← All incidents
Incident dossier · Rank #2

SK C&C Pangyo Data Center Fire and the National Kakao/Naver Outage

SK C&C (tenants: Kakao, Naver) 2022-10-15 127h 30m core impact FirePower

On 15 October 2022, a lithium-ion UPS battery on the third basement level (B3) of the SK C&C Pangyo Data Center ignited and spread, forcing firefighters and operators to cut power to the entire facility. Because tenant Kakao had concentrated core functions — including its sole authentication service — at this single site, the power cut cascaded into a nationwide outage of KakaoTalk, Daum, Kakao-linked banking, KakaoT and dozens of downstream services. Co-tenant Naver, which kept its own separate data center, was far less affected. KakaoTalk stayed down for more than 11 hours, the longest in its history, and full restoration of all Kakao services took roughly five days. The event triggered a CEO resignation, a three-way blame dispute, and South Korea's "Kakao Outage Prevention Act" mandating geographic redundancy for major platforms.

Failure cascade

Failure cascade: trigger → fault → downstream impactTriggerPrimary faultDownstream impactTrigger — Fire (2022-10-15)Trigger · Fire2022-10-152022-10-15Primary fault at SK C&C (tenants: Kakao, Naver) — SK C&C Pangyo Data CenterSK C&CSK C&C Pangyo Data CenterSK C&C Pangyo Data CenterDownstream service degraded by the fault: KakaoTalk (messaging)KakaoTalkDownstream service degraded by the fault: KakaoPageKakaoPageDownstream service degraded by the fault: Daum portal / Daum NewsDaum portal / Daum NewsDownstream service degraded by the fault: Brunch (blogging)Brunch+4 more downstream services

Trigger → primary fault → downstream blast radius, derived from the sourced root cause and affected-services record.

Facility & location

Operator
SK C&C (tenants: Kakao, Naver)
Data center
SK C&C Pangyo Data Center
Location
Seongnam (Pangyo), South Korea
Date
2022-10-15

Impact & scale

Users affected
Nationwide — the majority of South Korea's population relies on KakaoTalk; ~2M users reported to have left the app in the immediate aftermath (single-source, low confidence)
Financial
Over KRW 2 trillion in Kakao-affiliate market-cap loss on Oct 17 (single-source, low confidence); Kakao committed KRW 460B (~US$323M) to new Ansan capacity
Scope
National-infrastructure-grade outage
Services / systems down
  • KakaoTalk (messaging)
  • KakaoPage
  • Daum portal / Daum News
  • Brunch (blogging)
  • Kakao Map
  • Kakao Bank (partial — Kakao-auth-dependent functions)
  • KakaoT (ride-hailing)
  • Kakao Pay securities/real-estate functions

Impact data & metrics

Ignition-to-suppression interval~1 minute (spark 15:19 -> gas discharge 15:20)
BMS-reported battery temperature during active fire~20C (held steady; sensing failed)
Clean-agent gas discharged~1,500 kg (50 kg x 30 cylinders)
Battery packs consumed~60 packs
Fire-charred area vs room size~40 sqm charred of a 3,300 sqm battery room
Spark-to-firefighter-arrival18 minutes (15:19 -> 15:37)
Kakao servers de-energised by whole-building cut~32,000 servers (3만2000대)
Peak firefighting mobilization114 personnel, 46 apparatus
Occupants evacuated / casualties26 self-evacuated / 0 casualties
Total time to full extinguishment8 hours 13 minutes (15:33 -> 23:46)
Kakao service full-recovery time127 hours 30 minutes (>5 days)
Ignition cell service age~6 years (installed 2016)

Magnitude profile

Magnitude sub-scores (0–10)Magnitude sub-scores (0–10)Users 10Users affected (0–10) — breadth of the user/customer population impacted. — scored 10/10.Financial 9Financial impact (0–10) — direct + consequential cost. — scored 9/10.Duration 9Outage duration (0–10) — how long service was degraded/down. — scored 9/10.Blast 10Blast radius (0–10) — how wide the fault propagated across systems/regions. — scored 10/10.
Magnitude 9.6 = blast 10×0.35 + users 10×0.25 + financial 9×0.20 + duration 9×0.20 (sub-scores 0–10 · weighted composite)

Blast radius and user impact are at ceiling: KakaoTalk is de-facto national communications infrastructure, and the outage propagated into banking, ride-hailing, deliveries and grocery. Duration scores high on the 5-day-plus full-restoration tail (7,650 min ≈ 127.5 h) even though the headline KakaoTalk downtime was ~11 hours. Financial score reflects the reported >KRW 2 trillion affiliate market-cap loss (low-confidence, single-source) plus the KRW 460B remediation commitment.

