For most of the data center industry’s history, cooling was treated as a facility system that supported the IT load. Servers produced heat, mechanical systems removed it, and electrical teams made sure enough power was available to keep everything running. The two systems were connected, obviously, but they were usually planned as parallel workstreams.

That separation is starting to fail.

AI infrastructure is forcing power and cooling into the same design conversation. A rack that pulls 10 kW can be cooled, powered, and operated within a familiar data center model. A rack that pulls 100 kW is a different machine. A rack-scale AI system pushing beyond 120 kW is not just a denser IT cabinet. It is a tightly integrated electrical, mechanical, controls, and safety problem sitting in the middle of the data hall.

NVIDIA’s DGX GB rack-scale systems make this clear. The rack is not just a set of servers. It includes compute trays, NVLink switch trays, power shelves, a bus bar, and liquid cooling manifolds. The compute trays are cooled by liquid flowing through manifolds and cold plates attached to CPUs and GPUs, while other components remain air cooled. NVIDIA’s documentation lists rack power consumption at approximately 120 kW and describes power shelves converting AC into nominal 50 V to 51 V DC, then distributing power through a rack bus bar. In other words, power and cooling are already co-designed inside the rack. The facility has to catch up.

This is why liquid cooling is not just a mechanical upgrade. It changes the electrical architecture, the reliability model, the safety plan, the monitoring strategy, and the way switchboards, PDUs, bus systems, and modular infrastructure need to be designed.

Cooling is becoming the system constraint

AI has made compute denser, but density has a consequence: heat becomes concentrated in a smaller footprint. Data Center Knowledge reported from Data Center World 2026 that as AI workloads push rack densities toward the megawatt era, cooling is no longer merely a supporting system. The discussion has shifted from “moving heat out of a room” to designing a connected system spanning silicon, fluid, controls, and workload scheduling.

That framing matters. Traditional air cooling was built around moving large volumes of air through rooms, rows, aisles, and racks. It works well across a broad range of legacy and moderate-density workloads, and it will remain important for years. The problem is that air has limits. As rack power climbs, the amount of airflow required becomes harder to deliver, harder to control, and more expensive to operate.

Uptime Institute’s analysis of NVIDIA Blackwell rack systems shows how quickly the problem escalates. The flagship NVL72 system is rated at 132 kW in a rack-scale configuration. Uptime noted that only about 1% of data center operators in its 2024 survey reported racks above 100 kW, and that a few rows of Blackwell systems can consume as much power as a sizable data hall full of hundreds of normal racks. At full load, Uptime reported that cold plates handle around 100 kW of thermal power, while the air exhaust still exceeds 25 kW.

That last point is the trap. Liquid cooling does not automatically eliminate air cooling. It changes the balance. Even with direct liquid cooling handling the highest heat sources, the facility still has air-side heat from networking, storage, power supplies, and other components. That means the future is not simply “air or liquid.” It is hybrid.

Data Center Knowledge reached the same conclusion in its coverage of AI cooling strategy. Industry leaders at Data Center World pointed to hybrid air-liquid architectures as the near-term steady state, with single-phase liquid cooling expected to support the next several generations of chips while two-phase systems continue to develop.

That matters for data center owners because hybrid systems are more complex than either air-only or liquid-only design. They require coordination between electrical loading, coolant flow, CDU capacity, heat rejection, controls, telemetry, and failure response. If the facility treats liquid cooling as an add-on after the electrical system is already designed, it will create avoidable risk.

Liquid cooling changes the failure model

Air cooling has inertia. If a fan slows down or an air handler trips, operators often have some thermal buffer before equipment reaches critical limits. That buffer may be short, but it exists. Liquid cooling changes the response time.

Data Center Knowledge reported that while air-cooled environments may tolerate minutes of disruption before reaching thermal limits, liquid-cooled systems can tolerate only seconds. That forces new requirements around thermal buffering, continuity on the mechanical side, and tighter integration with power systems.

That is the part many electrical teams need to internalize. Once cooling becomes a seconds-level reliability issue, the electrical system supporting the cooling infrastructure is no longer auxiliary. Pumps, CDUs, controls, leak detection, valve actuation, sensors, and monitoring systems become uptime-critical loads.

If a rack depends on coolant moving through cold plates, then the power feeding the cooling loop is part of the compute availability chain. It is not just a mechanical support load. It is part of the mission-critical path.

