When the Power Station Becomes the Site

Power stations were built to move large amounts of energy through constrained sites. AI campuses are now looking for exactly that combination: power, land, cooling options and an infrastructure workforce. That does not make every generation site a data center site. It does, however, change the order in which developers should screen markets.

The latest signal comes from Japan. On 1 October 2026, JERA announced a collaboration with RHAELM and Dell Technologies to develop AI infrastructure at the Chiba Thermal Power Station. The first phase is described as having 400 MW of power capacity, with operations targeted around 2028 and investment exceeding $15 billion across the wider programme. Reuters reported that Apollo is expected to support financing.

That 400 MW figure matters, but so does its unit. It is power capacity, not a disclosed IT load. Treating the two as interchangeable would overstate deliverable compute and hide the conversion losses, redundancy, cooling loads and phase design that sit between them. This distinction is central to power station data center development.

Power station data center development changes the first question

The conventional search starts with a market, then land, then a grid application. A generation-linked model can reverse that sequence: start with an energy node and test whether it can support a digital campus.

This is attractive because grid queues and network reinforcement increasingly dominate delivery schedules. The International Energy Agency expects global data center electricity consumption to approach 950 TWh in 2030, while warning that grids and critical equipment can take longer to expand than data centers themselves. A power-station site may reduce one form of uncertainty by co-locating generation and demand.

But it does not remove the development problem. It moves the problem. Instead of asking only when the utility connection will arrive, the developer must also ask whether generation is firm enough, fuel is secure, the operating regime is compatible with a 24/7 campus, and the electrical interface can be permitted and financed.

That is why I would screen this model through five gates, in this order.

Gate one: establish the usable power envelope

The headline capacity is the beginning of the work, not the answer. The relevant number is the capacity that can be delivered to IT, at the required reliability, after station auxiliary loads, conversion losses, cooling demand and redundancy.

The first diligence pack should therefore separate generation MW, import/export rights, campus electrical capacity and IT MW. PowerlandMap's note on why data center capacity is not one number is especially relevant here. A reported 400 MW electrical envelope may support materially less IT load depending on topology and operating assumptions.

The test should also include availability by hour, outage history, maintenance windows, black-start capability, islanding rules and the fallback grid path. A behind-the-meter arrangement may shorten the route to initial power, but it can create concentration risk if generation and campus operations share the same critical failure modes. Our earlier analysis of behind-the-meter power sets out that trade-off in more detail.

Gate two: price the fuel and carbon exposure

A generation-linked campus inherits the economics of its energy source. The IEA's work on energy supply for AI shows why the mix matters: data center demand is growing into power systems with different combinations of gas, nuclear, renewables and network constraints.

For investors, the issue is not only the spot cost of power. It is the delivered cost over the lease or compute contract, including fuel indexation, carbon costs, network charges, taxes, reserve obligations and the price of backup capacity. Those components should stay separate in the model. Blending them into a single €/MWh assumption makes sensitivities harder to audit and can disguise which risk is actually transferable.

This is also where location policy becomes material. Japan's Ministry of Economy, Trade and Industry is advancing its Watt-Bit collaboration framework to coordinate electricity and communications infrastructure. That policy direction supports co-optimisation, but it should not be mistaken for a project-specific permit, tariff or connection right.

Gate three: prove that the industrial site can host digital infrastructure

Power stations often have industrial zoning, transmission infrastructure and experienced operations teams. They may also have legacy contamination, coastal or flood exposure, seismic constraints, constrained access, noise limits and complicated cooling-water rights.

The right question is not whether the land is industrial. It is whether the complete campus layout can satisfy planning, environmental, fire, water, logistics and resilience requirements without creating interfaces that cannot be financed. A structured data center permit due diligence should map each approval to the activity it authorises, its conditions and its critical path.

This matters particularly at operating power stations. Construction phasing must protect existing generation, while future data halls must remain accessible for equipment replacement. Shared substations, cooling corridors or control systems can save capital, but they can also blur responsibility between utility, developer, operator and tenant.

Gate four: separate capital announced from capacity deliverable

The JERA programme is reported at more than $15 billion, covering land, power, data center facilities and AI compute. That is a useful signal of ambition, not a unit-cost benchmark. The scope spans categories that should never be combined when comparing data center CAPEX.

Infrastructure outside IT, servers and GPUs, storage and networking, land, fuel assets and grid works each have different depreciation profiles and financing structures. I would not derive a cost per IT MW until the denominator and scope are explicit. The same discipline underpins our analysis that capital is not capacity.

The financing package also needs to follow the interfaces. If a utility owns generation, a developer owns the campus and a technology partner supplies compute, lenders will ask who carries completion risk, performance risk and demand risk at each milestone. A large sponsor commitment can improve credibility, but it does not eliminate the need for bankable contracts across the chain.

Gate five: match the campus to an actual deployment schedule

AI infrastructure programmes compress technology cycles. A site targeted for 2028 must make decisions today about density, liquid cooling, substations and fibre routes without assuming that today's rack design will remain static.

That is a development-readiness problem. The useful schedule is not a single ready-for-service date. It is a chain of evidence: land control, environmental scope, electrical design, fuel commitment, permits, long-lead orders, construction access, commissioning and tenant or compute deployment. Our development readiness milestones provide a practical way to stage that evidence, while the grid upgrade cost framework helps keep shared and project-specific works distinct.

I would also model at least two density scenarios. A facility can be physically complete but commercially mismatched if its cooling and electrical design cannot accommodate the selected compute generation. Conversely, designing every component for a speculative maximum can strand capital.

What is repeatable, and what remains local

The repeatable idea is straightforward: energy-rich industrial nodes can become candidates for AI infrastructure when they combine controllable power, developable land and a credible delivery consortium. The local execution is not repeatable by template. Fuel markets, grid codes, environmental regimes, cooling options and telecom routes remain jurisdiction-specific.

This is where market intelligence is useful before detailed engineering. A developer should compare candidate nodes across PowerlandMap's market coverage, test the public evidence behind each project and identify which missing facts can change the investment case. A short platform demonstration can show how supply, power and development signals are compared, while teams that need continuous screening can review subscription options.

My view

Power-station co-location is not a shortcut around development discipline. It is a different development stack. It can reduce exposure to a remote grid queue, but it increases the importance of fuel security, operating interfaces, environmental diligence and precise capacity definitions.

The JERA-RHAELM signal is significant because it brings utility, development, compute and capital roles into one announced programme. The next proof points will be more granular: disclosed IT capacity, contractual power arrangements, environmental approvals, phase design and procurement milestones. Until those are visible, the right interpretation is high-potential and early-stage, not fully de-risked.

For teams evaluating power station data center development, the practical next step is to compare candidate sites with consistent definitions and document the evidence behind each assumption. Request access to PowerlandMap to screen markets, projects, power conditions and development signals in one place.

*Matthieu Gallego*