AI is no longer competing for talent. It is competing for megawatts

Artificial intelligence has entered a less visible but far more decisive phase: physical constraint. In 2023 and 2024, executive debate focused on models, talent, use cases and productivity. In 2026, the strategic question is shifting: who can secure power, compute capacity, cooling, permits, grid access and operational resilience before competitors do? AI is no longer only a CIO or Chief Data Officer agenda. It is becoming a capital allocation, location, regulatory risk and cost-structure decision. For CEOs, funds, family offices and mid-market companies, the implication is material: adopting AI is not enough; leaders must understand who controls the infrastructure that makes AI scalable.

The core thesis is straightforward: competitive advantage in AI is moving from the algorithm to guaranteed access to electricity and compute capacity. Companies that treat AI as software will buy compute inflation; companies that treat it as critical infrastructure can capture productivity, resilience and a valuation premium.

Executive summary

• Power is becoming a strategic AI variable. The International Energy Agency expects data-centre electricity consumption to double by 2030, while AI-focused data centres could triple their consumption, following 17% growth in 2025. ([iea.org](https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions?utm_source=openai))
• The United States is the leading test case. The IEA reports that the US accounted for 45% of global data-centre electricity consumption in 2024 and that data centres may represent nearly half of US electricity demand growth through 2030. ([iea.org](https://www.iea.org/reports/energy-and-ai/executive-summary%C2%A0?utm_source=openai))
• The bottleneck is no longer only GPUs; it is time-to-power. Goldman Sachs projects US data-centre power demand could more than double from 31 GW in 2025 to 66 GW in 2027. ([goldmansachs.com](https://www.goldmansachs.com/insights/articles/us-data-center-power-demand-projected-to-double-by-2027?utm_source=openai))
• The risk is moving into prices, permits and social acceptance. In PJM, one of the largest US power markets, the independent market monitor attributed $6.5 billion, or 40%, of costs in a recent capacity auction to data-centre load. ([utilitydive.com](https://www.utilitydive.com/news/data-centers-pjm-capacity-auction/808951/?utm_source=openai))
• For investors, diligence is changing. AI-intensive businesses should be assessed not only by their technology roadmap, but by exposure to compute inflation, hyperscaler dependency, cloud contracts, data location, critical workloads, latency, business continuity and pricing power.
• For the mid-market, the opportunity is not to build infrastructure but to buy optionality. Smarter cloud contracts, hybrid architecture, inference optimisation, AI consumption governance and vendors with credible power access can become real sources of margin protection.

Why it matters now

1. AI has moved from a digital promise to an industrial demand curve

The early generative AI narrative was largely intangible: knowledge work, automation, assistants, copilots, new products. Scale has exposed a more physical reality. AI requires firm power, long-life assets, advanced cooling, electrical equipment, land, permits, fibre, in some designs water, and power contracts measured over 10, 15 or 20 years.

The difference versus previous technology cycles is that demand is not distributed evenly. It concentrates around power corridors, fibre nodes, permitting-friendly jurisdictions and utility systems able to absorb large loads. That concentration can create location advantage, but also congestion, community opposition and political risk. NERC’s long-term reliability assessment highlights resource adequacy concerns across North America and links part of the pressure to data centres and other emerging large loads. ([nerc.com](https://www.nerc.com/globalassets/our-work/assessments/nerc_ltra_2025.pdf?utm_source=openai))

2. Markets are starting to value the invisible infrastructure layer

Until recently, investors asked about AI adoption, expected productivity and exposure to foundation models. They now need to ask about the layer beneath it. Does the company have cloud contracts that protect against cost escalation? Can workloads move across providers? Which critical processes will depend on real-time inference? How much margin is lost if the cost per query, token, GPU-hour or storage unit rises by 20%? What happens if a provider limits capacity during peak demand?

The answer will vary by sector. In vertical software, professional services, insurance, banking, healthcare, logistics and retail, AI can become a structural variable cost. In manufacturing, defence, energy, pharmaceuticals and engineering, it can become a strategic capability linked to simulation, design, predictive maintenance and automation. In both cases, the point is the same: the cost of intelligence is starting to enter gross margin.

