One operating system, from the edge to the cloud.
AquaWatt DC OS is a resource operating system for AI data centers. It spans five layers — from the physical equipment on the floor to a governed control plane that verifies every action against the real world.
Facility Systems
The equipment and building systems that keep an AI data center running.
AquaWatt Edge
An on-premise layer that connects to equipment, buffers data locally, and enforces safety policies close to the machines.
Data & Intelligence Cloud
A cloud data plane that turns streaming telemetry into forecasts, digital twins, and multi-objective optimization.
Application Layer
The workspaces operators, engineers, and sustainability teams use every day.
Control & Verification
Every recommended action is validated, approved, applied, and then checked against the real-world outcome.
Speaks the language of your facility.
The AquaWatt Edge connects to existing building and IT systems through the protocols your equipment already uses — no rip-and-replace required.
From a sensor reading to a verified outcome.
Data flows outbound from the edge to the cloud for intelligence, but control always returns through a governed path: nothing changes on the floor without policy checks and an operator’s approval, and safety-critical logic stays local.
See the architecture applied to your facility.
Walk through how AquaWatt DC OS would connect to your systems and where it can create measurable value.
It has to read your facility before it can improve it.
A secure edge gateway collects data from on-premises equipment and connects the systems that already run your site. Nothing is inferred from a single source, and no stream influences a recommendation until it passes quality checks.
Supported protocols
Targeted protocol support for the edge gateway.
Building management systems
Facility-side conditions and cooling behaviour ingested from the BMS.
DCIM systems
Inventory, capacity, environmental and power data from the DCIM layer.
Server and accelerator telemetry
Device-level signals that tie compute demand to power and heat.
Workload platforms
Scheduling context so compute placement can be reasoned about, not guessed.
Utility and grid data
The external cost and carbon signals that make timing decisions matter.
Weather
Current and forecast conditions that drive cooling and water demand.
Water data
The water picture most platforms leave out entirely.
Trust the data before you trust the decision.
Data-quality monitoring
Every incoming stream is checked before it is allowed to influence a recommendation.
- Missing signals
- Frozen sensor values
- Invalid units
- Out-of-range readings
- Duplicate data
- Timestamp drift
- Unexpected sampling changes
Unit normalization
Measurements are normalized into a common model while the original values and units are preserved, so nothing is silently rewritten.
Integration health
Operators see the state of every connection rather than discovering a dead feed after the fact.
- Connection status
- Latency
- Last message time
- Error history
Integration scope reflects planned platform capability. Specific vendor systems, protocol coverage and control integrations are confirmed per site during pilot discovery.
Try it on the model before you try it on the hall
A live facility is the worst possible place to test a hypothesis. The digital twin gives the platform — and your operators — somewhere safe to be wrong.
What the twin represents
Nine interacting layers of the facility, calibrated against measured performance rather than design specifications.
Simulate the question you are actually asking
Select a scenario to see the dimensions every simulation is scored against — the comparison is always multi-objective, never a single number.
Temperature-setpoint changes — evaluated across all eight dimensions below:
Calibrated against reality
Models are tuned using measured facility performance, so the twin reflects how the building actually behaves rather than how it was specified to behave.
Uncertainty is disclosed
Every simulation states its confidence and its model limitations. A prediction presented without error bars is a guess wearing a suit.
Pre-action validation
Eligible automated actions are validated against the twin, or an approved surrogate safety model, before they are allowed to execute.
Describes planned digital-twin capabilities under development. The scenario selector above is an illustration of the simulation framework, not a live model of any facility.
Not one model — a governed portfolio of them
Forecasting a chiller is not the same problem as attributing water to a workload or answering an operator’s question in plain English. AquaWatt DC OS runs a portfolio of specialised models, and every one of them is registered, monitored and reversible.
Model categories in the platform
Every production model is on the record
A model without provenance has no business influencing critical infrastructure. Each one carries a full record before it ships.
- Model owner
- Version
- Training-data window
- Approved use
- Performance metrics
- Validation record
- Deployment date
- Monitoring thresholds
- Rollback version
- Known limitations
Watched continuously, not at review time
Models degrade quietly as facilities change. These signals are tracked in production so degradation is caught by the platform rather than by an operator.
- Prediction accuracy
- Drift
- Data-quality changes
- Confidence calibration
- False-positive rate
- False-negative rate
- Optimization outcome variance
- Safety-policy violations
- Business impact
Human oversight is structural
Models may propose actions. Execution authority belongs to customer policies and the AquaWatt safety layer — a separation that holds regardless of how confident a model is.
Learning from verified outcomes only
Forecasting and optimization improve using the measured results of real actions — and only after those outcomes have passed validation. Unverified results never become training signal.
Describes the AI and governance approach AquaWatt DC OS is being built to. Reinforcement-learning methods are treated as candidates for validated control environments only and are not part of the first release.