Capital concentration, grid fragility, premium power markets, and the emergence of fusion finance

‍

Over the past three years, AI capital has not diversified. It has concentrated. Capital has flowed into GPUs, hyperscale data centres, foundation models, and the firms capable of financing them at scale. What looks like a software boom has quietly become an infrastructure build-out measured in gigawatts, decades, and balance sheets.

‍

This is no longer just an AI story. It is an energy story. And increasingly, it is a sovereignty story.

‍

The Hidden Constraint Behind AI Scale

By the mid-2020s, AI workloads crossed a threshold. Training and inference moved from megawatts to gigawatts. Frontier data centres now rival cities in electricity demand. In the United States alone, data centres consumed 183 terawatt hours of electricity in 2024, representing more than 4 per cent of the country's total consumption.¹ This is roughly equivalent to Pakistan's entire annual demand. The International Energy Agency projects US data centre consumption will grow to 426 terawatt hours by 2030, a 133 per cent increase.

‍

The scale is staggering. PJM Interconnection is the largest regional grid operator in the United States, coordinating wholesale electricity across 13 states and serving 65 million people from Illinois to New Jersey.² Its territory includes Northern Virginia, which hosts the world's largest concentration of data centre infrastructure. PJM forecasts peak load growth of 32 gigawatts from 2024 to 2030. Nearly all of that is from data centres.

‍

The prevailing assumption was simple: if demand rises, capital will build more supply. That assumption is breaking.

‍

Electric grids do not scale like software. In the PJM region, the timeline from interconnection application to commercial operation has risen from less than two years in 2008 to over eight years in 2025.³ Ireland, where data centres now consume 22 per cent of the nation's electricity (up from 5 per cent in 2015),⁴ imposed a de facto moratorium on new data centre grid connections in Dublin from 2021 until December 2024.⁵ The Irish regulator now requires any new data centre to install on-site generation or battery systems capable of meeting its full electricity demand.

‍

Water is finite too. A typical large data centre consumes 3 to 5 million gallons of water daily, equivalent to the needs of a town of 50,000 residents.⁶ More than 160 new AI data centres have been built across the United States in the past three years in areas with scarce water resources. In Arizona, Google negotiated water rates of USD 6.08 per 1,000 gallons while residents paid USD 10.80.

‍

The industry is discovering a hard truth: compute scales faster than energy. This is the energy wall.

‍

Capital Concentration Is Structural, Not Accidental

AI funding is not just large. It is narrow. Four companies (Amazon Web Services, Google, Meta, and Microsoft) currently control 42 per cent of US data centre capacity.⁷ AWS alone plans to quadruple its current 3 gigawatt capacity to nearly 12 gigawatts.

‍

In the first six months of 2024, Microsoft, Amazon, Google, and Meta collectively spent more than USD 100 billion on AI and broader cloud infrastructure.⁸ The scale dwarfs public investment alternatives. The European Union's EuroHPC initiative, the largest government-backed programme, represents roughly 2 per cent of that figure.

‍

This concentration is not accidental. It is how capital behaves when infrastructure costs explode. But it creates systemic exposure.

‍

Gigawatt-scale data centres require tens of billions in upfront capital expenditure, multi-decade power commitments, and continuous reinvestment in short-lived hardware. GPUs depreciate in years, while energy infrastructure must be financed over decades. The result is visible in consumer bills: PJM's July 2024 capacity auction increased by USD 7.3 billion, an 82 per cent rise, driven almost entirely by data centre demand.⁹ The Natural Resources Defense Council estimates that consumers in the PJM region will pay an extra USD 163 billion through 2033 as new data centres exceed available power supply.¹⁰

‍

When power becomes scarce or politically constrained, optionality disappears quickly.

‍

When AI Stops Being Software and Becomes Governed Infrastructure

As AI industrialises, it changes category. What was once governed as software is now governed as infrastructure.

‍

Boards and governments increasingly face the same underlying question: who controls the inputs? At scale, AI depends on three assets that are no longer neutral or globally fungible: data, compute, and the jurisdiction in which the infrastructure operates.

‍

If an organisation does not control the jurisdiction governing its AI inputs, it does not fully control its AI strategy. Legal authority, not technical architecture, becomes the ultimate lever.

‍

This is the point at which AI stops being borderless in practice, even if it remains so in theory.

‍

The Global Race for Sovereign AI

Sovereign AI is often discussed as a response to data privacy or national security concerns. That framing understates what is happening.

