Why AI Data Center Capacity Needs a Confidence Score

AI data-centre capacity has become one of the most repeated numbers in infrastructure markets, and one of the easiest to misuse.

A site can be described as 100 MW when that number means critical IT load, total facility input, a utility application, an accepted connection, a future campus ceiling or simply a developer’s ambition. Those are not interchangeable states. My conclusion is straightforward: every capacity figure used for site selection, underwriting or market sizing should carry a confidence score that explains what the number measures, how mature it is and when it can become usable.

That is not a call to discard early signals. Early signals are valuable. It is a call to stop compressing unlike evidence into a single total.

The current Epoch AI data-centre dataset makes the issue visible. Its 20 October 2026 update covers 95 large AI data centres with an estimated 14.2 GW of IT power capacity. Epoch estimates total facility capacity at 18.5 GW after cooling and other infrastructure overheads, while noting that average consumption is typically 60–80% of capacity. The dataset is useful because it publishes definitions and uncertainty. It is also a reminder that even a carefully assembled total represents a modeled and incomplete view of a fast-moving market.

For an investor, operator or utility, the decision is therefore not “which total is largest?” It is “which part of this total can support a contract, a grid plan or a capital commitment?”

One number can contain three different power boundaries

The first control is the measurement boundary.

IT capacity is the electrical power available to servers, storage and networking. Facility power includes IT load plus cooling, pumps, fans, power conversion and other building systems. Grid or campus capacity may sit further upstream and can include reserve, common infrastructure or later phases.

The relationship is often expressed through power usage effectiveness:

Facility power = IT power × PUE

Epoch’s published figures imply a ratio of approximately 1.30 between 18.5 GW of facility capacity and 14.2 GW of IT capacity. That ratio is plausible as a portfolio-level estimate, but it should not be silently applied to an individual project. Climate, redundancy, cooling architecture, loading and measurement period can all change the result. Uptime Institute’s survey material also cautions that PUE is a facility-efficiency metric, not a universal basis for comparing unlike sites.

A reproducible example shows why this matters.

| Input | Base case | Low-use case | High-use case | |---|---:|---:|---:| | Design IT capacity | 100 MW | 100 MW | 100 MW | | Assumed PUE | 1.20 | 1.20 | 1.20 | | Facility capacity at full IT load | 120 MW | 120 MW | 120 MW | | Average utilisation assumption | — | 60% | 80% | | Average facility demand | — | 72 MW | 96 MW | | Annual electricity | — | 631 GWh | 841 GWh |

The calculation is simple: 100 MW × 1.20 gives 120 MW of facility input at full IT load. Multiplying by 60% or 80% utilisation gives 72 MW or 96 MW of average demand. Multiplying those values by 8,760 hours gives approximately 631 GWh or 841 GWh per year.

This is a model, not a forecast. It excludes outages, commissioning ramps, seasonal PUE variation, on-site generation, storage dispatch and curtailment. Its purpose is to stop a design ceiling from being repeated as annual consumption.

Capacity needs a maturity state

The second control is project maturity.

I would not give the same weight to an identified site and an operating hall. A practical sequence is:

1. Announced: a company has described a project or portfolio. 2. Land controlled: ownership, lease or enforceable option is documented. 3. Planning advanced: the relevant application is filed or approved. 4. Grid path documented: the connection point, scope, milestones and obligations are evidenced. 5. Construction committed: major works, equipment and capital commitments are visible. 6. Energised: electrical infrastructure is live. 7. IT commissioned: compute can operate within the stated boundary. 8. Contracted or occupied: demand is attached to the delivered capacity.

A project can move backwards. Applications are withdrawn, connection dates change, phases are reduced and customers defer occupancy. That is why PowerlandMap treats project status as a change log rather than a permanent label. The related analysis, A Data Center Project Status Needs a Change Log, sets out how dated evidence should replace a single “planned” field.

This maturity sequence is not a probability model. It is an evidence model. It tells the reader which dependency has been removed and which remains.

A confidence score should be explainable, not magical

I would use five transparent components, each scored from zero to two.

| Component | 0 | 1 | 2 | |---|---|---|---| | Source | unattributed or recycled | credible secondary source | first-party or regulatory evidence | | Power boundary | undefined MW | boundary inferred | IT/facility/grid scope explicit | | Project stage | marketing only | one material milestone | multiple dated milestones verified | | Timing | no delivery date | target year or broad window | phase-specific dated milestone | | Reconciliation | no cross-check | one corroborating source | conflicts and duplicates resolved |

The total produces a ten-point confidence score. It does not predict success. It describes the evidence quality of the specific capacity claim.

A 2/10 announcement may still matter as an early market signal. A 9/10 figure may still be delayed. The score prevents the first from being added to the second without disclosure.

I would also retain the raw source wording. Normalising a figure should never erase what the company or authority actually said. If one source says “300 MW campus” and another says “300 MW IT,” the database should preserve the conflict until the boundary can be reconciled.

Energy forecasts are not site pipelines

The third control is the time and metric frame.

The International Energy Agency projects global data-centre electricity consumption to reach around 945 TWh by 2030 in its base case, roughly double today’s level. The Lawrence Berkeley National Laboratory 2025 update estimates that US data centres could account for 11.8% of national electricity use in 2030, with a scenario range of 9.5–15.3%.

Those are annual energy forecasts. They should not be added to a pipeline of site MW. Energy in TWh measures consumption over time; power in MW measures an instantaneous rate or capacity boundary. A utility interconnection queue adds another layer because requests can overlap, change or never materialise.

The distinction matters commercially. An equipment supplier may care about design IT MW and deployment dates. A utility cares about coincident peak demand, ramp profile and network location. An investor may care about contracted critical load and capital required to reach ready-for-service. A market-intelligence product has to keep all four views available without pretending that they are the same dataset.

Where the alternative view wins

There is a serious counterargument: markets move too quickly for elaborate scoring, and a simple announced-MW total is often the only comparable number available.

That view wins for rapid signal detection, provided the label remains explicit. If the task is to identify where developer attention is moving, announced campus ceilings can be useful. They become misleading only when the same total is presented as near-term supply, delivered IT capacity or expected electricity demand.

The answer is not to wait for perfect information. It is to separate discovery from underwriting.

For discovery, accept low-confidence records with a clear source date and stage. For underwriting, require the evidence appropriate to the decision: a grid document for power readiness, a planning record for entitlement, a construction milestone for delivery and a customer or occupancy disclosure for demand.

The decision rule I would use

Before a capacity figure enters a shortlist, model or market total, I would require five answers:

• What exactly does the MW measure?
• Which physical site and phase does it belong to?
• What dated evidence supports the current stage?
• What is the earliest independently supportable delivery milestone?
• Could the same capacity already appear under another company, campus or source?

If any answer is missing, keep the value but lower its confidence and exclude it from delivered-capacity totals.

This approach changes the output. Instead of “the market has X GW,” the result becomes a layered statement: X GW announced, Y GW with a documented grid path, Z GW under construction and W GW commissioned, each with a defined IT, facility or grid boundary. That is less dramatic, but far more useful for capital allocation.

PowerlandMap’s role is to preserve those distinctions across locations, phases, sources and time. The objective is not a larger number. It is a number that can survive the next diligence question.

For teams comparing sites or validating a pipeline, start with the evidence behind the MW before ranking the MW itself.