Energy-First AI

Finland's Mäntsälä data centre pays £28 ($35, €33) per megawatt hour (MWh) for carbon-free Nordic hydro. Virginia operators negotiate £67 ($85, €80) per MWh and still face four-year grid queues. That 58% cost gap is not a rounding error. It is the entire AI infrastructure story.

Last week's Apple-Google partnership crystallised this reality. Apple will pay approximately £790m ($1bn, €950m) annually for access to Google's 1.2 trillion parameter Gemini model. A company with £250bn in cash reserves concluded that building AI infrastructure was not worth the energy headache.

The 2020s will be remembered as the decade when data centres became power plants that happen to compute.

This week, we examine the graphics processing unit (GPU) operators who recognised this first.

The neocloud market reached £19bn ($24bn, €23bn) in 2025. According to Mordor Intelligence, it is projected to surge to £137bn ($174bn, €165bn) by 2030. That represents a 48% annual growth rate. Yet this headline obscures a structural shift in where that value accrues.

Power Cost by Region (Industrial Data Centre Rates)

Source: Green Mountain Data Center, Cushman Wakefield Global Data Center Report

A 100 MW facility in Norway costs £38m ($48m, €46m) less annually than the same facility in Virginia. Over a 10-year power purchase agreement, that gap reaches £380m ($483m, €456m). At 500 MW, the figure exceeds £1.9bn ($2.4bn, €2.3bn).

Operators who secured Nordic power agreements in 2023 now command cost advantages that compound annually. Power represents 30% to 40% of GPU cloud operating expenses. A 58% reduction in that line item transforms unit economics entirely.

CoreWeave exemplifies this transformation. The company began as an Ethereum mining operation in 2017. It pivoted to AI infrastructure as demand exploded. Today, it operates 250,000 NVIDIA GPUs across 33 data centres with approximately 470 MW of active IT power. Its valuation reached £47bn ($60bn, €56bn) following a March 2025 initial public offering (IPO).

The crucial insight: Microsoft signed £7.9bn ($10bn, €9.5bn) in contracts with CoreWeave through 2030. The hyperscaler that operates Azure needed external GPU capacity. It could not secure enough power internally.

Microsoft's total neocloud investment reached £26bn ($33bn, €31bn) across multiple providers. OpenAI signed contracts worth £17.6bn ($22.4bn, €21.2bn) with CoreWeave despite Microsoft backing. When companies defining AI infrastructure still lack sufficient capacity, the constraint is not compute. It is megawatts.

Three distinct approaches have emerged for energy-first AI infrastructure. Each offers quantifiable advantages.

Model 1: Stranded Asset Conversion (Crusoe Energy)

Crusoe demonstrates the gas-to-nuclear pathway. Originally capturing flare gas from oil fields to power crypto mining, the company divested those operations in March 2025. In October 2025, Crusoe partnered with Blue Energy to develop a 1.5 GW nuclear-powered campus in Port Victoria, Texas.

The arbitrage is precise:

Crusoe is raising £1.08bn ($1.375bn, €1.3bn) at a valuation exceeding £7.9bn ($10bn, €9.5bn). The funding is not speculative. It reflects operational track record and locked-in arbitrage.

Model 2: Nordic Renewable Arbitrage (Nebius)

Nebius demonstrates the green energy procurement strategy. Headquartered in Amsterdam, the company invested over £790m ($1bn, €950m) in European AI infrastructure by mid-2025. Its flagship Mäntsälä facility in Finland houses 60,000 GPUs. Annual revenue potential exceeds £790m at full utilisation.

Finland offers structural advantages beyond headline pricing:

Nebius also deployed capacity in Keflavik, Iceland, powered entirely by geothermal energy at approximately £27 ($34, €32) per MWh. This represents the lowest rate among major GPU cloud deployments globally. The company secured Ark Data Centres capacity in Surrey, UK, with 4,000 NVIDIA Blackwell Ultra GPUs.

Model 3: Efficiency Arbitrage (Lambda Labs)

Lambda Labs demonstrates the developer-focused efficiency strategy. With annualised revenue reaching £397m ($505m, €477m) by May 2025, Lambda offers H100 GPUs at roughly £1.96 ($2.49, €2.36) per hour. CoreWeave charges £3.34 ($4.25, €4.02) for comparable access.

That 41% price advantage attracts researchers and startups needing flexibility without multi-year commitments. Lambda achieves 61% gross margins through operational efficiency rather than energy arbitrage. Its virtualisation software partitions clusters of 16,000+ GPUs into 15-minute billing increments.

