Industrial IoT (IIoT)
Data Center Growth Drives New Monitoring and Automation Technology
Published by IOT San Jose Research & Editorial Team

- Why Temperature Alone Stopped Being Enough
- From Dashboards to Automation
- The Power Problem Sits Underneath
- Where IoT Technology Fits In
- Business Value
- A Worked Example: Sizing the Telemetry for One Hall
- Implementation Advice
- Common Challenges
- Questions to Ask Before Signing a Monitoring Contract
- What Comes Next
The number that explains most of what is happening inside modern data centers is rack density. Industry analysts expect it to climb from roughly 20 to 40 kilowatts per rack toward 120 to 140 kilowatts in 2026 as AI hardware fills the halls. At 30 kilowatts, a cooling hiccup gives operators several minutes of slack. At 130 kilowatts, the slack shrinks to seconds, and the instrument that tells you a problem is developing has to be better than a thermostat on the wall.
That arithmetic is reshaping what data center operators buy. The market is not short of demand: CBRE reports primary-market vacancy fell from 3.3 percent in 2023 to under 2 percent in 2025, and roughly 80 percent of capacity under construction is already pre-leased. When every megawatt is spoken for, the competitive edge moves from building space to running it precisely. Monitoring and automation, long treated as back-office plumbing, are becoming a core part of the product.
Why Temperature Alone Stopped Being Enough
For decades a data center was monitored by watching air temperature and power draw. That worked when air carried away most of the heat. Dense AI racks increasingly use direct-to-chip liquid cooling, where coolant runs through plates mounted on processors. The US liquid cooling market is expected to exceed 2 billion dollars in 2026, and analysts describe direct-to-chip as the emerging dominant design, with immersion cooling relevant mainly for specific high-density edge deployments.
Liquid systems fail differently from air systems. Cooling vendors such as Baumer argue that a complete picture needs five measurements, flow, pressure, temperature, coolant quality, and system integrity. That grouping comes from sponsored vendor material, so treat it as one company's framing, but the logic holds up. A slow leak shows up as a pressure drop long before a temperature alarm. A clogged filter appears as reduced flow. Degraded coolant chemistry can corrode plates over months. None of these is visible to an air thermometer.
| What to watch | Typical sensor | Failure it reveals early |
|---|---|---|
| Coolant flow rate | Flow meter on each loop | Clogged filters, failing pumps, closed valves |
| Loop pressure | Pressure transducers | Leaks and air in the line |
| Supply and return temperature | Immersion or surface probes | Reduced heat transfer, fouled heat exchangers |
| Coolant quality | Conductivity and chemistry sensors | Contamination and corrosion risk |
| Leak presence | Rope or spot leak detectors | Fluid where it should never be |
| Power at rack and branch circuit | Metered rack power distribution units | Overloaded circuits and imbalance between phases |
From Dashboards to Automation
More sensors alone produce more alarms, which operators already ignore. The shift now underway is from watching to acting. Vendors of data center infrastructure management software, known as DCIM, are adding AI to forecast failures, automate maintenance scheduling, and tune power and cooling in real time. Digital twins let operators simulate a change, such as moving a block of GPU servers to another hall, before touching anything physical. We discussed the broader twin idea in our article on digital twins and predictive maintenance, and the logic transfers directly to a data hall.
A realistic example shows how the pieces fit. This is an illustrative scenario. A coolant distribution unit begins to show a slow decline in flow while pump power rises slightly. Neither reading crosses an alarm threshold. A model trained on the unit's history recognizes the pattern as an early sign of a partially blocked filter, opens a maintenance ticket, and shifts load to a redundant loop before temperatures change. A human approves the work for the next scheduled window. The outage that did not happen is invisible, which is exactly the point.
The Power Problem Sits Underneath
Cooling is half the story. Interconnection delays for new power are stretching to five to seven years in major markets, and the five largest hyperscalers are projected to spend as much as 602 billion dollars in 2026, according to one industry trends summary. Operators are responding with microgrids, behind-the-meter generation, and in some cases moves to second-tier markets. Where power is the scarce resource, wasting it is expensive, so energy monitoring at the rack and branch level becomes a financial tool.
