Summary
The electrical design of an AI data center can no longer stop at uptime inside the fence. A new Open Compute Project (OCP) base specification defines how hyperscale computational loads should behave when the utility grid experiences a voltage disturbance. The central idea is straightforward but operationally demanding: a very large data center should not abruptly remove hundreds of megawatts of demand during a short grid fault, then return that demand in an uncontrolled step.
The specification establishes voltage ride-through windows and post-fault active-power recovery targets for computing load at the point of interconnection. In important portions of the envelope, the computational load must remain connected, may reduce consumption in a controlled way, and should recover to 90% of its pre-fault level within two seconds after voltage returns to the normal band. That requirement reaches through utility studies, medium-voltage distribution, UPS behavior, rack power supplies, firmware, controls, and workload operations.
This is not another power-usage-effectiveness exercise. It changes the definition of a well-behaved AI facility from “stays online” to “stays online without destabilizing the system that supplies it.”
Why a Data Center Trip Becomes a Grid Event
Traditional data center resilience is organized around the application. If utility power deviates from acceptable limits, protection systems transfer or disconnect load to protect servers and preserve service. That strategy is rational for a small facility. At hyperscale, the same protective action can become a bulk-power-system disturbance.
OCP cites a simultaneous loss of 1,500 MW of load in North America in August 2024 and a later, larger event in the Eastern Interconnection in July 2026. Removing that much demand while generation remains online pushes frequency and voltage upward. A protection response intended to isolate one customer can therefore create a second problem for grid operators.
NERC’s work on emerging large loads explains the broader issue. Data centers and other power-electronic loads can be unusually large, can change consumption quickly, and are not necessarily visible to system operators at the fidelity needed for planning and real-time balancing. Their behavior differs from older industrial loads, and multiple sites may respond to the same voltage event in nearly the same way. Correlated behavior is what turns a facility setting into a regional reliability concern.
Backup generation does not answer this concern by itself. The grid cares about the amount and timing of demand that disappears from the point of interconnection, regardless of whether the servers continue running behind a UPS or generator.
What the OCP Specification Actually Requires
The OCP document focuses on high- and low-voltage ride-through and post-fault active-power recovery. It applies the performance commitment to the computational portion of load, which the specification says is typically about 70% to 90% of total facility demand. Cooling and auxiliary systems are treated separately because their electrical and mechanical dynamics are different.
The Normal and Moderate-Disturbance Bands
Between 0.90 and 1.10 per unit voltage at the point of interconnection, computing load operates continuously at constant power. For a high-voltage excursion above 1.10 and up to 1.20 per unit, the proposed minimum ride-through time is one second, also with constant-power behavior.
For a sag from 0.80 to below 0.90 per unit, the computational load should ride through for two seconds. It may reduce active power in proportion to voltage, but after voltage returns to the 0.90-to-1.10 band it should restore 90% of pre-fault load within two seconds. Between 0.50 and below 0.80 per unit, the minimum ride-through interval is 0.5 seconds with the same controlled recovery target.
Severe Voltage Sags
The specification allows greater reduction during deeper faults. From 0.35 to below 0.50 per unit, the minimum interval is 0.25 seconds; below 0.35 per unit, it is 0.15 seconds. In those regions, computing load may fall to any value, including zero, but the post-fault recovery commitment still applies once voltage returns to the normal band.
This is an important engineering compromise. Requiring IT equipment to consume constant power at nearly zero grid voltage would demand substantial stored energy and could create damaging current behavior when voltage recovers. OCP explicitly acknowledges that current OEM UPS and power-supply architectures cannot provide unlimited ride-through without redesign and possible new compliance risks.
The Architecture Boundary Moves Into the Rack
Grid interconnection was once primarily a facilities-engineering concern. Ride-through behavior makes it a system property.
A utility sees the aggregate response at the point of interconnection, but that response is produced by layers: substation protection, transformers, switchgear, medium-voltage UPS equipment where present, power distribution units, rack rectifiers, server power supplies, cooling systems, and workload controls. A compliant curve at one layer does not guarantee compliant behavior for the whole site.
PSU and Rectifier Firmware Become Infrastructure Policy
OCP identifies firmware changes to low-voltage rectifiers and power-supply units as one possible mitigation. Stored energy already present in capacitors may allow equipment to bridge brief disturbances or shape its power reduction and recovery. Feasibility, however, varies by model and vendor. A fleet assembled over several hardware generations may contain materially different protection thresholds and restart behavior.
