Choosing a site for an AI data center requires evaluating available power and the energization schedule alongside cooling, networking, land and building conditions. Co-location with a power plant and behind-the-meter supply may ease some power constraints, but available electricity must still be converted into facilities and GPU systems that can pass acceptance testing before compute can be delivered. Japan's Chiba project illustrates how to assess capacity, power arrangements and phased delivery.
Assessing Available Power Alongside Site Conditions
Available power is the capacity a site can obtain under agreed supply and equipment conditions. The energization schedule is when the site can formally receive power after the required applications and construction are complete. A nearby power plant, secured land or an order for substation equipment each represents only one part of progress.
Separate the initial phase from the ultimate scale early in site selection, and check whether power, cooling and networking can be completed together for each phase. If planned power becomes available only later, the initial facility must be designed around the capacity that can actually be delivered.
Generation Capacity, Available Site Power and IT Load
Generation capacity describes capability at the source, while IT load is the electricity used by servers, networking and storage equipment. Capacity allocation, supply constraints and facility energy consumption sit between the two, so they cannot be treated as equivalent.
| Capacity level | Meaning | Conditions to distinguish |
|---|---|---|
| Generation capacity | Rated output of generating equipment | Actual supply also depends on operation, maintenance and existing allocations. |
| Available site power | Power capacity the site can obtain | Check supply arrangements, energization works and the availability date. |
| Total facility consumption | Consumption of the data center as a whole | Includes IT, cooling, distribution losses and other facility uses. |
| IT load | Electricity used by IT equipment | Includes complete servers, networking and storage, a broader scope than GPUs alone. |
Power capacity is commonly expressed in megawatts (MW), while energy consumption is measured in MWh or kWh. Using 1 MW continuously for 1 hour consumes 1 MWh, or 1,000 kWh.
Power Usage Effectiveness (PUE) is total facility energy consumption divided by IT equipment energy consumption over the same period. The Green Grid uses it to measure data center energy efficiency.
Average IT power ≈ average total facility power over the same period ÷ PUE.
For example, assume a facility averages 12 MW under stable operation and has a PUE of 1.2. Average IT power is about 10 MW, with the remaining 2 MW used for cooling, power distribution and other facility functions. This is an illustrative assumption: 12 MW of incoming supply capacity does not necessarily deliver 10 MW of IT capacity. Capacity planning must also account for peaks, redundancy and headroom.
What Behind-the-Meter Supply Means and How to Assess It
Behind-the-meter (BTM) supply places generation or energy-storage resources on the customer's side of the meter used for utility settlement, allowing energy to serve loads on that side. It describes the electrical arrangement. Whether a power plant is nearby and whether the site is entirely off-grid are separate questions.
Behind-the-meter resources can include equipment such as on-site solar. Large data centers may also arrange dedicated supply paths from a power plant, but the metering boundary, supplier and relationship with the grid still need to be clarified.
Assess behind-the-meter supply in three areas:
-
Supply path: can existing generation and site works reduce the wait for new transmission or grid connections?
-
Continuous operation: how will the grid, other generation, UPS and backup equipment support the load during planned maintenance or a source failure?
-
Cost and requirements: compare fuel, equipment, backup, operations, maintenance and applicable charges together, and confirm local supply and interconnection requirements.
Bringing the source closer to the load may shorten the supply path. Reliability, electricity cost and emissions still depend on the particular source and contractual arrangements.
How Power-Plant Co-location May Shorten Delivery: Japan's Chiba Project
Where grid connection queues or transmission expansion constrain progress, using an existing source and arranging a direct supply path may reduce dependence on new grid works and shorten the wait for available power. This does not remove supply approvals, necessary grid assessments or backup design.
On October 1, 2026, JERA, Dell and RHAELM announced a memorandum of understanding. A project at a site adjacent to Chiba Thermal Power Station would provide a practical foundation for a standardized AI infrastructure model.
| Participant | Announced primary role |
|---|---|
| JERA | Contributes energy and power-plant site capabilities. |
| Dell Technologies | Provides standardized, rack-scale compute infrastructure through Dell AI Factory. |
| RHAELM | Leads project delivery, integrating energy, the site and compute infrastructure development. |
The project plans for up to 400 MW of power capacity, using a behind-the-meter approach with existing generation assets, and targets operation around 2028. These are development plans and targets. The 400 MW figure is not equivalent to GPU consumption or operational capacity, and is insufficient on its own to calculate a GPU count.
The useful lesson is coordinated planning across existing generation, the site and compute facilities. The project does not provide a fixed schedule that can simply be replicated elsewhere. Whether another site can shorten delivery still depends on source availability, construction, approvals and equipment delivery.
Four Delivery Checkpoints Between Available Power and Usable Compute
Moving from available power to operational compute requires completing power supply, facilities, equipment and integrated acceptance testing. The actual schedule depends on the last critical item that would prevent subsequent work from proceeding.
| Confirm power conditions Capacity · supply path · energization | Prepare the facility Distribution · cooling · networking | Deploy GPU systems Servers · interconnects · storage | Integrated acceptance and delivery Load tests · failure scenarios |
|---|
Power conditions require confirmation of capacity, the supply arrangement and energization works. Facility readiness covers distribution, cooling, racks and network paths. GPU deployment includes servers, interconnects, storage and the management environment. Integrated acceptance uses load tests, failure scenarios and agreed workloads to confirm that facilities and clusters operate together.
These are delivery checkpoints. When estimating the schedule, separately identify applications and approvals, long-lead equipment, construction, installation and testing. Keep the energization date distinct from the date services can be provided.
Assessing Sites and Phased Delivery for AI Data Centers in Taiwan
AI data center development in Taiwan can start with initial-phase demand, assess power conditions and site capability together, and then plan phased delivery to compare costs and expansion options across sites.
Estimate IT power, cooling and networking needs from the initial GPU systems and workloads, and confirm capacity and engineering conditions with the power supplier. In parallel, check permitted land and building uses, floor loading, heat rejection, water supply and telecommunications routes, and include equipment procurement and testing in the same schedule. Facility design and AIDC construction requirements must be validated against the target GPU systems.
As site and power conditions become clearer, the next step is to translate available capacity into rack-level power distribution, cooling and equipment configurations for each phase, with corresponding acceptance criteria. This is where facility construction needs to integrate site planning with GPU deployment.
KONST Group's infrastructure brand, Konstra AI, provides electrical and mechanical systems, environmental controls and GPU cluster construction, planning facility conditions together with compute equipment deployment. Enterprises preparing to enter the construction phase can contact KONST to discuss site conditions, the initial configuration and construction scope.
FAQ
Approximately How Many GPUs Can 400 MW Support?
First identify which capacity level the 400 MW figure represents, then estimate IT consumption and its allocation across complete servers, networking and storage. An illustrative formula is: power available to GPU servers ÷ design power per server × GPUs per server. Dividing 400 MW directly by an individual GPU's power consumption cannot determine the deployable GPU count.
Is Behind-the-Meter Supply Legal, and Can It Be Used in Taiwan?
Behind-the-meter describes equipment and metering arrangements; the label alone does not establish whether a specific project is lawful. A project in Taiwan should first clarify self-use versus supply or sale of electricity, generation and consumption equipment, interconnection, and applicable land and building requirements. Relevant authorities, the power supplier and professional designers then need to confirm the arrangements. An overseas case cannot serve directly as approval for a project in Taiwan.
What Does Phased Delivery Mean?
Phased delivery divides the overall planned capacity into stages that can be accepted and used independently. Each stage needs corresponding power, cooling, networking, equipment and operating arrangements, with the impact of later construction on existing operations also assessed.
