cut off were never its target. The AI Sovereignty Trilemma model holds that frontier capability bought cheaply surrenders sovereign control, and that surrender has moved up the stack: sovereign hosting, sovereign data, and sovereign infrastructure do not save a capability governed elsewhere. Model availability belongs on the Board's risk register with a named owner and a fallback that survives the specific event, jurisdictional as well as commercial. [When a Frontier Model Vanishes](/blog/ai-sovereignty-trilemma-reality/) sets out the questions. ### Should we secure long-term power agreements the way hyperscalers do? For most organisations, no, and the reason is scale, not timidity. On my count in [A New Grid Actor](/blog/ai-energy-infrastructure-grid-actor/), hyperscaler nuclear commitments now exceed 8GW, which is operators becoming grid actors, not customers buying certainty. Goldman Sachs' "Powering the AI Era" (2025) shows why the comparison fails: the binding constraint is power, not capital, and no financing structure manufactures it. The Board's questions are different: where the organisation's energy exposure sits, whether its providers have secured their capacity, and, if on-site generation is in play, where it crosses from backup to grid export and into a different regulatory regime. [The Goldman Sachs analysis](/blog/goldman-sachs-energy-diplomacy/) sets out the constraint. ### How should we read a capacity or nuclear announcement from our providers? As a capability signal first and a cost signal second. Capacity is being built for demand the providers can already see, so the first question is what becomes newly possible, then which dependencies deepen, then what governance needs to exist before anyone acts. New nuclear, SMRs above all, arrives in the mid-2030s, so the gap is being filled by gas and the emissions accountability travels with the contract. A provider building its own generation is becoming a grid actor, with concentration and ownership questions a grid the UK's size feels sooner than America's. [The Headroom Argument](/blog/headroom-argument-ai-efficiency/) and [A New Grid Actor](/blog/ai-energy-infrastructure-grid-actor/) supply the reading. ## References The external research this briefing draws on most heavily, from the institutions and analysts doing the serious work on AI's physical foundations. - **Goldman Sachs** (2 July 2025): [Powering the AI Era](https://www.goldmansachs.com/what-we-do/investment-banking/insights/articles/powering-the-ai-era). Powering the AI Era: the analysis concluding that power, not capital, is the binding constraint on AI progress. - **Financial Times** (8 December 2025): [The power crunch threatening America’s AI ambitions](https://ig.ft.com/ai-power/). The interactive analysis behind the projected 19GW US power shortfall by 2028. - **International Energy Agency** (12 November 2025): [World Energy Outlook 2025](https://www.iea.org/reports/world-energy-outlook-2025). World Energy Outlook 2025: global data centre demand rising 160% by 2030, and the SMR timeline. - **Wärtsilä and AVK** (15 September 2025): [Data centre dispatchable capacity: a major opportunity for Europe’s energy transition](https://www.wartsila.com/docs/default-source/energy-docs/technology-products/white-papers/data-centre-dispatchable-capacity-avk-wartsila_white-paper_2025.pdf). The white paper on data centre microgrids as dispatchable capacity for Europe's constrained grids. - **US Department of Energy** (11 September 2023): [DOE Releases New Report on Pathways to Commercial Liftoff for Virtual Power Plants](https://www.energy.gov/lpo/articles/doe-releases-new-report-pathways-commercial-liftoff-virtual-power-plants). Pathways to commercial liftoff for virtual power plants, targeting 80-160GW of capacity by 2030. - **UK Government** (8 April 2025): [AI Energy Council to ensure UK’s energy infrastructure ready for AI revolution](https://www.gov.uk/government/news/ai-energy-council-to-ensure-uks-energy-infrastructure-ready-for-ai-revolution). The AI Energy Council, established to ready UK energy infrastructure for AI demand. - **Office for National Statistics** (18 May 2025): [The impact of higher energy costs on UK businesses](https://www.ons.gov.uk/economy/economicoutputandproductivity/output/articles/theimpactofhigherenergycostsonukbusinesses/2021to2024). The official data behind the fourfold UK energy cost disadvantage against US competitors. - **Deloitte** (18 November 2025): [TMT Predictions 2026: The AI gap narrows but persists](https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions.html). TMT Predictions 2026: inference on track to account for around two-thirds of all AI computing power. - **Leopold Aschenbrenner** (29 May 2024): [IIIa. Racing