電源断でもシステムは稼働：Metaが瞬間停電への対応準備を検証Lights Out, Systems On: Validating Instant Power Loss Readiness

Meta has introduced a new testing approach it calls Instantaneous PowerLoss Storm, a deliberate effort to verify that its data center infrastructure can keep operating through sudden, zero-notice power disruptions. The work, described on Meta's engineering blog, matters because modern data centers underpin services used by billions of people, and even brief electrical interruptions can cascade into hardware faults, corrupted state, or extended outages if systems are not prepared to absorb them.

The core idea is to treat instantaneous power loss not as a rare edge case to be avoided, but as a condition that should be repeatedly and intentionally rehearsed. Conventional resilience testing often assumes some warning before a power event, allowing systems to drain workloads, flush caches, and shut down gracefully. An instantaneous or "zero-notice" loss removes that buffer entirely. By simulating these abrupt failures across its fleet, Meta aims to expose weaknesses in how servers, storage, and networking equipment behave when power simply vanishes, and to confirm that critical systems either ride through the event or recover cleanly afterward.

This effort sits within the broader discipline known as chaos engineering, the practice of deliberately injecting faults into production or production-like environments to build confidence in a system's resilience. The approach was popularized in part by Netflix's Chaos Monkey, which randomly terminated cloud instances to ensure applications could tolerate the loss of individual components. Meta's PowerLoss Storm appears to extend that philosophy from software-level failures down to the physical power layer, where the consequences are harder to model and the blast radius can be larger. Testing at this level is technically demanding because it involves real electrical infrastructure rather than software toggles, raising the stakes for any test that goes wrong.

The technical context centers on the layered systems that data centers use to guard against power problems. Facilities typically rely on uninterruptible power supplies, which use batteries or capacitors to bridge the gap between a utility failure and backup generators coming online. They also depend on redundant feeds, automatic transfer switches, and increasingly on battery systems integrated directly at the rack or server level. A storm-style test is likely designed to probe what happens when these protections are bypassed or fail to engage in time, validating assumptions about how quickly hardware loses power and whether software has flushed important data to durable storage beforehand.

For a company operating at Meta's scale, the motivation is partly statistical. When an organization runs millions of servers, events that are individually improbable become near-certainties somewhere in the fleet on any given day. A power anomaly that affects only a small fraction of machines can still translate into a meaningful number of failures, so the ability to tolerate instantaneous loss without data corruption or prolonged downtime becomes an operational requirement rather than a theoretical nicety. Rehearsing these scenarios in a controlled way also helps teams refine recovery procedures and reduce the time it takes services to return to normal.

The timing aligns with intensifying industry attention to data center power. The rapid expansion of AI training and inference has driven sharp increases in electrical demand, pushing operators to build denser, higher-power racks that draw more current and generate more heat. Higher power density tends to make systems more sensitive to electrical disturbances, and it strains the grid connections and backup systems that facilities depend on. Other large operators, including hyperscalers and cloud providers, have similarly invested in power redundancy, on-site generation, and energy storage, though their specific testing methodologies are not always disclosed publicly.

Meta indicates it is sharing details of the paradigm, which suggests the company intends the work to inform engineers beyond its own organization, consistent with its history of publishing infrastructure practices. The full post reportedly outlines what the testing covers and how it is conducted, though the available summary is truncated. As with many resilience initiatives, the lasting value will depend on how consistently such tests are run and how the findings are folded back into hardware design, software behavior, and operational playbooks. If the approach proves effective, it could contribute to a wider conversation about how the industry validates power resilience as workloads grow more demanding and the cost of unexpected downtime continues to rise.