Autonomous AI cooling control in a live data center

FLUIX's A.I.M.I. controller took direct, autonomous control of cooling at TenHats, a 2 MW mission-critical, AI-ready facility, inside guardrails set by the site's operators.

Key facts

Site	2 MW, Knoxville, TN
Period	Dec 2025 – Mar 2026
PUE	1.36 → 1.27 (6.6%)
SLA excursions	Zero under full AI control
Cooling response	~15 min → ~5 min
Outcome	Continuation signed

What happened

The controller ran on-premises, integrated with the existing building management and HVAC systems, with no new sensors and no cloud dependency in the control loop. It learned the site's thermal behavior in under two weeks.

Preliminary results, reviewed by TRC, show PUE improving from 1.36 to 1.27 and modeled HVAC savings of about 0.4 MWh per day during full control periods. FLUIX projects up to about 130 kW of additional power headroom at a fully loaded 2 MW profile. TVA EnergyRight supported the research and deployment.

Before the live deployment, FLUIX validated the controller in a purpose-built mock data hall. In medium-load runs it cut HVAC power from 3.08 kW to 1.44 kW versus the unit's native controller, with a safety layer that checks every action against thermal and actuator limits and reverts to the native controller on any anomaly.

Results reported by FLUIX AI. Live-site figures are preliminary; lab figures are from a mock data hall.

What it proves for Aidni

AI can operate real data-center systems safely when it is bounded by operator guardrails, hard limits and fail-safes, and the efficiency it recovers turns directly into power headroom. That is the operating principle behind Conduit, and the headroom is exactly what Grid plans compute against.

Aidni Conduit →Aidni Grid →Aidni Thermal →

Exhibits

The controller and its lab validation.

Diagram of the A.I.M.I. 1.0 closed-loop controller: data center telemetry feeds a reinforcement learning controller with a safety and constraint layer, which issues continuous cooling setpoints. — **Fig. 1 ·** A.I.M.I. 1.0 control loop with its safety and constraint layer.Image: FLUIX AI

Two power charts: HVAC power under the native CRAC controller versus under A.I.M.I. control, with IT load steady at about 2.8 kW. — **Fig. 2 ·** Lab study, medium load: HVAC power under the native controller (left) vs. A.I.M.I. (right).Image: FLUIX AI

PUE overlay chart comparing the native controller run and the A.I.M.I. run over about 250 minutes. — **Fig. 3 ·** Lab study: PUE overlay, native controller vs. A.I.M.I.Image: FLUIX AI

Computer room air conditioning unit in the FLUIX test lab, remotely operated by a local server. — **Fig. 4 ·** FLUIX test lab: CRAC unit under AI control.Image: FLUIX AI

FLUIX lab containment with a cold-aisle flooded zone and an isolated hot-air return plenum inside a warehouse. — **Fig. 5 ·** FLUIX test lab: containment with isolated hot-air return.Image: FLUIX AI

FLUIX A.I.M.I. dashboard showing site capacity, live PUE and a log of setpoint changes made by the AI. — **Fig. 6 ·** A.I.M.I. operator dashboard with an audit log of AI setpoint changes.Image: FLUIX AI

Sources

About the partner

FLUIX AI builds A.I.M.I., an autonomous controller for data-center cooling. Aidni and FLUIX work together on governed AI operations for data centers.

fluix.ai →

Next case study →

AI cooling control at hyperscale

Industry precedent · Google DeepMind · Google data centers

Start here

Bring us one site. We will show you its real capacity.

A six-week Site Assessment maps the power envelope, the cooling design and the compute plan for one facility, delivered by the engineers who build the platform.

Book a site assessment Design Partner Program