Securing the OT that keeps AI data centers running — the power, cooling, safety, and control systems behind every GPU rack.
Power, cooling, and safety all run on OT, and GPU racks have little thermal margin, so an OT fault quickly becomes a compute outage.
Rack density has jumped from roughly 10 to 15 kW to well over 100 kW. A single liquid-cooled GB200 NVL72 AI rack draws about 132 kW — twelve times a typical rack — and twelve times the power in one footprint means twelve times the heat to remove. At that density, power and cooling stop behaving like building services and start behaving like tightly coupled industrial processes.
Those processes are run by operational technology: protective relays and switchgear, generator and UPS controllers, chiller plant PLCs, coolant distribution units, fire panels, access control, and the BMS, EPMS, SCADA, and DCIM platforms that supervise them all. These systems are increasingly converging into one supervisory platform, and they are reachable through vendor remote access, integration gateways, and business-class internet connections that sit outside IT governance.
Security has to cover both sides. ProtexICS brings an OT-native approach to AI data center cybersecurity: we start with what could go wrong physically — loss of cooling, loss of power, a false suppression release — and trace the cyber pathways that make it possible.
Typical data center rack, 2026 average
One GB200 NVL72 AI rack, liquid cooled
The power — and the heat to remove — in one footprint
Exposed ICS and building automation devices found near US data centers
An AI data center is built from six core operational units. Only two of them are IT.
Utility substation and switchgear, backup generators, UPS and batteries, busway and rack PDUs.
Chiller plant and cooling towers, CDUs and the liquid-cooling loop, and CRAH / AHU units.
Fire detection and suppression, plus physical security: access control, CCTV, and intrusion detection.
BMS, EPMS, and DCIM, with SCADA HMIs, historians, and remote access over every other OT system.
Liquid-cooled GPU racks of about 120 to 132 kW — fed entirely by the OT power and cooling systems.
NVLink inside racks, plus InfiniBand or Ethernet at 400 to 800 Gb/s between them, with an out-of-band management network for BMCs and DPUs.
OT controls the electrical chain from utility feed to rack, and the chilled water process that carries heat away from the GPUs.
A manipulated EPMS reading or relay setting can hide an overload or trip an entire hall offline.
Liquid-cooled racks have little thermal margin, so a cooling fault quickly causes an outage.
Every stage runs on PLCs, VFDs, and BMS controls, so each one is an OT pathway to a compute outage.
Makeup water prep from utility or reclaimed sources: strainers for debris, media filters for solids, softening or RO.
Water chemistry control: corrosion and scale control, biocide dosing, side-stream filtration, and pH, ORP, and conductivity monitoring.
The central plant: chillers cool the loop, towers or dry coolers reject heat, with waterside economizers and thermal storage.
To IT equipment: variable-speed pumps, headers to data halls, CDUs to cold plates, and CRAH and rear-door heat exchangers.
What each system does, what controls it, the protocols on the wire, and what an attacker could do with it.
Takes high-voltage utility power, steps it down, and protects every downstream circuit with relays and breakers.
Start automatically on utility loss and carry the site while transfer and paralleling gear shifts the load.
Bridge the gap between utility loss and generator pickup, while battery storage smooths GPU power swings.
Deliver conditioned power from the electrical room to each rack — the last OT hop before the GPUs.
Reject heat from the facility water loop using chillers, cooling towers, economizers, and thermal storage.
Pump filtered coolant to GPU cold plates and isolate the rack loop from facility water.
Remove residual heat from air-cooled networking and storage gear, and hold hall temperature and humidity.
Detects smoke early and releases clean agent or pre-action water. Life safety overrides every other system.
Controls who enters the site, halls, cages, and OT rooms, and records what happens there.
The supervisory layer that monitors and commands systems 1 to 9 through HMIs, historians, and remote access.
Protocols vary by vendor and site. Passive discovery confirms what is actually on the wire.
Factory-integrated modules arrive with pre-programmed PLCs, remote-access paths, and default settings. Security requirements must enter at design and commissioning.
Campus siting follows grid capacity. The utility interconnect and substation come first.
A powered shell, with prefabricated power and cooling skids or modules.
Halls, racks, busway, manifolds, fiber, and OT network cabling installed.
Integrated systems testing, BMS/EPMS/SCADA integration, and handover.
A BMS compromise can cascade: lost cooling, overloaded circuits, false suppression release.
Researchers found 6,300 exposed ICS and building automation devices near US data centers, some on business-class internet outside IT governance.
Altered CDU, pump, or valve commands can starve racks of coolant and trip GPU clusters.
Manipulated EPMS, relay, or UPS data can hide overloads or force unplanned transfers and trips.
Maintenance links and gateways can bypass segmentation and reach BMS, EPMS, and CDU networks.
We start with one question: what could go wrong physically, and which cyber pathway makes it possible? Aligned to IEC 62443, NIST CSF, NIST SP 800-82, and ISO 27001.
Passive observation, documents, and interviews — never active scans of live control systems.
Map threats, barriers, and consequences for loss of cooling, loss of power, and false suppression discharge.
Add cyber causes to facility hazard reviews so reliability, safety, and security share one risk view.
Rate the maturity of zones and conduits across BMS, EPMS, SCADA, and CDU networks, with a multi-year roadmap.
SPAN/TAP capture builds an asset inventory of live 24/7 systems with zero packets injected.
You leave with a ranked risk list and a roadmap that facilities and security teams can both act on.
Agree on sites, systems, and stakeholders. Collect drawings and network documents.
Passive SPAN/TAP capture, document review, and interviews with facilities staff.
Bowtie and CyberPHA workshops, plus IEC 62443 and NIST gap ratings.
Ranked risks, OT asset inventory, and a multi-year remediation roadmap.
Sources: Uptime Institute survey, 2026; HPE GB200 NVL72 QuickSpecs, 2025; NVIDIA DGX GB200 guide; NVIDIA 800 VDC architecture, 2025–2026; TrendAI Research, An Invisible Attack Surface, 2026. Unfamiliar terms are defined in the acronym list below.
The terms facilities, controls, and security teams use when they talk about power, cooling, and building systems.
What could go wrong physically, and which cyber pathway makes it possible? Let’s scope your sites, systems, and stakeholders.
Schedule a Consultation