AI racks turned data centers into plumbing systems

For most of the industry’s history, a server rack was an electrical and airflow problem. AI changed that. The densest racks now carry coolant to every chip, and the white space has become a plumbing system that has to be watched like one.

From 8 kW to 132 kW

Uptime Institute’s 2024 survey put average rack density below 8 kW, and relatively few facilities had racks above 30 kW. The reference design Schneider Electric developed with NVIDIA for the GB200 NVL72 runs at 132 kW per rack, and it is liquid cooled.

Air can’t carry that much heat away from a single rack. So coolant goes where the heat is: through cold plates on the processors, fed by manifolds and hoses at the rack, and circulated by coolant distribution units (CDUs) on the floor.

What’s new in the white space

A liquid-cooled AI hall adds parts that most data centers never had next to IT equipment:

  • CDUs with pumps, heat exchangers, filters and expansion tanks.
  • Manifolds and hoses running to every rack.
  • Quick disconnects at each server, handled every time hardware is moved or swapped.
  • Cold plates with fine internal channels that clog if the coolant carries particles.

Every joint, seal and pump is a place where liquid can escape or flow can fall. Multiply that by every rack in the hall.

The failure modes are already documented

In May 2026, ASHRAE’s technical committee for data centers (TC 9.9) published a bulletin on keeping direct-to-chip loops healthy. It opens with a blunt line: “It is too easy to violate direct-to-chip cooling system operational and warranty requirements.” The four problems it highlights:

  • Coolant drift. When the fluid moves away from its design mix, pressure drop, pump behavior and heat transfer all change.
  • Trapped air. Air cuts heat transfer, makes pumps cavitate and speeds up corrosion. It can show up weeks after startup.
  • Damaged stainless steel. Welding and grinding strip the protective film, and loose iron can foul cold-plate channels.
  • Pressure ratings. Loops need to be designed for every condition, including filling, flushing and connecting or disconnecting hardware.

None of these starts as a puddle. Each one starts as a small change in pressure, flow, vibration or chemistry.

Why the stakes are higher

Outages are already expensive. In Uptime Institute’s 2025 outage analysis, 54% of operators said their most recent significant outage cost more than $100,000, and 20% said it cost more than $1 million. Liquid next to power and IT equipment adds a new way to get there, as the 2023 Google Cloud outage in Paris showed.

Monitoring has to change with the hardware

Air-era monitoring watches temperature and power. A liquid-cooled hall also needs to watch:

  • Pressure at pumps, filters and across cold plates.
  • Flow through each loop and branch.
  • Vibration on pumps and valves.
  • Coolant chemistry, such as conductivity and pH.

Fixed alarm thresholds catch a failure once it is obvious. The slow drift that comes first, like a seal starting to weep or a filter loading up, stays inside the limits until it doesn’t. Catching it means learning what normal looks like for each loop and flagging the change early.

That is what MayimFlow is built to do: learn each loop’s normal rhythm and flag the early signs of a liquid failure. See how it works, or read leak detection vs. leak prediction.

Sources

  1. Upsite Technologies: Rack density rises while PUE and outage frequency remain flat (Uptime Institute 2024 survey, Jan 2025)
  2. DCD: Schneider Electric co-develops GB200 NVL72 reference architecture with Nvidia (Dec 2024)
  3. StorageReview: ASHRAE TC 9.9 flags four ways direct-to-chip cooling loops go wrong (Aug 2026)
  4. Data Center Knowledge: Data center outages decline for fourth straight year (Uptime Institute Annual Outage Analysis 2025)

See what MayimFlow would watch in your facility.

Predictive liquid-risk intelligence for AI data centers and infrastructure.