In short
- L1 static model · L2 connected · L3 simulating · L4 optimising · L5 autonomous.
- A level counts only when all its items and all lower levels are in place.
- Planning assumptions for cooling energy saved: 3 / 8 / 15 / 20 % at L2 to L5.
- Climb one level at a time; L3 is where capacity is released, L4 where energy savings grow.
The five levels
| Level | Answers | Cooling saved | Stranded capacity released |
|---|---|---|---|
| L1 Static model | What is where? | — | — |
| L2 Connected | What is happening now? | 3 % | 5 % |
| L3 Simulating | What if? | 8 % | 30 % |
| L4 Optimising | What should we change? | 15 % | 40 % |
| L5 Autonomous | Change it within limits | 20 % | 50 % |
Savings and released shares are the planner's editable assumptions, set inside reported results. The levels synthesise common maturity models (descriptive, predictive, prescriptive, autonomous); terms follow ISO/IEC 30173:2023. Explore them on the maturity ladder.
Where your site is today
| Item | Level |
|---|---|
| As-built 3D model of the hall (BIM or scan to BIM) | L1 |
| Asset register linked to the model | L1 |
| BMS and power meter data streamed to the model | L2 |
| Rack-level power and inlet temperatures | L2 |
| Calibrated CFD model of the halls | L3 |
| Electrical what-if (failure and capacity) scenarios | L3 |
| Recommendations from analytics or ML | L4 |
| Operators act on recommendations in a workflow | L4 |
| Setpoints changed automatically within limits | L5 |
| Alarms, guardrails and manual override tested | L5 |
Static and connected
L1 is an accurate as-built model and asset register: the base every other level reads from. L2 streams live BMS, power and rack data into it. The value at L2 is visibility: live PUE, hot spots, power per rack. The work is mostly data engineering: consistent tags (Brick, Haystack), a single naming scheme and reliable feeds.
Simulating: where capacity is released
A CFD model calibrated against measured temperatures, plus electrical what-if scenarios, lets operators prove headroom instead of guessing. Enterprise data centers use about 56 % of their capacity (451 Research via Cadence), often because nobody can show it is safe to add more. L3 is where that stranded capacity starts to come back.
Optimising and autonomous
L4 adds analytics or machine learning that recommend setpoint changes, with the reasoning shown, and a workflow in which operators act on them. L5 lets the system change setpoints itself within limits, with guardrails, alarms and a tested manual override. Google's DeepMind work cut cooling energy by up to 40 %, a best case; Telefónica Germany expects 15–20 %.
Why twins stall
- Skipping layers: optimisation on top of unreliable data produces recommendations nobody trusts.
- Naming chaos: the same sensor called five different things across BMS, EPMS and DCIM.
- Model drift: the as-built model is not updated after changes, so the twin stops matching the hall.
- No owner: a twin needs an operations owner, not only a project team.
- Security late: write-back at L5 needs cybersecurity designed in, not added after.
Questions about maturity
Which level should we aim for?
Most sites gain the most by reaching L3 for capacity and L4 for energy. L5 makes sense once operators trust L4 recommendations and the guardrails are proven.
How long does it take to climb a level?
Usually months per level, driven by data quality more than software. L1 to L2 is mostly integration and tagging; L2 to L3 is calibration.
Can a new build start at L3?
Yes, if the design model, commissioning data and live feeds are handed over together. New builds can skip the clean-up that existing sites need.
Is L5 safe?
Only with limits, alarms and a tested manual override. Automatic changes stay inside an envelope the operators define.
Sources
- ISO/IEC 30173:2023 terms; level structure is a synthesis of common maturity models.
- Google DeepMind (2016); Telefónica Germany via dotmagazine (2026); 451 Research via Cadence.
- Savings and released shares by level are planning assumptions, not quotes.