‍

The numbers tell the story. The United States has allocated USD 470.9 billion to AI in 2025. China has committed USD 119.3 billion.¹¹ Saudi Arabia's HUMAIN entity, a subsidiary of the Public Investment Fund, is investing to create up to 500 megawatts of AI computing capacity over the next five years, with an 18,000 GPU supercomputer as the first phase. The UAE's MGX consortium, which includes BlackRock, Microsoft, and Nvidia, is acquiring Aligned Data Centers in the United States for approximately USD 40 billion, securing over 5 gigawatts of capacity.

‍

Across Europe, a wave of sovereign large language model projects is underway: France's Mistral, Switzerland's Apertus, Spain's Alia, Portugal's Amalia, and the Netherlands' GPT-NL.¹² Norway's Telenor launched a sovereign AI factory in late 2024 running on 100 per cent renewable energy.

‍

True AI sovereignty is not primarily about models. It is about resilience under stress. At scale, sovereign compute requires firm domestic electricity, predictable pricing over decades, insulation from geopolitical leverage, and political legitimacy for industrial energy use.

‍

Once AI underpins defence, healthcare, logistics, and government services, electricity stops being a commodity. It becomes a national input.

‍

This is where many sovereign AI strategies quietly struggle. Renewables help, but intermittency pushes reliance onto storage, backup generation, and grid complexity. Fission helps, but siting, waste, and public acceptance constrain speed. Fossil fuel stopgaps conflict with climate commitments and long-term independence.

‍

The question becomes unavoidable: how do you power AI at scale without locking in fragility, emissions, or dependency?

‍

The Energy Wall as Selection Mechanism

Much commentary warns of an imminent collapse when AI "hits the energy wall." That framing is misleading.

‍

There is no single moment when AI stops. What actually happens is selection. Some regions become unviable for large-scale AI. Ireland's four-year moratorium demonstrated this. Power-secure jurisdictions gain disproportionate advantage. On-site and dedicated generation become default assumptions. Compute becomes valuable only when paired with energy certainty.

‍

This is not collapse. It is reallocation. And reallocation reshapes winners and losers.

‍

The World Economic Forum estimates that extreme heat, drought, and other climate hazards could drive cumulative annual costs at data centres globally up by USD 81 billion by 2035, rising to USD 168 billion by 2065. Total cumulative losses could potentially reach USD 3.3 trillion.¹³ Around one quarter of today's facilities, and nearly one third of those under construction, are in regions projected to face greater water scarcity by 2050.

‍

Premium Power: Risk Pricing, Not ESG Theatre

A quiet shift is already underway in power markets. Large AI buyers are increasingly willing to pay above-market rates for long-term, carbon-free, firm electricity.

‍

The evidence is unambiguous. In September 2024, Microsoft signed a 20-year power purchase agreement to restart Three Mile Island's Unit 1 reactor in Pennsylvania, securing 835 megawatts to power its data centres. Meta announced a 20-year agreement with Constellation Energy for 1.1 gigawatts from the Clinton Clean Energy Center in Illinois. Amazon committed to purchasing 1.9 gigawatts through 2042 from Talen Energy's Susquehanna nuclear plant in Pennsylvania.¹⁴

‍

Google signed agreements with Kairos Power for 500 megawatts of small modular reactor capacity by 2030, and committed early-stage capital to Elementl Power for three United States reactor sites totalling 1.8 gigawatts. Amazon partnered with Energy Northwest for four advanced small modular reactors with a combined capacity of up to 960 megawatts, and invested USD 500 million in X-Energy's 320 megawatt small modular reactor project.¹⁵

‍

This is not ESG theatre. It is risk pricing.

‍

For AI operators, the cost of outages, regulatory intervention, carbon non-compliance, or stalled expansion far exceeds the marginal cost of power. In this context, a premium power purchase agreement is not a subsidy. It is insurance.

‍

AI is repricing electricity not by kilowatt-hour, but by certainty.

‍

Fusion Meets Its Market

Fusion energy has long been framed as a distant solution. Elegant physics. Long timelines. Uncertain markets.

‍

That framing no longer fits the buyer that AI has created.

‍

The fusion industry is now at a pivotal moment. Private fusion companies have raised USD 9.77 billion as of mid-2025.¹⁶ The US Department of Energy's October 2025 Fusion Science and Technology Roadmap targets commercial fusion power on the grid by the mid-2030s. MIT modelling projects fusion generation rising from 2 terawatt hours in 2035 to 375 terawatt hours in 2050.

