Skip to main content

Data-Center Cooling Intelligence

On-Prem

5-12% cooling cost reduction, commissioned in seconds on a single CPU, inside your facility

5-12%
Operating cost reduction potential
Seconds
To commission on 1 CPU
100%
On-prem, telemetry never leaves

Trained in seconds on a CPU, on-prem, with calibrated bounds your operators can audit. The AI buildout has made every megawatt count, and we make learned cooling optimization something any facility can commission without a data-science team.

The Problem

Fixed worst-case setpoints

Most facilities run static cooling setpoints sized for the worst case, burning energy on every normal day

AI control demands trust and time

Deep-RL cooling control works, but requires months of per-site onboarding and black-box trust

No visibility into drivers

Operators cannot see which pumps, temperatures, or schedules actually drive total consumption

Data sensitivity

Colocation and edge operators cannot send facility telemetry to the cloud for analysis

Bottom line: AI cooling control demonstrably works (Google proved it in production), but the deep-RL route keeps the savings on the table for everyone else.

The Product: An Advisory Cooling Optimizer

Connects to existing BMS, BACnet, and DCIM systems, with no new sensors and no data-pipeline project

A learned thermal-power model is fit in seconds on a single CPU, fully on-prem

Every 15 minutes the system recommends the setpoint that minimizes energy and carbon

Every recommendation carries a calibrated confidence interval and stays inside the thermal envelope

Operators keep final authority; closed-loop only after savings are measured

What-if analysis: operators can drag any lever and see the predicted KPI impact before acting

Existing telemetry
BMS, BACnet & DCIM native
1 CPU, seconds to fit
No GPU, no cloud
Every 15 minutes
Auditable setpoint advice

Results from a Live Facility Analysis

Anonymized findings from a real data-center deployment, with site details withheld for client privacy.

Model Fit Accuracy
97.8%
Learned from ~46,000 historical records across 180 telemetry variables at a live facility
Top Drivers Identified
~58% explained
Two controllable variables, pump load and supply temperature, explained most of total consumption
Operating Cost Potential
5-12% reduction
Primarily via setpoint optimization, scheduling, and chiller water inlet temperature tuning
Commissioning
Seconds on 1 CPU
Comparable deep-RL benchmarks train on 100+ CPU servers with far more data

Operational Insights Delivered to the Team (Examples)

Example insights from this analysis: not a black box, but a concrete, prioritized action list operators can execute today.

1

Prioritize the dominant pump load driver: peak-load reduction and load balancing across units

2

Widen the supply-temperature setpoint band gradually to prevent overcooling

3

Align operating schedules: early/late starts and night loads reviewed against actual demand

4

Verify auxiliary pumps only run when needed, avoiding premature staging

5

Tune chiller water inlet temperature toward the efficiency sweet spot

Start With a Read-Only Savings Estimate

No pilot budget, no installation, nothing touching your plant. Send historical telemetry and get back the same analysis shown above for your own facility. The model itself fits in seconds.

You send exports
3 months of BMS/DCIM history (CSV is fine)
Fit runs in seconds
Data twin, top drivers, and a savings estimate, reviewed with you same week
No commitment
NDA first, data deleted on request
Request a savings estimate

A Data Twin of Your Facility at Affordable Compute

We build a data twin of your facility from its own telemetry: a learned model you can test setpoints against before touching the plant. Deep reinforcement learning is the approach hyperscalers use for cooling. Google's DeepMind, acquired for around $500M in 2014, cut data-center cooling energy by up to 40% with it. It also needs a research team, a training cluster, and a retraining cycle. We match published deep-RL results on the same open benchmark, using their code and their metrics, on hardware that already sits in your facility.

Savings vs fixed-setpoint baseline

SustainDC benchmark, 5 held-out months (% saved)

HVAC energyIT energy0%3%6%9%12%

Commissioning compute

Hardware needed to train the controller (CPUs)

Deep RLIoTGPT0306090120
Trains & retrains on-prem
1 CPU, no cloud
Light enough to refit inside the facility as seasons, loads, and hardware change — deep-RL controllers need a training cluster and a research team for every update.
Why we never target PUE
−8.8% energy
A PUE-optimized controller "improves" the ratio 1.3% while raising total energy +11%. We report PUE, never target it.

Compared on the SustainDC data-center benchmark (HP Labs, NeurIPS 2024): a data twin of a facility, evaluated over 5 simulated months of workload and weather not used for fitting. Published deep-RL numbers are cited from that paper, not rerun. All results are from the data twin, at affordable compute, and a real-site advisory pilot is the next step. We report PUE, never target it; optimizing the ratio alone can inflate IT power.

On-Prem by Construction

Telemetry never leaves

All learning and inference happen inside the facility.

No GPU, no cloud

A single CPU is enough, with no data-pipeline project required.

Deterministic & auditable

Calibrated bounds instead of black-box trust; every recommendation is inspectable.

Ready to See Results in Your Data Center?

Book a free demo and we'll show you exactly how IoTGPT can improve your KPIs — using your own operational data.