That has practical consequences. Electrical design now has to account for cooling continuity in the same way it accounts for compute continuity. The power system must support not only the GPU load, but also the equipment keeping those GPUs thermally stable. Redundancy strategy, breaker coordination, UPS support, monitoring, control wiring, and emergency shutdown logic all need to consider what happens when cooling equipment loses power or behaves abnormally.

Leak detection also becomes part of the infrastructure conversation. NVIDIA’s rack-scale documentation states that leak detection in the DGX GB liquid cooling system helps prevent equipment damage, maintain reliability and uptime, protect data integrity, preserve safety, and avoid costly repairs and downtime.

That is not a small operational detail. It is a sign that the rack itself is becoming a combined power, fluid, and controls system.

The design vocabulary is expanding

The industry is still working through which liquid cooling architectures will dominate. Direct-to-chip cooling, immersion cooling, rear-door heat exchangers, in-row CDUs, in-rack CDUs, and hybrid cooling blocks all solve different problems.

Direct-to-chip cooling places cold plates on the highest-heat components, typically CPUs, GPUs, and accelerators. Coolant removes heat at the source before it becomes a room-level air problem. This is the architecture most closely tied to current high-density AI systems.

Rear-door heat exchangers use cooling at the back of the rack to remove heat from exhaust air before it enters the room. They can support higher densities while preserving more of the traditional server configuration, but they still depend on airflow and rear access.

Immersion cooling places equipment in dielectric fluid. It can support very high heat densities, but it changes serviceability, maintenance practices, equipment compatibility, and operating procedures.

CDUs, or coolant distribution units, become the bridge between the facility water system and the technology cooling system. They control flow, temperature, pressure, filtration, and separation between loops. Once CDUs enter the design, the data center has new equipment that must be powered, monitored, protected, and physically integrated into the layout.

This is why OCP and ASHRAE formed a 2025 alliance focused on data center liquid cooling standards and best practices. Their announcement specifically references the intersection between facility water systems, technology cooling systems, direct-to-chip cooling, immersion cooling, CDUs, door heat exchangers, power usage effectiveness, and sustainability. It also notes that OCP’s cooling-focused community is working on Cooling Environments, Cold Plate, Immersion, CDU, Door Heat Exchanger, Technology Cooling Systems, Data Center Facilities, and Sustainability.

That is the industry telling us something important: liquid cooling is not a product category. It is an infrastructure layer.

CoreSite’s 2026 outlook points in the same direction. It reports that experts expect liquid cooling to ramp up in 2026 because AI power and thermal requirements are making it mainstream, and that skidded modular liquid cooling units starting at 2 MW, and going beyond that, are expected to become common for high-density data center builds by late 2026. CoreSite also notes that new two-phase direct-to-chip solutions are expected as rack densities climb up to and beyond one megawatt.

A 2 MW cooling unit is not something you “bolt on” casually. It has electrical load, controls, physical clearance, service access, water connections, sensors, alarms, and failure modes. If the electrical infrastructure is not designed around that reality early, the project team ends up forcing a high-density cooling architecture into a low-density planning process.

That is how expensive redesign happens.

The electrical layout changes with the thermal strategy

Once liquid cooling becomes part of the data hall, electrical layout decisions become more sensitive.

A traditional electrical plan might focus on service entrance, switchboards, UPS, PDUs, rack power, feeder routing, and redundancy. In a liquid-cooled AI environment, that same plan now has to account for coolant distribution, CDU placement, leak detection zones, maintenance access, pipe routing, containment, pump power, control panels, and mechanical clearances.

The rack row is no longer just a power and network row. It is a power, network, fluid, and controls row.

That creates tradeoffs. Locating CDUs close to the load can improve efficiency and simplify coolant delivery, but it may affect electrical room layouts, aisle spacing, floor loading, and service access. Rear-door heat exchangers can extend air-cooled infrastructure, but they change rear clearance and plumbing requirements. Immersion systems can support extreme density, but they introduce different physical footprints and maintenance procedures. Direct-to-chip cooling may reduce room-level heat load, but it increases dependency on pumps, manifolds, quick connects, leak sensors, and controls.

All of this changes how switchboards and PDUs should be planned. The electrical system must support not only the IT load but the cooling infrastructure that makes the IT load survivable. It must also account for how those loads behave together. If AI workloads ramp quickly, thermal demand can ramp with them. If workload scheduling becomes more dynamic, cooling systems may need to pre-stage capacity. Data Center Knowledge reported that operators are increasingly aggregating telemetry into centralized platforms and that the next step is bidirectional control, where workload signals inform cooling behavior in real time.

