Crusoe Just Ordered 5 GW of AI UPS: What That Deal Tells Us About the Future of Backup Power

The State of the Union: AI Has Turned Backup Power Into Grid Infrastructure

Crusoe and ON.energy announced on July 21, 2026, that they will deploy 5 gigawatts of AI UPS™ technology across multiple hyperscale campuses. Commissioning is expected to begin in 2026 and continue into 2027, according to Crusoe’s announcement.

That is not a routine UPS purchase. It is a signal that backup power is moving from a facility-level safeguard to a strategic component of the electrical grid. AI campuses can contain tens or hundreds of megawatts of rapidly changing GPU demand. During synchronized training workloads, thousands of processors may ramp up or down together, creating power fluctuations that traditional data center designs were not built to absorb.

The public announcement describes a partnership and planned deployment rather than a conventional equipment purchase order. The distinction matters, but the practical message is clear: hyperscale operators now view power protection, voltage ride-through, load smoothing, and grid compatibility as one integrated engineering problem.

For data center operators, the 5 GW figure is less important as a headline than as a planning benchmark. It demonstrates how quickly the definition of “backup power” is expanding: and why operators of smaller enterprise facilities should reassess their own power architecture before AI-driven loads expose its limits.

Why Now: The Status Quo Is Failing

Traditional UPS systems were primarily designed to bridge a utility outage, filter disturbances, and provide enough runtime for generators to start or systems to shut down safely. That model still works for many conventional server rooms. It is not sufficient by itself for high-density AI environments.

Latency is now a power-quality issue

AI workloads are sensitive to interruptions that may be too brief to register as a conventional outage. A voltage sag, frequency deviation, or momentary fault can interrupt GPU jobs, trigger hardware protection, or force a distributed workload to restart.

The operational impact is not limited to a few unavailable servers. A failed training run can consume hours or days of compute time. In cloud environments, an electrical disturbance can also affect service-level agreements, customer workloads, and downstream applications.

A modern UPS strategy must therefore address both runtime and response time. Operators should evaluate:

• Transfer and ride-through performance during voltage faults
• Inverter response to fast load changes
• Compatibility with generators, switchgear, and on-site generation
• Battery autonomy at the actual critical load
• Monitoring of voltage, frequency, harmonics, and battery condition

Redundancy must include the power path

A Tier III facility is designed for concurrent maintainability, while a Tier IV facility targets fault tolerance through independent and physically separated systems. Neither designation eliminates the need to validate how AI loads behave inside the power architecture.

An operator may have redundant UPS modules but still depend on a common upstream transformer, switchboard, control network, or cooling system. For AI facilities, redundancy should be analyzed across the complete chain:

1. Utility and on-site generation
2. Medium-voltage distribution
3. Transformers and switchgear
4. UPS modules and energy storage
5. Power distribution units and rack-level distribution
6. Monitoring, controls, and communications
7. Cooling and thermal management

A redundant design that cannot isolate a fault is not resilient. The Crusoe-ON.energy model emphasizes this system-level view by placing medium-voltage UPS infrastructure inline between power sources and the data center.

Thermal Management changes the electrical design

AI racks can consume far more power than conventional enterprise racks. Depending on the GPU platform and configuration, current deployments may range from approximately 30 kW to more than 100 kW per rack, with future designs pushing higher. Some AI data halls are planned around 1 MW or more per rack cluster, rather than treating the individual rack as the primary unit of capacity.

That density affects everything from busway sizing to battery discharge profiles. Liquid cooling adds pumps, heat exchangers, controls, and potentially a second class of critical loads. If cooling fails, the IT load may need to be curtailed even when electrical capacity remains available.

Power and cooling should therefore be modeled together. A UPS system sized only for the nameplate load of the GPUs may not provide enough capacity for pumps, controls, network equipment, and the operational margin required during a thermal event.

What the 5 GW Deployment Tells Us

1. The UPS is becoming a load-management platform

The next generation of UPS systems will do more than supply battery power during an outage. They will help control how the facility interacts with the grid.

A medium-voltage UPS can provide a buffer between an unstable or constrained power source and a highly dynamic AI load. It can help the facility remain online during disturbances while presenting a more predictable electrical profile to the utility.

This is particularly relevant in regions where large-load interconnection requirements are becoming stricter. Crusoe and ON.energy state that the AI UPS™ has been validated against ERCOT large-load interconnection requirements, including NOGRR 282 and low-voltage ride-through protocols. Operators should not assume that one vendor’s validation automatically applies to their project, but they should expect grid-code compliance to become part of the power protection specification.

2. Energy storage will be segmented by function

A single battery system may not be the right answer for every power event. Data center operators increasingly need to separate energy storage by purpose:

• Milliseconds to seconds: ride-through for voltage disturbances and switching events
• Minutes: bridge power while generators start or loads are managed
• Longer durations: support extended outages, peak shaving, or grid services
• Rack-level protection: protect individual IT loads from localized faults

Lithium-ion batteries are common in modern UPS installations because of their energy density and monitoring capabilities. Lead-acid remains relevant in some applications because of established maintenance practices and lower initial cost. Flow batteries, iron-air systems, and other long-duration technologies may serve different applications than the short-duration UPS battery string.

The correct architecture depends on the event profile, not simply the desired runtime.

