INS-01Energy modelling · Analysis
Two bottom-up models, the same country and the same year, roughly 50% apart. The gap does not live in the scenario labels. It lives in three conversion constants, and only one of them is about AI.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
The short answer
First principles
Published projections are usually quoted as headline totals with the assumptions stripped off, which is what makes them look like disagreements about AI. They are not. Every figure in the table below is a deterministic output of a small set of stated parameters, and the honest way to read a range is to ask what would have to be true for each endpoint to be the wrong one. The four-case framework was set out in the IEA's 2025 Energy and AI report[2] and carried forward, updated, into its 2026 edition.
Both models compute the same identity. Annual energy equals installed capacity multiplied by 8,760 hours and by a capacity factor; total facility energy equals IT energy multiplied by PUE. That means three independent levers set the answer: how much accelerated capacity is installed, how hard it runs, and what the infrastructure overhead costs. Public argument concentrates almost entirely on the first.
The second lever is the one to watch. Across all four of the IEA's cases and every projection year, the published fleet capacity factor sits between 48% and 50% — it is effectively a constant, and it is never sensitivity-tested.[1]Berkeley Lab independently uses 50% when it converts energy to interconnection capacity, and is candid that the figure is weakly evidenced: current utilization of interconnection capacity "is not well documented but is estimated to be around 50% due to high redundancy requirements and maintenance needs."[3] Two institutions converging on the same lightly-sourced constant is not corroboration.
The third lever is quietly load-bearing too. The IEA's Base Case has fleet PUE improving from 1.38 in 2025 to 1.29 in 2030.[1]Uptime Institute's 2025 survey of more than 800 operators reports industry-average PUE essentially flat for about six years.[6] Those two statements are not compatible without a large mix shift toward new-build hyperscale capacity — which is a real effect, but it is an assumption, not an observation.
Sources for Table 1: IEA Key Questions on Energy and AI, Annex A[1]; LBNL 2025 Update, Tables 1–2[3]
Table 1 · Published projections
The table states each projection with the assumption doing the most work and the observation that would falsify it.
| Ref | Projection | Scope | Value | Load-bearing assumption | What would make it wrong |
|---|---|---|---|---|---|
| DC-01 | IEA Base Case | World, 2030 | 945 TWh | Accelerator shipments rise by a factor of 2.2 from 2025 to 2030, in line with IT-industry projections; fleet capacity factor holds near 48%; fleet PUE falls from 1.38 to 1.29. | The fleet runs harder than 48% as the AI share grows, or accelerator shipments outrun the IT-industry forecast the model is fed. |
| DC-02 | IEA Lift-Off Case | World, 2030 | 1,008 TWh | Stronger AI adoption, and bottlenecks in chips, energy equipment and grid connection are effectively resolved over time. | The bottlenecks persist. The IEA's own 2026 update reports a high-bandwidth-memory shortage expected to last through at least end-2027 and places Lift-Off close to the Base Case in the near term. |
| DC-03 | IEA Headwinds Case | World, 2030 | 833 TWh | AI monetisation disappoints, investment slows, and local constraints delay data center development. | Capital keeps arriving. The same report puts announced 2026 capital expenditure by leading hyperscalers and neo-clouds at USD 715 billion, a 75% increase on 2025. |
| DC-04 | IEA High Efficiency Case | World, 2030 | 868 TWh | Demand follows the Base Case, but rightsizing of models, continued model-efficiency gains and a shift toward edge computing offset it. | Efficiency is spent rather than banked — cheaper inference per token buys more tokens, leaving facility energy flat or higher. |
| DC-05 | IEA Base Case | United States, 2030 | 426 TWh | Same model and same shipment inputs as DC-01, resolved to the US region. | See DC-06 — a second bottom-up model puts the same quantity 52% higher. |
| DC-06 | LBNL Reference Case | United States, 2030 | 649 TWh | Shipment data through November 2025; accelerator service life 5 years; AI server utilization 80% for training and 20% for inference; AI idle power 20% of rated. | Accelerators retire a year sooner (LBNL's own case: −9.1%) or shipments consolidate 15% lower (−10.8%). |
| DC-07 | LBNL High Inference Energy | United States, 2030 | 782 TWh | Same installed equipment as DC-06, but AI idle power rises to 30% of rated and inference utilization to 30%. | Inference stays lean. Note what this case does not change: installed capacity is identical to the Reference Case. |
| DC-08 | LBNL Compounded Uncertainty | United States, 2030 | 521–843 TWh | Every parameter simultaneously at its extreme. LBNL states it does not assume these variables are correlated; the range is a stress test, not a distribution. | Nothing. It is a bound. Treating its midpoint as a forecast is a category error, and its low bound still sits 22% above the IEA Base Case for the same country and year. |
Figure 1 · Original analysis
The following perturbs the IEA's own published Base Case parameters one at a time and compares each result against the spread the IEA itself publishes between its scenarios.