3. Power pressure creates a new valuation premium

The premium will not be limited to utilities and data-centre operators. It will also accrue to companies that can demonstrate three attributes: disciplined AI usage, resilient compute contracts and flexible technology architecture. In M&A, this changes the due diligence conversation. A company whose growth depends on AI but lacks control over inference costs may have lower earnings quality than headline growth suggests. Conversely, a company that has reduced consumption per process, redesigned models, compressed prompts, used smaller models where sufficient and negotiated multi-cloud capacity may deserve an operational premium.

McKinsey has noted that time-to-power is a critical factor for data-centre operators and that constraints exist across power, electrical infrastructure and specialised labour. ([mckinsey.com](https://www.mckinsey.com/industries/private-capital/our-insights/how-data-centers-and-the-energy-sector-can-sate-ais-hunger-for-power?trk=article-ssr-frontend-pulse_publishing-image-block&utm_source=openai)) For an investment committee, the translation is clear: AI infrastructure is no longer an IT line item; it is a value-creation assumption that must be tested.

Implications for CEOs, investors and leadership teams

For CEOs: AI needs its own P&L

Most companies still manage AI as a portfolio of initiatives. That is no longer sufficient. AI needs a clear economic account: use cases, compute consumption, unit cost, savings generated, revenue enabled, dependency risks, security and scalability. The CEO should require every material initiative to move from pilot logic to margin logic.

Consider a plausible example: a distribution company deploys AI agents for customer service, demand forecasting and local pricing. The pilot shows savings. But as it scales, inference consumption grows with every interaction, the cloud contract is renegotiated on less favourable terms and general-purpose models are used for tasks that a cheaper specialised model could handle. Productivity exists, but captured margin is lower than expected. The answer is not to slow AI; it is to industrialise it.

For CFOs: AI cost should be budgeted as digital energy

The CFO should treat compute as a critical input. In some companies, AI spend will remain modest for now; in others, it will grow silently inside SaaS, cloud and productivity licences. The risk is that AI inflation appears across multiple budget lines rather than as a consolidated exposure.

The control question is not how much we spend on AI. It is how much it costs to produce one useful unit of intelligence: an approved recommendation, a processed invoice, a resolved call, a validated line of code, an audited claim or a completed simulation. Without that metric, the committee sees activity, not value.

For investors: technology diligence must include indirect power diligence

Funds assessing data-intensive assets should review supplier dependency, elasticity of compute demand, contractual clauses, model architecture and EBITDA sensitivity to inference cost. In software, BPO, digital health, fintech, cybersecurity and analytics assets, this can be decisive.

Diligence should also examine supplier geographic concentration. If a critical part of the service depends on cloud regions exposed to power congestion or regulatory constraints, operational risk is not theoretical. Goldman Sachs estimates that data centres could rise from 4.1% of US summer peak power demand in 2025 to 8.5% in 2027. ([goldmansachs.com](https://www.goldmansachs.com/insights/articles/us-data-center-power-demand-projected-to-double-by-2027?utm_source=openai)) That pressure can flow into pricing, availability and capacity allocation priorities.

For family offices: infrastructure, energy and picks-and-shovels are back at the centre

The thesis does not require picking the winning foundation model. Capital can target the enabling layer: firm generation, storage, electrical equipment, liquid cooling, fibre, efficiency software, energy management, engineering services, industrial land with grid access and companies that reduce inference cost.

But the simple version of the thesis is dangerous. Not every power-related asset becomes a winner because AI consumes electricity. Permitting, interconnection, counterparty quality, contract duration, merchant exposure, social acceptance and technological obsolescence matter. The right question is not whether AI uses power; it is who can capture scarcity rents without assuming disproportionate risk.

For the mid-market: advantage sits in architecture, not size

A mid-market company does not need to compete with hyperscalers for megawatts. It does need to avoid being trapped in an expensive architecture. That means selecting use cases with clear returns, using smaller or specialised models where possible, caching responses, reducing unnecessary calls, automating quality evaluation, negotiating data portability and maintaining supplier alternatives.

In practical terms, AI efficiency can become an operating capability. Just as lean manufacturing reduced physical waste, lean AI will reduce computational waste: oversized prompts, overpowered models, duplicated tools, poor reuse of knowledge and pilots that consume budget without scaling.