3. Monitoring will become as important as hardware

A 5 GW deployment cannot be managed through periodic inspections and local LCD panels. Operators need real-time visibility into battery state of health, module availability, load balance, power quality, thermal conditions, and alarm history.

Remote monitoring also changes how maintenance is performed. Predictive analytics can identify battery degradation, abnormal temperature rise, repeated voltage events, or declining module efficiency before a failure occurs.

The most effective approach combines equipment telemetry with facility-level context. A battery alarm should be correlated with load changes, cooling conditions, generator status, and utility events. That is where AI-driven analytics can provide an advantage: not by replacing engineering judgment, but by surfacing relationships that are difficult to identify manually.

The AI UPS Roadmap

Facility managers do not need a 5 GW campus to apply the lessons from this deployment. The following steps can be started today.

1. Build a time-based load profile.
   Record normal, peak, idle, startup, and rapid-ramp conditions. For AI equipment, a static nameplate estimate is not enough. Measure load changes over seconds and minutes, including the impact of cooling systems.

2. Map every critical power path.
   Document utility feeds, generators, transformers, switchgear, UPS modules, batteries, PDUs, rack distribution, and control networks. Identify single points of failure and common-mode dependencies. Compare the result with the resiliency objectives of a Tier III or Tier IV design.

3. Set measurable UPS requirements.
   Specify more than kVA. Define target efficiency: many modern double-conversion UPS systems advertise efficiency in the 95% to 99% range, depending on operating mode and load: and require documentation for efficiency at the facility’s expected load range. Include voltage ride-through, overload capacity, short-circuit performance, bypass operation, harmonic limits, and battery autonomy.

4. Create a battery lifecycle plan.
   Track battery age, temperature, impedance, state of charge, and state of health. Budget for replacement, disposal, testing, and firmware updates before the end of useful life. A battery system is not resilient if its condition is unknown.

5. Connect monitoring to operational response.
   Establish alert thresholds, escalation procedures, and maintenance workflows. Remote monitoring should reach the people responsible for facilities, IT, cybersecurity, and executive continuity: not just an unattended dashboard. Segment monitoring networks appropriately and control vendor access.

Ace Real Time Solutions helps organizations translate these requirements into practical designs. Its power delivery and protection services include UPS systems, surge protection, voltage regulation, transformers, automatic transfer switches, PDUs, installation, maintenance, and emergency support. For institutional, enterprise, and government buyers, the power protection procurement page provides a starting point for coordinated sourcing.

What This Means for Enterprise Data Centers

Most enterprise operators will not deploy medium-voltage AI UPS systems at hyperscale. They may, however, install GPU clusters, support accelerated computing, or expand workloads that create more volatile power demand.

The practical response is not to purchase the largest UPS available. It is to align the design with the failure modes that matter:

• If the risk is short utility disturbances, prioritize ride-through and power quality.
• If the risk is generator startup, validate runtime and transfer coordination.
• If the risk is a single power path, invest in maintainable redundancy.
• If the risk is battery failure, improve telemetry and lifecycle management.
• If the risk is high rack density, coordinate electrical and cooling capacity.

Product selection should also reflect the operational environment. Ace supports solutions from major manufacturers including APC by Schneider Electric, CyberPower, Vertiv, and Minuteman Technologies. APC’s portfolio, for example, includes Smart-UPS systems with network connectivity and service bypass equipment for maintainability. Larger projects may require custom UPS, battery, rack, cable-management, airflow, and remote-control architectures rather than catalog equipment alone.

The standard for modern infrastructure is no longer simply “the lights stay on.” The standard is predictable, observable, maintainable power under changing load conditions.

The Bottom Line

Crusoe’s planned 5 GW AI UPS deployment shows where critical power is headed. Backup power is becoming an active layer of the data center’s electrical and grid strategy.

For operators, the message is immediate:

• Design for dynamic load behavior, not only peak load.
• Treat medium-voltage distribution and UPS protection as one system.
• Coordinate power, cooling, and controls.
• Measure efficiency and ride-through performance under real operating conditions.
• Use remote monitoring to turn maintenance from a calendar task into a risk-management process.

Whether you operate a hyperscale campus, a regional colocation facility, or a business data center, Real-Time Solutions begin with a clear understanding of the load and the consequences of interruption.

Visit acerts.com to request a power audit, discuss a solution design, or ask for the appropriate technical specification sheet for your facility. You can also submit an enterprise request for quote for a customized power protection project.

Frequently Asked Questions

What is an AI UPS?

An AI UPS is a power protection system designed for the high-density, rapidly changing electrical loads created by artificial intelligence infrastructure. It may include medium-voltage power conversion, battery energy storage, voltage ride-through, load smoothing, and grid-interactive controls in addition to conventional backup functionality.

How does an AI UPS protect a data center?

An AI UPS provides conditioned power during utility disturbances and can isolate sensitive IT equipment from voltage faults. In larger architectures, it can also buffer rapid GPU load changes so the facility presents a more stable demand profile to the grid and on-site generation assets.

How should a data center size UPS capacity for AI workloads?

Start with measured load profiles rather than GPU nameplate ratings alone. Include rack density, cooling, networking, startup behavior, redundancy requirements, battery runtime, power factor, efficiency, bypass capacity, and future expansion. The final design should be validated against the facility’s Tier III or Tier IV objectives and applicable utility interconnection requirements.