This is especially relevant in Santa Clara County, which has one of the densest concentrations of data centers in the country. Many of those facilities were built for lower densities, and operators are retrofitting them for AI loads. Retrofit work favors wireless and clamp-on instrumentation that can be added without shutting down halls, a point that connects to our coverage of edge computing and data centers in San Jose.
Where IoT Technology Fits In
A data center is, in effect, an industrial plant that happens to house computers, and the tools look familiar to anyone who has read about industrial IoT and predictive maintenance. The new part is scale and speed. A single large hall can hold tens of thousands of sensor points, and the decisions have to happen in seconds. That pushes intelligence toward the edge of the facility, where gateways collect readings and run local models, a theme from our article on IoT gateways becoming edge AI computers. It also raises the importance of secure device management across thousands of instruments, covered in our piece on IoT device management.
Business Value
- Uptime. Early leak and flow warnings convert emergencies into scheduled maintenance.
- Capacity. Accurate, granular power data lets operators safely run closer to limits and place new racks where headroom exists.
- Efficiency. Tuning cooling to actual heat output cuts the energy spent on overcooling.
- Contracts. Customers buying capacity for expensive AI hardware ask for proof of environmental control, and good telemetry supports service-level commitments.
A Worked Example: Sizing the Telemetry for One Hall
Consider a 2 megawatt data hall holding 80 liquid-cooled racks. This is an illustrative calculation, not a described facility. Each rack has metered power at the branch level, supply and return temperature probes, a flow meter, and a leak sensor. Add pressure points on the coolant distribution units and the secondary loops, and the hall quickly passes 1,000 sensor points. Sampling each once per second produces roughly 86 million readings a day, most of which say nothing is wrong.
That volume is why edge processing matters. A local gateway can store raw data for a few days, compute rolling averages and rate-of-change values, and forward only exceptions and summaries to the central platform. The platform then compares each loop with its own history and with its peers, which catches the slow drift that fixed thresholds miss. Operators who skip this design usually end up with a data lake nobody queries and an alarm list nobody trusts. Operators who plan it well get a short daily list of items that deserve a human decision.
The same design helps during incidents. If the wide-area link fails, local controllers keep logging and enforcing safe limits, and the history is synchronized once the link returns. Testing that behavior before go-live is a cheap way to avoid an expensive surprise.
Implementation Advice
- Map the failure modes first. List the ways each system can fail and the earliest measurable sign of each. Buy sensors to cover those signs, not to fill a catalog.
- Instrument liquid loops as a system. Flow, pressure, temperature, and quality readings are far more useful together than any one alone.
- Normalize the data. Equipment from different vendors speaks different protocols. Agree on one data model before layering analytics on top.
- Automate in stages. Start with alerts and recommendations, then allow automated actions only for well understood, reversible steps.
- Secure the monitoring network. Instrumentation that controls cooling is a target. Segment it, authenticate devices, and keep firmware current.
Common Challenges
| Challenge | Why it appears |
|---|---|
| Alarm overload | Thousands of points with fixed thresholds generate noise that staff learn to ignore |
| Vendor fragmentation | Cooling, power, and security systems often use different protocols and platforms |
| Retrofit constraints | Live halls cannot be shut down, so installation windows are short |
| Trust in automation | Operators hesitate to let software adjust critical systems without a clear audit trail |
| Vendor-sourced claims | Much published material on cooling intelligence comes from suppliers and deserves independent testing |
Questions to Ask Before Signing a Monitoring Contract
Vendors will happily demonstrate dashboards. Better questions concern what happens after installation. Ask how the system handles sensors from other manufacturers, how long raw data is retained, and whether the analytics can be exported if you change platforms. Ask who owns the models trained on your facility's data. Ask for a reference customer with a liquid-cooled deployment of similar density, and speak to their operations staff, not just their executives. Finally, ask how the product behaves when its cloud connection is down, because a monitoring system that fails during a network outage fails at the worst moment.
What Comes Next
Expect liquid cooling instrumentation to become standard equipment, expect DCIM platforms to compete on how well their automation can be trusted, and expect power data to move from the facilities team to the finance team. The operators who invest early in clean, well-organized telemetry will be able to add AI-driven control gradually and safely. Those who delay will find that denser racks give them less time to learn. For related thinking on the compute side, see our article on NVIDIA's physical AI push.