Traditional nuclear development requires 10 to 15 years from announcement to operation. Blue Energy claims its shipyard-manufactured approach reduces deployment to 36 months. The difference determines market capture.

Consider the trajectory:

Seven years from inception to market-defining scale.

A traditional utility planning equivalent 470 MW capacity today will not deliver power until 2033 at earliest. By then, the market structure will be locked.

Policy is catching up to market reality across multiple jurisdictions.

United Kingdom: AI Growth Zones launched to accelerate planning approvals for data centre developments. The forthcoming National Policy Statement EN-7 will expand regulatory scope. It will include small modular reactors (SMRs) and advanced modular reactors. British regulators explicitly acknowledge that AI sovereignty requires energy sovereignty.

South Korea: The 11th Basic Power Supply Plan, finalised in February 2025, adds 700 MW of SMR capacity by 2038 alongside two large-scale reactors. Korea Hydro and Nuclear Power partnered with Amazon and X-energy. Together they plan to deploy over 5 GW of advanced reactors by 2039, targeting £39bn ($50bn, €47bn) in investment.

Japan: Plans announced to relocate data pools near offshore wind and nuclear sites. Currently, 90% of data centres cluster in Tokyo-Osaka areas. Clean energy concentrates in Hokkaido and Kyushu. The geographic mismatch drives policy intervention.

The International Energy Agency (IEA) projects nuclear and renewables will provide nearly 60% of data centre electricity by 2030, up from 35% today. This is not speculation. This is demand meeting supply along the path of least resistance.

For nuclear operators, three immediate priorities emerge.

Priority 1: Contact GPU cloud procurement teams directly.

CoreWeave, Nebius, and Crusoe seek anchor power agreements in the 50 to 200 MW range with 10 to 15 year terms. Target contacts: Chief Infrastructure Officer, VP of Data Centre Operations, Head of Power Procurement. These conversations should happen before grid-based offtake negotiations.

Priority 2: Model behind-the-meter configurations for existing sites.

A 500 MW nuclear facility with 50 MW spare capacity represents:

Multiply across fleet. EDF, Constellation, KEPCO, and Rosatom each operate portfolios where aggregate incremental value exceeds £100m annually.

Priority 3: Engage regulators on co-location frameworks.

The UK's AI Growth Zones and Korea's integrated SMR-AI planning demonstrate templates that accelerate deployment. Propose pilot programmes. Regulators respond to specifics, not abstractions.

For infrastructure investors, the neocloud market offers exposure to AI growth through energy fundamentals. The £137bn projected 2030 market requires approximately £1.9tn ($2.4tn, €2.3tn) in cumulative capital expenditure. Energy-first operators capture disproportionate value because power access determines compute access.

The value chain is inverting. Traditional infrastructure investing assumed: land, then permits, then grid connection, then compute deployment. Energy-first investing assumes: power access, then compute deployment, then customer acquisition, then regulatory normalisation.

CoreWeave's 50%+ Model FLOPS Utilisation on Hopper-class GPUs runs roughly 20% higher than public benchmarks. At £3.34 per GPU-hour, this efficiency gap generates approximately £8,800 ($11,200, €10,600) additional revenue per GPU annually. Multiply by 250,000 GPUs. The operational advantage exceeds £2.2bn annually.

For nuclear operators specifically, the GPU cloud sector represents near-term anchor demand with superior economics to grid sales. Constellation, EDF, KEPCO, and CGN should evaluate direct data centre partnerships. Grid-based offtake agreements may never materialise at acceptable terms.

When Apple, with effectively unlimited capital, chose to pay Google £790m annually rather than build AI infrastructure, it validated a thesis: energy economics now dominate compute economics.

The GPU operators who recognised this first now command valuations that rival traditional utilities. They built around power rather than around chips. They transformed energy access into compute access.

The arbitrage window remains open. Norway costs £24 per MWh. Virginia costs £67 per MWh. A 100 MW facility captures £38m annually on the spread.

The operators who move fastest win.

Next week: 150 Sites. 6 Continents. What Press Releases Hide.

Stop typing reproductions and start vibing code. Wispr Flow captures your spoken debugging flow and turns it into structured bug reports, acceptance tests, and PR descriptions. Say a file name or variable out loud and Flow preserves it exactly, tags the correct file, and keeps inline code readable. Use voice to create Cursor and Warp prompts, call out a variable like user_id, and get copy you can paste straight into an issue or PR. The result is faster triage and fewer context gaps between engineers and QA. Learn how developers use voice-first workflows in our Vibe Coding article at wisprflow.ai. Try Wispr Flow for engineers.

Start flowing free