When Typhoon Sinlaku Made a 200-km Cable Take a 12,000-km Detour: Anatomy of a BGP Reroute During a Cat-5 Storm

Shortly after dawn UTC on April 14, 2026, our monitoring detected something that looked, at first glance, like a submarine cable failure. Round-trip time between Saipan and Guam — two islands separated by roughly 200 kilometres of Pacific water — jumped from 8.35 ms to 109.98 ms. A 13× latency spike on a cable whose physics floor is just 2.6 ms, during the exact window that Category 5 Super Typhoon Sinlaku was crossing directly over the Northern Mariana Islands. The obvious hypothesis was the obvious hypothesis: the storm had broken the cable.

It hadn't. The cable was fine. What broke was something subtler — and, in a way, more interesting.

The 12,000-km detour

The Mariana-Guam Cable, lit in 1997 and operated by PTI Pacifica, is the only submarine cable connecting the Commonwealth of the Northern Mariana Islands to the rest of the world. Its 268 kilometres of fibre make a short arc from Tanguisson Point on Guam's northwest coast up through Rota, Tinian, and Saipan, carrying every byte of non-satellite internet for about 55,500 people.

Our Saipan probe — RIPE Atlas anchor 6923, hosted by the Pacific transit carrier OneQode (AS140627) — was measuring round-trip time to an endpoint on Guam Exchange (AS152735, prefix 103.142.152.0/24) every thirty minutes. Here is the pre-shift traceroute, captured April 13:

Eight hops. Six of them inside PTI Pacifica's network. End-to-end: 8.35 milliseconds. That is roughly the physics of light travelling twice through 835 km of optical fibre — the expected signature of a direct Saipan → Tanguisson → target journey.

Here is the same measurement 14 hours later, after Sinlaku had passed:

One extra hop. But a single leg — from Guam to Los Angeles and back into OneQode's core — had added 109 milliseconds. The packet was no longer staying on the island. It was crossing the Pacific, entering US transit, handing off to Arelion (AS1299), being carried back across the Pacific, landing in Guam again, and finally reaching PTI for the last leg down the Mariana-Guam Cable to the target. Round-trip distance, in fibre terms: approximately 12,000 kilometres, for what should have been a 200-km trip.

The asymmetry reveal

If the cable had broken, both directions would degrade. We measured the reverse — Tanguisson Point on Guam, back to Saipan — and got the unambiguous answer:

Guam-to-Saipan traversed the Mariana-Guam Cable directly, the same way it had the day before. The cable was functioning exactly as designed. The anomaly existed only in one direction, and only for one ISP's customers. This was not physical damage; it was a routing change.

What the typhoon actually did

Typhoon Sinlaku reached Category 5 intensity on April 12, 2026, with the Joint Typhoon Warning Center recording peak one-minute sustained winds of about 300 km/h and the Japan Meteorological Agency measuring a central pressure of 897 hPa. By several metrics it was the strongest tropical cyclone ever recorded that early in any calendar year. At approximately 14:30 UTC on April 14, its eye passed directly over Tinian and Saipan.

The wind at landfall had weakened to about 230 km/h — still Category 4. Saipan recorded sustained gusts of 209 km/h and 135 mm of rainfall in a single day. Roughly 43,000 people across the Commonwealth of the Northern Mariana Islands lost electrical power. Every piece of telecommunications infrastructure on the surface of those islands — cellular towers, microwave links, cable-landing power plants, peering equipment in ISP racks — experienced the same battering that the buildings housing them did.

Here is what our measurements show the typhoon did not do. These are the other five major Pacific submarine cables that our infrastructure monitors via probes with line-of-sight to the affected region:

Every other Pacific cable in the affected region carried on within its normal variance. Sinlaku, a storm powerful enough to take down a significant portion of the built environment on two islands, produced no measurable degradation in five of the six cables our probes could see. The structural engineering of modern submarine cable landings — buried beach manholes, hardened beachfront buildings, deep-sea routes well below storm influence — worked as designed.

The synchrony

What failed was above ground, and the timing matters. Three RIPE Atlas observations anchor the chronology:

PTI Pacifica's own probe on the island dropped off the RIPE Atlas mesh 13 hours before OneQode's path shifted — the first signature of trouble in PTI's Saipan-side infrastructure. By the time OneQode's forwarding changed, PTI had been partially unreachable from the Atlas network for half a day. The eye of the storm arrived ten hours after the reroute, meaning the BGP damage preceded the highest winds. What drove failures on April 13 was the storm's outer bands, the precursor weather, and presumably the loss of commercial mains power to any equipment not on hardened UPS.

