For years, data center design had a fairly stable center of gravity. Rack densities moved up, but not so violently that the entire facility architecture had to be reconsidered every hardware cycle. A 5 kW rack was normal. A 10 kW rack was meaningful. A 20 kW rack required more planning, but it still fit inside the mental model most engineers, owners, and contractors already understood.

AI changed the scale of the problem.

The conversation is no longer about whether a data hall can support 10 kW or 20 kW racks. The market is now dealing with 100 kW rack-scale systems, 132 kW liquid-cooled NVIDIA GB200 NVL72 systems, and emerging design roadmaps that point toward hundreds of kilowatts per rack. NVIDIA describes the GB200 NVL72 as a rack-scale, liquid-cooled system connecting 36 Grace CPUs and 72 Blackwell GPUs in a single 72-GPU NVLink domain, and NVIDIA’s own DGX GB system guide lists rack power consumption at approximately 120 kW. Supermicro’s GB200 NVL72 configuration lists total power at 132 kW and operating power between 125 kW and 135 kW.

That is not incremental density growth.

That is a structural reset.

Once racks move into the 100 kW range, every old facility assumption starts getting tested at the same time. Power distribution changes. Cooling changes. Cable management changes. Floor loading changes. Switchboard sizing changes. Fault current exposure changes. The rack is no longer just an IT endpoint. It becomes an industrial load, and the entire electrical path from utility to rack has to be designed accordingly.

This is the point many projects miss. They treat rack density as a server-room issue, when it is actually a whole-building infrastructure issue.

The rack is becoming the new unit of infrastructure

In traditional data centers, the rack was important, but it was rarely the organizing unit for the entire facility. Power design started with utility service, transformers, switchgear, UPS, distribution, PDUs, and then eventually the rack. The rack was the destination, not the architecture driver.

AI flips that sequence.

The facility now has to be designed backward from the rack. If the rack is going to draw 100 kW, 132 kW, or eventually several hundred kilowatts, then everything upstream has to be revalidated. It is not enough to say the site has enough megawatts. The question is whether those megawatts can be delivered into a small physical footprint safely, efficiently, and repeatedly.

Introl’s 2026 high-density rack analysis frames the change bluntly: traditional racks designed for 5 to 10 kW loads cannot safely support modern GPU server power requirements without fundamental architectural changes. The same analysis notes that AI facilities often need 40 to 250 kW per rack, and that 100 kW-capable infrastructure may cost $200,000 to $300,000 per rack, while retrofitting existing facilities to 40 kW density can cost $50,000 to $100,000 per rack.

Those numbers matter because they show that rack density is not a server purchasing decision. It is a capital planning decision.

A project team can buy next-generation GPUs and still fail if the upstream infrastructure cannot deliver power, remove heat, handle weight, route fiber, and maintain serviceability. At AI density, stranded capacity does not only mean unused floor space. It can mean unusable compute sitting inside a facility that technically has room, but does not have the electrical or thermal pathway to support it.

That is the expensive version of “almost ready.”

Open Rack V3 and Open Rack Wide are signs of where the industry is going

The emergence of Open Compute Project rack standards is not a niche standards-body exercise. It is a signal that hyperscale infrastructure is being redesigned around AI from the rack outward.

Open Rack V3 was built around interoperability and scale. The OCP Open Rack V3 base specification says Open Rack provides the structure and power for interoperability of Open Rack-based IT systems, and it defines requirements for components within the Open Rack Frame V3 family so that IT gear and power systems developed by the community can intermate. It also includes 48 V busbar geometry and rack-to-busbar interfaces, which shows how deeply rack structure and power delivery are now linked.

That is a big conceptual difference from the traditional view of a rack as a metal frame filled with equipment. In the OCP model, the rack is part of the electrical architecture. Power shelves, busbars, battery backup modules, and rack geometry are all connected.

Delta’s 33 kW Open Rack V3 power shelf is a good example of the direction. OCP describes the shelf as delivering over 97.5% power efficiency during AC-to-DC conversion, with 48 V or 50 V DC output, higher than traditional 12 V DC, to reduce distribution losses for OCP ORv3 racks and high-power AI servers.

Now the industry is moving beyond Open Rack V3 into Open Rack Wide. S&P Global’s coverage of the 2025 Open Compute Project Summit notes that AI requirements are pushing scale both logically and physically, with server rack width doubling in the Open Rack Wide specification to support greater density and better serviceability. It also notes that liquid cooling has moved from “nice to have” to required, and that energy delivery is shifting as higher-voltage DC power enters the conversation.

That matters for Moonshot because rack standards are no longer just about the rack. They are about the interface between compute, cooling, and power distribution. As the rack gets wider, heavier, hotter, and more power-hungry, the facility’s electrical system has to become more intentional.

The days of treating rack density as a downstream IT variable are over.

The real stress shows up upstream

When a rack moves from 10 kW to 100 kW, the visible change is inside the data hall. The real stress, however, shows up upstream.

A 100 kW rack pulls the same load as ten traditional 10 kW racks, but concentrated into one location. That concentration affects feeder sizing, busway strategy, PDU selection, floor layout, heat rejection, cable routing, and maintenance access. If several rows of racks are all in that range, the data hall starts behaving less like a conventional IT environment and more like a dense industrial process load.

The electrical implications compound quickly.