‍

The commercial signals are tangible. Type One Energy selected a retired coal plant in Tennessee for its Infinity Two project, repurposing existing grid infrastructure for its stellarator-based fusion technology. Helion Energy signed the first fusion power purchase agreement with Microsoft for at least 50 megawatts, with a subsequent agreement with Nucor. Commonwealth Fusion Systems, the most heavily funded company at over USD 2 billion, signed a power purchase agreement with Google and announced its first commercial plant near Richmond, Virginia, in collaboration with Dominion Energy, also targeting the early 2030s.¹⁷

‍

In October 2025, Google DeepMind and Commonwealth Fusion Systems announced a research partnership to use AI to accelerate fusion development. The US Department of Energy released its fusion roadmap the same day, a coordinated signal of the convergence between AI demand and fusion supply.

‍

In a compute-constrained, sovereignty-aware world, fusion competes on different terms than traditional energy sources. Dispatchable, always-on power matters more than spot pricing. Land footprint and water intensity matter as much as cost. Political durability matters as much as efficiency. Reliability over decades matters more than marginal economics.

‍

Fusion is not competing with intermittent renewables on today's wholesale curves. It is competing with grid fragility and strategic dependence.

‍

The First-of-a-Kind Financing Shift

‍The critical unlock is not technological alone. It is financial.

‍

A viable structure for first-of-a-kind (FOAK) fusion plants is emerging. Strategic offtakers with large, non-discretionary power needs (hyperscalers, steel manufacturers, industrial consumers) commit to long-term contracts at premium pricing. That contracted revenue underwrites early capital risk. Utilities and infrastructure operators provide grid integration, regulatory cover, and financing credibility. Capital flows based on contracted cash flows, not speculative price exposure.

‍

This is how airports, LNG terminals, pipelines, and defence infrastructure are financed. The Microsoft-Helion power purchase agreement, the Google-Commonwealth Fusion Systems partnership, and the Amazon-Dominion SMR agreements are not public relations exercises. They are the financial architecture of energy transition.

‍

AI has quietly pulled energy into the same financing category.

‍

But the first-of-a-kind financing structure only holds when the capital architecture is designed correctly from the outset. Equity sequencing, covenant design, government co-investment alignment, and institutional narrative must be engineered as a system rather than assembled reactively. This is precisely where infrastructure-scale energy projects face their most acute and least visible risk.

‍

The Real Bottleneck Ahead

By the late 2020s, AI will not be limited by models, chips, or capital availability. BloombergNEF forecasts US data centre power demand will rise from almost 35 gigawatts in 2024 to 78 gigawatts by 2035, more than doubling.¹⁸ Between 2024 and 2028, the share of US electricity going to data centres may triple, from 4.4 per cent to 12 per cent.

‍

The constraints will be power generation capacity, grid resilience, water and cooling availability, and political permission to build.

‍

The winners will not be those who scale compute the fastest. They will be those who align compute, energy, and sovereignty most effectively.

‍

Conclusion

AI has not failed. It has industrialised.

‍

The era of treating energy as a background assumption is ending. The era of treating compute as critical infrastructure has begun.

‍

Sovereign AI and the energy reckoning are not separate debates. They are two sides of the same transition: from software optimism to physical reality.

‍

In that world, clean, firm, dispatchable power moves from the edge of the AI conversation to its centre. Not because it is fashionable. But because physics, politics, and scale demand it.

‍

The question is no longer whether energy will constrain AI. It is which jurisdictions, technologies, and financing structures will break through the wall, and which will be left behind.

‍

The capital architecture behind these decisions, including the governance frameworks, sequencing systems, and institutional narratives that make energy infrastructure financeable at scale, is being designed now, before the first institutional term sheet. For developers, infrastructure funds, and sovereign platforms alike, that design work is not optional.

‍

The wall is real. The window is open. The architecture determines who passes through it.

‍

About the Author

Carly Martin is the Founder of Elenara Capital, a specialist Capital Systems Architecture firm that designs capital stacks, governance frameworks, and sequencing systems for sovereign AI and energy infrastructure platforms approaching institutional capital raises. Elenara Capital operates upstream of transactions, working with energy developers, data centre platforms, fusion companies, and infrastructure funds before they engage their bank mandate.

To discuss a mandate or request a copy of the Capital Architecture Checklist for Infrastructure Fund Due Diligence, contact: carly.martin@elenaracapital.com.au

‍

References

Sources are grouped thematically. Full citation details available on request.