Sequence of events (SOE)

Phased sequence of events2016 · TRIGGER — Lithium-ion UPS batteries installed in the B3F electrical/battery room; per MSIT no Li-ion-specific fire measures were ever added over the ~6 years that followed (latent condition)TRIGGER20162022-10-15 15:19 · TRIGGER — CCTV captures a spark from a single Li-ion cell on a UPS battery rack in the B3F electrical room; internal short from age-related insulation breakdown initiates thermal runawayTRIGGER2022-10-15 15:192022-10-15 ~15:19 · DETECTION — BMS fails as a detection layer — captures no abnormal warning and holds battery temperature at a steady ~20C even as the cell burnsDETECTION2022-10-15 ~15:12022-10-15 15:20 · DETECTION — Separate automatic fire-detection loop triggers ~1 minute after the spark; automatic suppression commanded to dischargeDETECTION2022-10-15 15:202022-10-15 15:20 · MITIGATION — Clean-agent/halogen GAS suppression discharges ~1,500 kg (50 kg x 30 cylinders); system activates correctly but does not extinguish the fireMITIGATION2022-10-15 15:202022-10-15 ~15:20 · MITIGATION — Gas suppression proven ineffective — sealed Li-ion cells in an enclosed basement room prevent oxygen-displacement agents from penetrating; only water (cooling) can arrest the runawayMITIGATION2022-10-15 ~15:22022-10-15 15:33 · DETECTION — Official fire-department report time logged for the B3F electrical-room fire at SK Pangyo Campus Building ADETECTION2022-10-15 15:332022-10-15 15:37 · MITIGATION — Firefighters arrive on scene, 18 minutes after the 15:19 sparkMITIGATION2022-10-15 15:372022-10-15 ~15:30s · IMPACT — All occupants self-evacuate — 26 workers exit under their own power with zero casualtiesIMPACT2022-10-15 ~15:32022-10-15 16:52 · MITIGATION — Because water needs the power off (electrocution risk), fire services formally request SK C&C cut electricity to the B3F zoneMITIGATION2022-10-15 16:522022-10-15 ~16:52 · CASCADE — Selective isolation is impossible — the B3F UPS also feeds the whole center — so a WHOLE-BUILDING power cut is executed, de-energising all ~32,000 Kakao serversCASCADE2022-10-15 ~16:52022-10-15 evening · IMPACT — Containment holds: the fire stays within the B3F electrical/battery room and never reaches the server halls; the national outage is caused by loss of POWER, not fire damage to serversIMPACT2022-10-15 eveni2022-10-15 ~17:19 · MITIGATION — Main flames knocked down roughly two hours after ignition once water is applied to the ~40 sqm affected area (of a 3,300 sqm room); ~60 battery packs consumedMITIGATION2022-10-15 ~17:12022-10-15 23:46 · RECOVERY — Full extinguishment declared — an 8h13m operation from the 15:33 report; peak mobilization 114 personnel and 46 apparatusRECOVERY2022-10-15 23:462022-10-20 ~19:00 · RESTORED — Kakao services fully normalized after 127 hours 30 minutes of degraded/failed operation — recovery driven by restarting a single-site fleet with no geographic redundancyRESTORED2022-10-20 ~19:02022-12-06 · RECOVERY — MSIT announces investigation results: BMS detection failure (held ~20C), cabling routed over batteries, no Li-ion-specific measures since 2016, UPS not fully separated; fault assigned to both SK C&C and KakaoRECOVERY2022-12-062023-03-30 · RESTORED — National Forensic Service announces mechanism: internal short from age-related insulation breakdown as probable ignition cause, external factors unlikely, precise trigger still under studyRESTORED2023-03-302022-10-15 ~15:20 · IMPACT — National-scale service disruption begins as Kakao and Naver platforms fail following the power loss at the single Pangyo siteIMPACT2022-10-15 ~15:2