Operators will need an inventory that ties server and rack configurations to electrical response, not just wattage and efficiency. Firmware lifecycle management also gains a new risk dimension: an update can alter aggregate grid behavior even if application benchmarks and ordinary failover tests remain unchanged.
Medium-Voltage UPS Is Not a Universal Retrofit
An in-line medium-voltage, double-conversion UPS can decouple facility load from some utility disturbances and enforce a more predictable response. It may be attractive in a greenfield design. OCP cautions that retrofitting one into an operating campus can be too costly or disruptive because of footprint, interconnection, and construction constraints.
That distinction matters to capital planning. New campuses may be able to design around a ride-through envelope. Existing sites may require grandfathering, incremental compliance for added load, utility-side mitigation, or negotiated operating limits rather than a simple equipment swap.
Cooling Cannot Be Hand-Waved Away
The computational load is the immediate focus because it dominates demand and its power electronics can respond rapidly. Cooling is not irrelevant. Chillers, pumps, fans, and control systems have inertia, restart sequences, and thermal consequences that differ from servers.
If computing returns to 90% of pre-fault power within two seconds while cooling remains offline or restarts slowly, thermal headroom becomes part of the disturbance plan. The correct design question is not only whether server power can recover, but how many events can occur within a given period before thermal buffers, UPS energy, or protection counters force a transfer.
OCP leaves cooling-load compliance and counter-based protection schemes for further work. Infrastructure architects should treat that as an open design dependency rather than assume the first specification completes the problem.
What Operators Should Add to Their Engineering Program
Model the Site as a Dynamic Load
Steady-state peak MW is insufficient. Interconnection studies need voltage-dependent behavior, transfer thresholds, recovery ramps, reactive-power characteristics, and multiple-disturbance behavior. Models should represent actual equipment generations and operating modes. They should be versioned alongside major facility and firmware changes.
Instrument the Point of Interconnection
High-resolution telemetry at the utility boundary is essential for proving what the site did during a fault. Rack and facility telemetry are needed to explain why. Time synchronization across electrical, cooling, and IT systems determines whether engineers can reconstruct a sub-second event rather than correlate unrelated samples.
Test Recovery, Not Only Failover
Conventional tests often verify that protected load survives a source transfer. Ride-through validation must also measure the demand visible to the grid during the event and the ramp after it. Tests should cover partial voltage sags, repeated events, high-voltage excursions, and degraded equipment states—not only a complete utility outage.
Put Electrical Behavior Into Procurement
Requests for servers, PSUs, UPS systems, and controls should include ride-through curves, configurable thresholds, recovery behavior, firmware governance, and model data suitable for utility studies. Without those artifacts, operators cannot know whether nominally interchangeable equipment is electrically interchangeable at fleet scale.
Coordinate Workload Flexibility Carefully
Schedulers may help limit ramp rates, defer batch work, or reduce noncritical demand. But software flexibility must not conflict with facility controls or turn thousands of nodes into a synchronized step load. Any workload response should be bounded, observable, and tested with the electrical control hierarchy.
What the Specification Does Not Settle
Version 1.0.0 is a baseline, not a complete global interconnection code. It does not eliminate regional utility requirements, and it is not itself a NERC Reliability Standard. OCP identifies future work on frequency ride-through, rate of change of frequency, phase jumps, counter-based protection, ramp rates, intra-minute variation, and forced oscillations.
Nor does the document prove that every existing AI campus can comply. Its suggested implementation approach recognizes the need for effective dates, grandfathering, incremental treatment of expansions, and best-effort pathways for existing facilities. That is a practical acknowledgment that firmware, power electronics, and campus topology constrain what can be changed after construction.
The Enterprise Decision
For CTOs and infrastructure leaders, the consequence is organizational as much as technical. Capacity planning, hardware qualification, utility interconnection, facilities controls, and cluster operations must share one model of site behavior. Grid performance cannot remain a document held only by the electrical engineering team while IT independently changes server populations and workload schedules.
The strongest near-term action is to establish an electrical configuration baseline: equipment inventory, firmware versions, protection settings, measured disturbance response, and ownership of every recovery control. New capacity requests should include ride-through and ramp-rate requirements early enough to influence architecture and procurement.
AI infrastructure is becoming large enough that reliability has two directions. The grid must reliably supply the data center, and the data center must behave predictably enough for the grid to remain reliable. OCP’s specification turns that reciprocal obligation into concrete engineering work.