Higher rack density means higher total current in localized distribution paths. Larger power blocks can increase available fault current. More aggressive transformer sizing can push interrupting and withstand requirements upward. Higher continuous load means bus temperature rise matters more. More power electronics can increase harmonic considerations. More liquid cooling equipment adds pumps, CDUs, controls, leak detection, and service access into the infrastructure model.

NVIDIA’s DGX GB system guide illustrates how tightly these systems are now integrated. The rack includes compute trays, NVLink switch trays, management switches, power shelves, liquid cooling manifolds, cable cartridges, and a power bus bar. Power shelves convert AC power into nominal 50 V to 51 V DC and distribute it through the bus bar to rack components, with multiple shelves used for redundancy.

That kind of rack is not simply plugged into a facility. It is integrated into one.

This is where traditional distribution assumptions begin to fail. A facility designed for average utilization, moderate rack diversity, and conventional air-cooled layouts may not survive the jump to dense, liquid-cooled AI clusters without major redesign. The rack may be ready for the workload. The facility may not be.

The most dangerous situation is when the first phase works just well enough to hide the problem.

A deployment starts with lower-density racks. The electrical system passes inspection. The owner plans to add higher-density AI hardware later. Three years in, the next hardware cycle arrives and the team discovers that the upstream switchboards, distribution paths, heat rejection systems, and floor layouts were never designed for the future state. The space is available, but the infrastructure is not.

That is stranded capacity in its most frustrating form.

The Moonshot connection: utility-to-rack design has to survive the next hardware cycle

Moonshot sits in the part of the system that becomes more important as rack density rises: the critical layer between utility power and the rack.

That layer includes switchboards, PDUs, transformers, modular data center infrastructure, rack systems, and the engineering decisions that connect all of it. Moonshot’s UL891 switchboard page emphasizes engineering around supply chain constraints and flexibility, the ability to use major breaker manufacturers, and site design evaluation to provide code-compliant, cost-effective, fast solutions. Moonshot’s rack systems also include pre-wired rack infrastructure, mounted PDUs, hot aisle isolation, and power distribution integration. Its modular data center work includes NEC and UL compliance, 415/240 V systems, electrical panels with main disconnects, smart PDUs, monitoring, sensors, power management, and UL field listing options.

That is the right footprint for this moment.

The market does not need another vendor saying “we support high density.” It needs partners who can translate high-density compute requirements into actual electrical infrastructure. That means asking practical questions early, before the project is boxed in:

Can the service and distribution architecture support future rack densities, or only the first phase? What happens to available fault current if the owner increases transformer capacity? Does the switchboard strategy account for sustained load instead of historical diversity? Are PDUs, bus systems, and downstream distribution being selected around the next generation of racks or the current one? Is cooling equipment being planned as part of the electrical load model? Are liquid-cooled racks, CDUs, pumps, controls, and monitoring included in the real infrastructure plan?

Those questions cannot be answered at the end of procurement. They need to be answered when the architecture is still flexible.

This is where Moonshot can position itself clearly. Moonshot is not just building equipment to match today’s drawings. It is designing infrastructure that can survive the next five to ten years of density increases. In AI, that matters because the compute lifecycle is moving faster than the electrical lifecycle. Servers can turn over quickly. Racks can be upgraded. GPU architectures can change. But switchboards, transformers, PDUs, and modular infrastructure are expected to remain in place much longer.

If the electrical backbone is designed only for today’s rack density, tomorrow’s compute plan may be dead on arrival.

The real takeaway

The jump from 10 kW racks to 100 kW racks is not a bigger version of the same problem. It is a different problem.

At 10 kW, a rack is an IT load. At 100 kW, it starts to behave like a major electrical asset. At several hundred kilowatts, it becomes a power, cooling, structural, and serviceability event. At 1 MW, it forces a new architecture.

That is why Open Rack V3, Open Rack Wide, 48 V and 50 V DC rack power, 800 VDC discussions, liquid cooling manifolds, high-capacity CDUs, and rack-scale AI systems all belong in the same conversation. They are not separate trends. They are all symptoms of the same underlying shift: AI is compressing more compute into less space, and the infrastructure around that compute has to become more integrated.

For data center owners, the risk is designing around the rack density they have today instead of the rack density they will need next. For engineering teams, the risk is treating rack density as a downstream equipment issue instead of an upstream distribution issue. For procurement teams, the risk is buying infrastructure that clears the first submittal but cannot support the second phase.

Moonshot’s opportunity is to own that gap.

The company already operates at the intersection of switchboards, PDUs, rack systems, modular data centers, transformers, and high-density power distribution. That is exactly where AI infrastructure is heading. The winner in this market will not be the company that simply builds equipment quickly. It will be the company that helps customers design infrastructure that remains useful as rack density continues to climb.

If you are planning AI infrastructure, the question is not only whether your facility can energize the first racks.

The better question is whether your power architecture will still make sense when those racks double, triple, or move into the next density class.

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

Because in AI infrastructure, the rack is no longer the endpoint.

It is the thing the entire power system is being designed around.

Sources:

https://introl.com/blog/high-density-racks-100kw-ai-data-center-ocp-2025

https://www.opencompute.org/documents/open-rack-base-specification-version-3-pdf

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

https://www.nvidia.com/en-us/data-center/gb200-nvl72/

https://www.opencompute.org/products/430/delta-33kw-1ou-open-rack-v3-power-shelf

https://www.spglobal.com/market-intelligence/en/news-insights/podcasts/next-in-tech/next-in-tech-ep-242-open-compute-project-summit

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

https://moonshotus.com/rack-systems/

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