Root cause

TRIGGER. On 15 October 2022 at approximately 15:30 KST (오후 3시30분경), a fire broke out in the basement electrical room of SK C&C's Pangyo data center in Seongnam, South Korea. The Korean-language record states the ignition point as "지하 3층 무정전전원장치에서 발화" (ignition in the UPS at basement level 3, B3F), and ZDNet Korea's same-day report independently places the origin in the underground electrical room ("지하에 위치한 전기실에 발생한 것으로 알려졌다"). The seat of the fire was thus the uninterruptible power supply (UPS) / battery installation on the B3F electrical floor — the very equipment meant to provide backup ride-through power. MECHANISM — WHY THE OUTAGE HAPPENED WITHOUT SERVERS BURNING. Critically, the fire did not physically destroy the tenant server halls. ZDNet Korea confirms the fire did not spread to the server/computer rooms ("화재가 서버실과 전산실까지 확산되지 않아"). The outage was instead caused by a deliberate fire-safety measure: to prevent the fire from spreading and to allow firefighters to work safely around energized electrical equipment, power to the entire building was cut ("건물 전체 전원 공급을 차단"). This building-wide de-energization — not flame or smoke damage to compute hardware — is what took the servers offline. On-site UPS/battery redundancy could not bridge the load because the UPS installation itself was the seat of the fire and the whole building was intentionally de-energized. (The frequently cited "~15:29 power cut to Kakao's zone" is not supported by the retrieved sources and is omitted here.) WHY REDUNDANCY DID NOT ABSORB THE LOSS. The loss of a single facility should, in a properly distributed architecture, have been survivable via geographically dispersed failover. The Korean record identifies the core organizational root cause as the absence of an effective distributed data-center architecture: the facility operated as a single point of failure and Kakao's servers were insufficiently dispersed. The record is pointed that nothing had changed since a materially similar single-data-center-dependency outage a decade earlier (April 2012): "10년 전 사건이 일어났을 때 이후로 서버 분산 등의 달라진 점이 없다." Because a critical share of Kakao's core services (and other tenants including Naver) was concentrated in this one center without working cross-site failover, the single-facility power loss cascaded into a nationwide service outage. ESCALATION AND SLOW RECOVERY. The fire was suppressed the same night at 23:46 KST, 8 hours 13 minutes after ignition ("8시간 13분 후인 그날 23시 46분에 진화되었다"). Restoring power nonetheless proved slow because the fire-damaged electrical room could not be safely re-energized immediately. Recovery therefore stretched far beyond suppression: mobile KakaoTalk was partially restored overnight, but full service restoration was not completed until 20 October 2022 ("10월 20일 23시, 모든 서비스와 기능들이 복구 완료됐다") — roughly 127 hours after the fire, one of the longest outages in the company's history. MSIT (과학기술정보통신부) coordinated the government response and Kakao co-CEO Namkoong Whon resigned over the outage (confirmed by the English Wikipedia Kakao article: "co-CEO Namkoong Whon resigned after a service outage caused by a data center fire"). SOURCING CAVEAT (unchanged and important). The B3F-UPS ignition finding and the "power-cut, not fire-damage" mechanism are drawn from ZDNet Korea (same-day) and Korean/English Wikipedia aggregations — press/tertiary sources, not a primary forensic document. The verbatim MSIT official cause-investigation report (reported published 2022-12-06) and any NFS/police final forensic conclusion (in particular whether the precise battery-ignition mechanism was conclusively determined or left undetermined) were NOT independently retrieved. The specific attribution of the forensic origin investigation to the NFS (국립과학수사연구원) and the Gyeonggi Nambu Provincial Police is not supported by the retrieved sources and should be treated as unverified. The exact ignition mechanism inside the UPS/battery installation therefore remains reported-but-not-primary-verified. Official confirmation. Official source genuinely established this pass. The MSIT investigation press release ('디지털서비스 장애 원인 조사결과 및 시정조치 요구', 6 Dec 2022) was read via korea.kr, the official Korean-government policy-briefing portal (KOGL type-1) that republishes MSIT releases verbatim — a legitimate primary official/regulatory document, hence officialPostmortem=true. MSIT's own msit.go.kr served a maintenance/bot-block page for every request, so the ministry-hosted copy could not be read directly. The fire-origin conclusion comes from the National Forensic Service (NFS) 감정 결과 and the police+NFS+fire-authority joint on-site exam; these are official-body findings but were only accessible as verbatim Yonhap wire quotations (labeled 'press' here for honesty), since NFS/police did not publish standalone public reports. The exact reason for the battery cell's internal insulation breakdown remained officially undetermined and no criminal charges were filed as of the reporting. There was no prior detailed root cause supplied; the finalRootCause was constructed from the official findings.

Contributing factors

Correction of errors (COE)

Lessons learnt

Improvements & remediation

Comprehensive analysis

A small fire, a national outage

The Pangyo incident is a textbook case of impact decoupled from physical damage. Only ~40 sqm of a 3,300 sqm battery room charred, ~60 packs burned, and flames never reached the server halls — the ~32,000 Kakao servers and Naver gear were intact throughout. The outage came entirely from the whole-building power cut needed to fight the fire safely. The lesson is that a data center's blast radius is defined by its electrical topology and its dependency graph, not by how much of it burns.