Energy demand, grid stress, and interconnection delays

¹  International Energy Agency. Energy and AI. Paris: IEA Publications, April 2025. See also: Pew Research Center, "What we know about energy use at US data centers amid the AI boom," October 2025.

²  PJM Interconnection. Generation Interconnection Fact Sheet. June 2025. / Rocky Mountain Institute. "PJM's Speed to Power Problem and How to Fix It." November 2025.

³  Rocky Mountain Institute. "PJM's Speed to Power Problem and How to Fix It." November 2025.

Grid regulatory intervention and water constraints

⁴  Central Statistics Office Ireland. Data Centres Metered Electricity Consumption 2024. June 2025.

⁵  Commission for Regulation of Utilities (Ireland). New Connections Policy for Large Energy Users: Decision Paper CRU/2025/30. December 2024.

⁶  Brookings Institution. "AI, data centers, and water." November 2025. / Environmental Law Institute. "AI's Cooling Problem." 2025. / University of Tulsa. "Data centers draining resources in water-stressed communities." July 2024.

Capital concentration and consumer cost exposure

⁷  BloombergNEF. Power for AI: Easier Said Than Built. July 2025.

⁸  Stanford University Human-Centered Artificial Intelligence. AI Index Report 2025. Stanford, CA: HAI, 2025.

⁹  Monitoring Analytics, LLC. State of the Market Report for PJM. October 2025.

¹⁰  Natural Resources Defense Council. "Data Center Growth and PJM Consumer Costs." December 2025.

Sovereign AI investment and global strategy

¹¹  Stanford HAI and OECD AI Policy Observatory. Global AI Investment Tracker 2025. / TRENDS Research and Advisory. "Funding the Future: Global Investment Strategies in AI." 2025.

¹²  Euronews. "Europe is trying to write a new sovereign AI map." December 2025. / European Commission. Digital Decade Policy Programme 2030. Brussels: EC, 2024.

Climate exposure at data centre infrastructure

¹³  World Economic Forum. "The $3.3 trillion question: Can data centres take the heat?" October 2025. / MSCI. "When AI Meets Water Scarcity." October 2025.

Premium power procurement and nuclear-backed PPAs

¹⁴  Constellation Energy. Press Release, September 2024. / Corporate disclosures: Amazon Web Services, Google, Meta, Microsoft. SEC Filings, 2024-2025. See also: Trellis Group, "Amazon, Google, Meta and Microsoft go nuclear," June 2025.

¹⁵  IEEE Spectrum. "Big Tech Embraces Nuclear Power to Fuel AI and Data Centers." December 2024. / Amazon Web Services. "AWS to invest in small modular reactor development." Press Release, October 2024.

Fusion energy: investment, roadmap, and commercial deployment

¹⁶  Fusion Industry Association. 2025 Global Fusion Industry Report. July 2025. / US Department of Energy. Fusion Science and Technology Roadmap. October 2025.

¹⁷  Helion Energy and Commonwealth Fusion Systems. Corporate announcements and PPAs, 2023-2025. / Type One Energy. "Type One Energy Selects Tennessee Site for Infinity Two Project." Press Release, 2024. / World Economic Forum. "How AI can help get fusion from lab to energy grid by the 2030s." December 2025.

Future power demand projections

¹⁸  BloombergNEF. Power for AI: Easier Said Than Built. July 2025. / MIT Technology Review. "We did the math on AI's energy footprint." September 2025.

‍

Disclaimer

This article is intended for wholesale investors as defined under the Corporations Act 2001 (Cth) and is not intended for retail investors. The information provided herein is for general informational purposes only and does not constitute financial, investment, or professional advice.

‍

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of Kings Gate Capital Partners. While every effort has been made to ensure the accuracy of the information, Kings Gate Capital Partners makes no representations or warranties, express or implied, as to the completeness, accuracy, reliability, suitability, or availability of the information contained in this article. Any reliance you place on such information is therefore strictly at your own risk.

‍

Investing involves risk, including the potential loss of principal. Past performance is not indicative of future results. Before making any investment decision, you should seek independent financial, legal, and tax advice tailored to your specific circumstances.

‍

This article may contain forward-looking statements that are subject to risks and uncertainties. Actual results may differ materially from those expressed or implied in such statements. Kings Gate Capital Partners disclaims any obligation to update or revise any forward-looking statements to reflect new information or future events.

‍

For more detailed information contact us directly.

‍

‍