to the Trillion-Dollar Cluster](https://situational-awareness.ai/racing-to-the-trillion-dollar-cluster/). Situational Awareness: the trillion-dollar cluster projections that first framed energy as the constraint. - **BCG** (September 2025): [The Widening AI Value Gap](https://media-publications.bcg.com/The-Widening-AI-Value-Gap-October-2025.pdf). The Widening AI Value Gap: only 5% of firms achieve AI value at scale, and why capability matching matters. - **Social Market Foundation** (9 February 2025): [High energy prices threaten UK AI world-leading status, as data centres can’t keep up with AI ambitions](https://www.smf.co.uk/high-energy-prices-threaten-uk-ai-world-leading-status-as-data-centres-cant-keep-up-with-ai-ambitions/). The 2025 analysis behind the fourfold gap: UK industrial electricity prices against US rates, the constraint the energy articles in this briefing turn on. - **Quartz** (16 October 2024): [Amazon is joining Google and Microsoft in going big on nuclear power](https://qz.com/amazon-google-microsoft-nuclear-power-ai-data-centers-1851673653). Amazon, Google, and Microsoft's 2024 nuclear power agreements for AI data centres. - **S&P Global** (30 May 2025): [Generative AI experiences rapid adoption, but with mixed outcomes – Highlights from VotE: AI & Machine Learning](https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning). S&P Global's 2025 finding that 42% of companies abandoned most of their AI initiatives, up from 17% the year before. - **Talen Energy Corporation** (11 June 2025): [Talen Energy Expands Nuclear Energy Relationship with Amazon](https://ir.talenenergy.com/news-releases/news-release-details/talen-energy-expands-nuclear-energy-relationship-amazon). Talen Energy's 2025 expansion of its nuclear supply relationship with Amazon, one of the hyperscaler power deals the grid-actor piece follows. ## The ideas beneath this briefing Ideas I've named and matured writing about AI Infrastructure: what each one means, and where it started. - **Sovereignty Paradox**: The bind in which high domestic energy costs push UK businesses toward foreign-hosted AI services, creating strategic dependencies that undermine national AI-first ambitions even as they reduce immediate operational cost. [Read more](https://mariothomas.com/blog/uk-ai-sovereignty-energy/) - **Maturity Arbitrage**: The strategic advantage of combining decades-proven AI disciplines like machine learning and computer vision with newer generative capabilities to balance risk while still capturing innovation at the edges. [Read more](https://mariothomas.com/blog/return-of-traditional-ai/) - **Sovereign Control**: The ability to keep an AI capability running on terms the organisation sets, rather than terms a distant provider, or a government, can revise without consultation; more than data residency. [Read more](https://mariothomas.com/blog/ai-sovereignty-trilemma-reality/) - **Adaptive Localisation**: A strategy of running different AI approaches in different markets, tuned to each market's regulatory, cultural, or competitive conditions, deliberately trading consistency for regional advantage. For a Board, choosing this stance means accepting real complexity costs and being ready to answer why the organisation treats one market's rules differently to another's. [Read more](https://mariothomas.com/blog/ai-sovereignty-board-trilemma/) - **Cultural Sovereignty**: The often-overlooked dimension where American, European and Chinese AI cultures embody different worldviews, disruption, deliberation and harmony, shaping architecture, governance and stakeholder engagement beyond regulation. [Read more](https://mariothomas.com/blog/ai-sovereignty-board-trilemma/) All concepts: https://mariothomas.com/glossary/concepts/ ## More Board Briefings More complete resources on AI and emerging technology for the Boards that need the full picture. - [AI Risk](https://mariothomas.com/briefings/ai-risk/): The AI risks that bite are seldom on the register, and the bill for sovereignty shocks, readiness gaps, verification costs, and model risk arrives later. - [AI Regulation](https://mariothomas.com/briefings/ai-regulation/): AI regulation has split into incompatible regimes, and the Board's task is not compliance with each but a deliberate position across all of them. - [AI Transformation](https://mariothomas.com/briefings/ai-transformation/): Most organisations are stuck at the pilot stage; crossing the divide is a Board judgement about delegation, architecture, and governance, not a technology purchase. - [Data Strategy](https://mariothomas.com/briefings/data-strategy/): Every AI system runs on data the balance sheet cannot see, so Boards should govern, structure, and monetise it as the asset it is.