In short
- Energy saving = IT × (PUE − 1) × cooling share × saving gained between levels × 8,760 h.
- Capacity released = (design capacity − IT load) × share released; its value is the new build it defers.
- A 10 MW site at PUE 1.45 moving from L2 to L4 saves about 3,300 MWh and €500k a year, with payback near 1.2 years.
- The released 2.3 MW defers about €23M of new capacity, often the larger number.
Cooling energy saved
The new PUE follows: 1 + (overhead − cooling × Δs) ÷ IT. Money = ΔE × electricity price; CO₂ = ΔE × grid factor.
Stranded capacity released
If a site uses 60 % of its design capacity, the design is IT ÷ 0.6 and the stranded part is the difference. A calibrated twin proves where racks can go safely, releasing a share of it. The value is the cost of building that capacity new, typically around €10M per MW as a planning figure.
Cost
The planner uses one-off costs for a 10 MW site of about €120k at L1, €250k at L2, €450k at L3, €700k at L4 and €950k at L5 (cumulative), scaled with the square root of site size, and a running cost of 25 % of the one-off per year. These are planning assumptions; real quotes depend on what data and models already exist.
10 MW, PUE 1.45, from L2 to L4
| Quantity | Value |
|---|---|
| Overhead = 10 × 0.45 · cooling share 70 % | 4.5 MW · 3.15 MW |
| Saving gained (15 % − 3 %) | 12 % |
| Cooling energy saved = 3.15 × 0.12 × 8,760 | ≈ 3,310 MWh/yr |
| Saved at €150/MWh · CO₂ at 0.35 t/MWh | ≈ €497k/yr · ≈ 1,160 t/yr |
| PUE after | 1.45 → 1.41 |
| Stranded capacity at 60 % use · released (40 % − 5 %) | 6.7 MW · 2.3 MW |
| New build deferred at €10M/MW | ≈ €23M |
| One-off cost (€700k − €250k) · running 25 % | €450k · €112.5k/yr |
| Net saving per year | ≈ €384k |
| Payback · cumulative after 5 years | ≈ 1.2 years · ≈ €1.47M |
The energy case alone pays back quickly; the capacity case is what usually wins the budget.
What has been measured
| Claim | Figure | Source |
|---|---|---|
| Cooling energy, ML control | Up to 40 % less cooling energy, 15 % less PUE overhead | Google DeepMind, 2016 · operator |
| Cooling energy, AI twin | 15–20 % reduction expected across sites | Telefónica Germany via dotmagazine, 2026 · operator |
| Stranded capacity | Enterprise data centers use 56 % of capacity | 451 Research via Cadence · research |
| Adoption | 33 % have or plan a digital twin | Eaton / 451 Research, 2023 · research |
| AI factories | Up to 30 % more GPU throughput in the same power | NVIDIA Omniverse DSX, 2025 · vendor |
Vendor figures are best cases. See the evidence section.
Making the case internally
- Measure the baseline: PUE, cooling share and capacity used.
- Value the stranded capacity at your own new-build cost.
- Price the next level only, not the whole ladder.
- Agree how savings will be measured before the work starts.
- Name an operations owner for the twin.
Questions about twin ROI
What savings are realistic?
Telefónica Germany expects 15–20 % less cooling energy with an AI-based twin. Google's DeepMind work reached up to 40 %, a best case.
How does a twin release capacity?
By proving cooling and power headroom with calibrated what-if models, so racks can be placed where operators previously did not dare.
Does the twin's own cost include sensors?
The planner's figures assume typical existing BMS and metering. Sites that need many new rack sensors should add that cost.
Does a lower PUE always mean lower bills?
Yes for the cooling energy saved, at constant IT load. If released capacity is filled with new IT, total consumption rises but cost per kW of IT falls.
Sources
- Planner model on this site: savings 3/8/15/20 %, released 5/30/40/50 %, cost by level, running cost 25 %.
- Google DeepMind 2016; Telefónica Germany via dotmagazine 2026; 451 Research via Cadence; Eaton / 451 Research 2023; NVIDIA Omniverse DSX 2025.
- All figures are planning assumptions to edit, not quotes.