That is where the industry is heading: coordinated power, thermal, and compute control.

The old model had layers reacting to each other. The new model requires the building to act like one system.

Where Moonshot fits

This is a major positioning opportunity for Moonshot because Moonshot already sits at the intersection of power distribution, modular infrastructure, controls, and data center deployment.

Moonshot’s modular data center offering is manufactured in Texas and designed around NEC and UL compliance. Its standard and integrated features include electrical panels with main disconnects, Moonshot PDUs, pre-wired network switches, filtered intake walls, high-CFM exhaust fans, smart PDUs, interior and exterior temperature sensors, differential pressure sensors, VFD fan motor speed control, automated system monitoring, reporting, and power management.

That matters because AI data centers increasingly need companies that understand infrastructure as a system, not as isolated pieces of gear.

A company that only sees the electrical one-line may miss the cooling dependencies. A mechanical vendor that only sees heat rejection may miss fault current, breaker coordination, UL listing implications, and power distribution constraints. A rack vendor may focus on IT hardware integration but not the facility-level consequences. Moonshot has the opportunity to sit across those boundaries.

Moonshot’s UL891 switchboard page also emphasizes that its engineers design solutions around supply chain constraints and flexibility, can use major breaker manufacturers, and evaluate site design to provide code-compliant, cost-effective, fast solutions.

That combination is important. High-density AI projects need electrical infrastructure that is code-compliant, flexible, fast to deploy, and designed around real operating conditions. Liquid cooling adds one more reason why early engineering matters. If switchboards, PDUs, modular data center blocks, rack systems, sensors, and cooling equipment are all being coordinated separately, the project is exposed. If they are coordinated early, the system has a much better chance of working as intended.

Moonshot should not position itself as “the company that does cooling.” That is too broad and not the point. The stronger positioning is this:

Moonshot understands how high-density power distribution has to interact with thermal infrastructure.

That is a credibility gap against competitors that only talk in terms of amperage, breaker availability, or lead time. In AI infrastructure, the buyer does not just need a switchboard. They need an electrical backbone that supports a thermal strategy that may evolve over multiple hardware generations.

That means asking questions early:

  • Where will CDUs sit relative to racks, PDUs, and service access?
  • Are cooling support loads treated as critical loads in the power model?
  • What happens to compute availability if a pump, CDU, or control panel loses power?
  • How are leak detection, alarms, sensors, and emergency shutdown logic integrated?
  • Does the electrical distribution design account for future cooling upgrades?
  • Are modular power and cooling blocks being designed together or forced together later?

These are not academic questions. They determine whether a high-density deployment is resilient or fragile.

The real takeaway

Liquid cooling is not replacing electrical design. It is making electrical design more important.

As AI racks move from tens of kilowatts to hundreds of kilowatts, the data center can no longer separate the power conversation from the cooling conversation. The rack is pulling more power, generating more heat, and depending on more integrated support systems than traditional infrastructure was built to handle. Cooling equipment now has to be treated as part of the critical power chain. Power distribution has to be designed with thermal behavior in mind. Controls, sensors, leak detection, and redundancy planning have to be coordinated from the beginning.

That is the convergence.

The winner in AI infrastructure will not be the team that simply delivers the biggest electrical service or the most advanced cooling technology. It will be the team that can make power, cooling, controls, and modular deployment work together without creating hidden failure points.

For Moonshot, this is exactly the kind of conversation that elevates the company beyond commodity gear. Moonshot can help customers think through the actual infrastructure required to support high-density AI environments, from UL891 switchboards and PDUs to modular data center systems, monitoring, power management, and future scalability.

If you are planning a liquid-cooled AI deployment, the question is not only whether the cooling system can remove the heat.

The better question is whether the power infrastructure has been designed to keep the entire thermal system alive, coordinated, and ready for the next density jump.

https://moonshotus.com/request-form/

Because in AI data centers, cooling is no longer downstream of power.

It is part of the power system.

Sources:

https://www.datacenterknowledge.com/cooling/cooling-becomes-the-system-constraint-as-ai-reshapes-data-center-design

https://www.coresite.com/blog/data-center-outlook-2026-power-and-cooling-challenges-and-solutions-are-top-of-mind

https://journal.uptimeinstitute.com/density-choices-for-ai-training-are-increasingly-complex/

https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html

https://www.opencompute.org/blog/open-compute-project-foundation-and-ashrae-form-new-alliance

https://moonshotus.com/modular-datacenters/

https://moonshotus.com/ul-891-switchboards/