TWh shift · cyan = unargued conversion constant · blue = the IEA's own scenario spread
Table 2 · The numbers behind Figure 1
Each row changes exactly one published parameter and leaves the rest of the identity untouched.
| Ref | Change from the published Base Case | 2030 world electricity | Difference |
|---|---|---|---|
| S-00 | Identity reproduced from the IEA's own published Base Case parameters | 950 TWh | reference (published total: 945 TWh; 0.5% rounding) |
| S-01 | Fleet capacity factor 48% → 55%, installed capacity unchanged | 1,089 TWh | +139 TWh (+15%) |
| S-02 | Fleet capacity factor 48% → 60%, installed capacity unchanged | 1,188 TWh | +238 TWh (+25%) |
| S-03 | PUE stays at its 2025 level of 1.38 instead of falling to 1.29 | 1,010 TWh | +60 TWh (+6%) |
| S-04 | For comparison — the IEA's own Base Case → Lift-Off Case | 1,008 TWh | +63 TWh (+6.7%) |
| S-05 | For comparison — the IEA's own Base Case → Headwinds Case | 833 TWh | −112 TWh (−11.9%) |
Reading the result
The result is uncomfortable for the way this debate is normally conducted. Moving the assumed fleet capacity factor by seven percentage points — from 48% to 55%, a change nobody argues about in public — shifts the 2030 world figure by more than twice the entire distance between the IEA's central case and its most bullish one. Holding PUE flat, which is roughly what Uptime's operator survey has observed for six years,[6] is worth about as much as the whole Base-to-Lift-Off gap on its own. The scenario labels absorb the attention; the conversion constants absorb the uncertainty.
The separable risk
Berkeley Lab's 2025 update makes the cleanest demonstration of this available anywhere, and it is easy to miss. Its High Inference Energy scenario raises projected 2030 US data center electricity by 20.6% — the largest single-parameter move in the study, larger than a 15% swing in accelerator shipments — purely by raising AI server idle power from 20% to 30% of rated and inference utilization from 20% to 30%. Because the installed equipment is unchanged, the report states plainly that total capacity needs would be unchanged; those facilities would simply run at higher utilization.[3]
For a developer that is a design rule, not a footnote: energy exposure and capacity exposure respond to different parameters and should be priced separately. It also means the two families of number circulating in this debate are not comparable at all. Berkeley Lab models consumption from installed equipment and puts requested interconnection capacity — and the generation that must be added to serve it — explicitly out of scope;[3] a utility interconnection queue is the opposite, a register of requests. Comparing a queue figure to a consumption forecast sets requested optionality against modelled physics.
The magnitudes are worth keeping in view. Applying its 50% utilization assumption, Berkeley Lab's Reference Case implies 148 GW of US interconnection capacity serving data centers in 2030, an average addition of 17.4 GW per year from 2024.[3] Over the same horizon NERC raised its ten-year summer peak demand growth projection from 132 GW to 224 GW in a single annual revision — a 69% increase between consecutive editions of the same assessment.[5]
The scenario labels absorb the attention; the conversion constants absorb the uncertainty.
+238
TWh added by moving capacity factor 48% → 60%
Practice
The chain from TWh to your interconnection megawatts runs through a capacity-factor assumption that neither institution tests, and the 48-50% figure both use is a fleet average that includes enterprise server rooms. A dedicated accelerated-compute facility is not that fleet.[1][3]
Three numbers move these projections: accelerator shipments, accelerator service life, and inference idle power plus utilization. Only the first is outside your control.[3]
A scenario that adds 20.6% to energy and nothing to megawatts is survivable under a capacity reservation and expensive under an energy-linked contract.[3]
Idle power fraction and per-node utilization are directly observable through out-of-band telemetry from day one. An operator who instruments this knows their own capacity factor within a quarter, while the national models are still estimating it.
Berkeley Lab's 2024 report gave 2028 as a 325-580 TWh band; its 2025 update puts the Reference Case for that year at 464 TWh and extends to 649 TWh in 2030. Any commitment whose economics depend on a specific point on that curve will be re-litigated before it is commissioned.[4][3]
Honest limits
Two published models were read side by side and one of them was perturbed. Where that stops:
In the product
That last point is the one PODOS is built around. A PODOS Pod is designed as a standardized 1 MW building block, which makes capacity a quantity added in increments as the forecast resolves rather than a single bet placed years ahead of the load. See monitoring and controls for the telemetry layer, power architecture and site power readiness for the interconnection questions, the glossary for terms, and the configurator to turn a target load into a unit count.
Summary
Bring the megawatts you can actually interconnect, not the terawatt-hours a national model projects. Engineering will size the rest.