Priority matrix

0-90 days

• Build an executive inventory of AI initiatives, related cloud spend and processes where inference will become recurring.
• Identify the 10 use cases with the highest potential consumption and calculate unit cost per business outcome, not per token or licence.
• Review cloud and SaaS contracts for escalation clauses, lock-in, capacity limits, data ownership and portability.
• Set an architecture principle: do not use large models when a smaller model, rules engine, semantic search or traditional automation can solve the problem.
• Include AI cost in 2026-2027 budgets as a cross-functional exposure, not as experimental innovation spend.

3-12 months

• Design a layered architecture: foundation models for complex tasks, specialised models for repetitive processes and classical automation for deterministic workflows.
• Negotiate capacity commitments, cost transparency, multi-region options and reasonable exit rights with cloud providers.
• Integrate AI efficiency metrics into management dashboards: cost per resolution, cost per prediction, reuse rate, latency, quality and net savings.
• Assess partnerships with vendors that can demonstrate robust access to power and credible continuity plans.
• Include compute inflation scenarios in financial planning, pricing and project valuation.

12-24 months

• Redesign high-volume processes to minimise unnecessary inference and maximise reuse of corporate knowledge.
• Build internal AI FinOps capabilities: consumption governance, model optimisation, observability and cost allocation by business unit.
• Incorporate AI and infrastructure diligence into acquisitions, carve-outs and value-creation plans.
• Explore strategic agreements with technology providers, regional data centres or sector platforms to secure capacity.
• Recalibrate the location strategy for critical operations, taking into account latency, regulation, energy resilience and data sovereignty.

Signals to watch

• Capacity auctions and wholesale power prices. Moves in markets such as PJM can foreshadow cost inflation in data-centre-dense regions. ([utilitydive.com](https://www.utilitydive.com/news/data-centers-pjm-capacity-auction/808951/?utm_source=openai))
• Interconnection timelines and electrical equipment lead times. Delays in transformers, substations and grid connections can constrain new capacity and lift cloud service prices.
• Nuclear, firm renewable and onsite generation deals by hyperscalers. Amazon, Microsoft, Google and Meta have explored firm power arrangements to support AI growth, indicating that the market increasingly views energy access as a strategic asset. ([aboutamazon.com](https://www.aboutamazon.com/news/sustainability/amazon-nuclear-energy-smr-plans?utm_source=openai))
• Local opposition to new data centres. Social acceptance can become as relevant as grid capacity when communities perceive pressure on bills, water or land use.
• Cloud and SaaS pricing changes. AI may appear as higher licence fees, additional consumption charges, capacity premiums or functionality limits.
• Model efficiency. Progress in small models, specialised chips, quantisation and inference optimisation can reduce unit cost and reshape winners.
• Regulation on resilience and cost allocation. If regulators require large loads to fund dedicated infrastructure, the economics of data centres and digital services will change.

Dextra recommendations

1. Move AI into the capital allocation agenda. The committee should decide where AI creates true advantage, what structural consumption it implies and what return threshold it must meet.
2. Build a compute exposure map. Identify the processes, vendors, geographies, contracts and business units most sensitive to AI availability and price.
3. Implement AI FinOps before scaling. Measure cost per outcome, eliminate redundant consumption and assign spend to accountable business owners.
4. Design for optionality. Avoid captive architectures; require portability, open standards and separation between data, models and applications where feasible.
5. Redefine technology due diligence. In M&A, review not only the stack and cybersecurity, but also inference economics, contractual scalability and dependence on external capacity.
6. Adopt a proportional model policy. Use the smallest model that meets the required standard for quality, risk and traceability.
7. Link AI to pricing. If AI creates customer value but also adds variable cost, the commercial model should reflect that through packages, limits, usage tiers or dynamic pricing.
8. Evaluate vendors by physical resilience. A digital SLA is not enough; leaders must understand regions, redundancy, energy access, continuity and capacity allocation priorities.
9. Prepare stress scenarios. Simulate 20%-50% increases in compute cost, capacity constraints or regional outages.
10. Turn AI efficiency into a value narrative. For investors and buyers, demonstrating net productivity and consumption discipline can become a differentiating advantage.

Closing

The first AI wave rewarded speed: test, launch, automate. The second will reward discipline: secure capacity, control costs, design architecture and convert intelligence into margin. In that transition, many companies will discover that they are not competing only for data or talent, but for access to scarce physical infrastructure. The sentence that should enter boardroom discussions is uncomfortable but accurate: AI will scale only as far as its scarcest megawatt allows.