Why gas suppression could not work

The clean-agent/halogen system discharged correctly within ~1 minute of the spark, yet the fire burned on. This was physics, not malfunction: gaseous agents extinguish by displacing oxygen, but a Li-ion cell in thermal runaway generates its own heat and oxygen and is sealed in casing, so the agent has no gap to penetrate. In an enclosed basement room, only water — which cools as well as smothers — can arrest the runaway. Reporting underscored the mismatch: ~1,500 kg of gas failed to stop a fire over roughly 40 sqm.

The detection blind spot

The BMS, nominally the early-warning layer, reported a steady ~20C throughout an active fire and flagged no anomaly before ignition. Whether the sensing scheme simply could not see a cell-internal short or the loop had failed silently, the operational result was false assurance. Detection that is not validated against the exact failure mode it must catch — internal short and thermal runaway in an aged cell — is a control that exists on paper only.

Topology as the amplifier

The single most consequential design choice was coupling: the UPS feeding the B3F fire zone also fed the whole center, cabling ran directly above the batteries and burned through at onset, and some UPS units were not fully separated from the battery room. When firefighters needed the power off at 16:52 to apply water, staff could not identify which switches would isolate only the fire zone, so a whole-building shutdown was the only option. Good design would have made selective isolation a labeled, drilled, minutes-long action.

Single-site dependence and the regulatory gap

Kakao's near-total reliance on one site, with no geographic redundancy and no drills for a full power-loss scenario, turned a facility fire into a 127h30m recovery. MSIT assigned fault to both operator and tenant. Compounding it, enhanced data-center suppression requirements effective Feb 2022 were not applied retroactively to the 2016 building, leaving the highest-risk room under the weakest standard — a reminder that grandfathering legacy critical infrastructure can institutionalize the very gap that fails.

Technical deep-dive

The Pangyo event is best understood as a small fire that caused a national outage — the damage vector was electrical de-energisation, not combustion. The fire itself was tiny in physical terms: only about 40㎡ of the 3,300㎡ B3F battery room was charred, roughly 60 battery packs were consumed, and the fire never reached the server halls. The ~32,000 Kakao servers and Naver equipment were physically intact throughout. What took them down was the whole-building power cut, executed because the fire could not be suppressed any other way. The forensic chain runs like this. At 15:19 an aged Li-ion cell (internal short from insulation breakdown per NFS) sparked. The BMS — the layer meant to see exactly this — did not register it, holding a steady ~20°C reading through an active fire, meaning either the sensing scheme could not see cell-internal runaway or the monitoring loop had failed silently. Detection instead came from a separate automatic fire-detection loop tied to the suppression system, which triggered at 15:20, about one minute later, and discharged the gas — approximately 1,500 kg (50 kg × 30 cylinders). The gas system activated correctly and on time but did not extinguish the fire, and this failure was physics, not malfunction. Gaseous agents asphyxiate by displacing oxygen; a Li-ion cell in thermal runaway generates heat and oxygen internally and is sealed in casing, so the agent has no gap to penetrate. Compounded by the enclosed B3F electrical room, only water — which adds a cooling effect the gas lacks — could arrest the runaway. Reporting noted that despite 1,500 kg of gas discharged, knocking down a fire over roughly 40㎡ still took about two hours. But water and live UPS busbars are incompatible: electrocution risk. At 16:52 firefighters formally requested SK C&C cut power. Here the topology became decisive. The UPS feeding the burning B3F zone also fed the whole center, cabling ran above the batteries and had already burned through, and some UPS units were not fully partitioned from the battery room. Staff could not identify in time which zones/switches needed selective isolation, so no selective cut was possible. A whole-building shutdown was executed, instantly de-energising the Kakao/Naver fleet — "카카오 서버 전체인 3만2000대의 전원 공급이 끊겨." Firefighters arrived at 15:37 (18 minutes after the spark); main flames were knocked down about two hours after ignition (~17:19), and full extinguishment was declared at 23:46 — an 8h13m operation measured from the 15:33 report — with peak mobilization of 114 personnel and 46 apparatus. Kakao's services did not fully normalize until 127h30m later, a recovery time driven not by the fire but by the scale of restarting and re-syncing 32,000 servers with no geographic redundancy.

References & provenance

Sourced from public post-incident reports. Quotes are short attributed excerpts for provenance only; the analysis above is original and substantially shorter than its sources. Last verified 2026-07-31.

Root access required

The DC Incidents dossier is a root-only module. Sign in with an authorized account to continue.

Back to Home