What the public BGP table shows

We pulled the RIPE Routing Information Service archive for the two prefixes involved. The RIS stores BGP updates seen by a global mesh of collectors, and its record for this window is instructive:

Neither prefix lost its origin during the window. Both kept being announced to the global internet as normal. What the RIS captured is much smaller — short bursts of intermediate-path churn and a handful of brief withdrawals from a few peers, consistent with isolated peering sessions flapping. A casual reader of the public BGP table for April 13–15 would have seen nothing alarming. Both networks, viewed from the outside, stayed reachable.

That is precisely the point. The reroute was not a global BGP event. It was a local one — happening inside the border between two ISPs on a single island, and invisible to the public routing table. OneQode had lost its direct peering session with PTI Pacifica in Saipan. With the local route gone, OneQode's routers fell back to their next-best path to reach Guam's address space: the default route through their transit carrier, Arelion, in Los Angeles. Packets still got where they were going. They just crossed the Pacific twice to do it.

Topology is not geography

This is the part most engineers already know, but that every storm re-teaches: internet topology is an economic graph laid on top of a physical one, and the two do not have to match. A packet going from Saipan to Guam does not travel "the shortest way" in a metric-distance sense; it travels the way the BGP policies of the networks in its path prefer at that moment. Those policies reflect peering contracts, transit bills, route preferences, and hundreds of small engineering choices made years ago in quiet rooms somewhere else. Most of the time, those choices happen to align with geography, because the cheapest peering is usually the closest one. When the cheapest peering goes offline, the second cheapest takes over, and sometimes the second cheapest lives 11,000 kilometres away.

The Pacific makes this especially vivid. An island's internet is only as robust as the number of distinct local peering relationships its ISPs maintain. If there is one peering between the two main operators, and it runs through a single piece of equipment in a single building, then any single-point failure at that layer — a roof off, a generator out, a link down — will force a trans-oceanic trombone. The submarine cable was never the weak point. The cable is, in fact, the part of the infrastructure most engineered for exactly the kind of storm Sinlaku brought.

The 33-hour stuck alert

Alert 2026041403 in our system fired at 05:01:55 UTC on April 14 and stayed in active/critical status for 33 hours. This was a real ship-in-the-bottle moment: the event we were built to catch, caught, and then the alert stuck. We debugged the lifecycle live and shipped two fixes against our anomaly detector while the storm was still blowing.

The first fix addressed a subtle gating condition: our auto-resolve logic required a successful ping round-trip to move an alert out of active. For ICMP-blocked targets — where traceroute works but ping does not — that condition was unsatisfiable, and alerts could never close. We replaced it with COALESCE(ping_rtt_ms, trace_rtt_ms), so trace-only evidence now counts. An unrelated stuck alert (Batam–Sarawak) auto-resolved on the next run.

The second fix was specific to Sinlaku's signature. Our alert captures the baseline RTT at detection time and compares current measurements against that frozen number. The rolling baseline inside the health-check table adapts to persistent changes within about fifteen hours, so fifteen hours after the reroute the rolling baseline had shifted from 25 ms to 100 ms. But the alert's frozen baseline (25.57 ms) hadn't moved. Current versus frozen kept reporting a 330% spike — with no obvious way to auto-close — even though in the new regime everything was perfectly stable. We added a baseline-shift check: when an alert is older than 24 hours, the current measurement is within 20% of the rolling baseline, and the rolling baseline is itself at least 1.5× the frozen one, we de-escalate from critical to monitoring and log the drift ratio. Alert 74 de-escalated cleanly at drift 3.95.

The underlying lesson is that anomaly detection has to distinguish two different failure modes: a cable breaks and latency spikes forever until physical repair; or a topology shift establishes a new normal and the monitoring needs to recognise it. We had been writing code that assumed only the first mode. Sinlaku gave us the second.

What the islands teach

In the days after the storm, on every half-hour our infrastructure has data for, the new 110 ms normal for Saipan → Guam has held steady. PTI Pacifica's probe stays offline. OneQode continues to route via Los Angeles. Whether this restores to a local peering next week, next month, or never depends on how quickly the ground-level infrastructure of one small Pacific island can be rebuilt — and on whether anyone involved decides the single-peering topology was itself the problem worth fixing.

The cable is still there. Its fastest round trip is still 8.35 ms. For the direction Guam sends to Saipan, that is still what our measurements see. For the direction Saipan sends to Guam, through one specific ISP, the number is 110 ms for now. Two islands, 200 kilometres apart, one submarine cable running between them — and for a few million packets a second, in one direction, a 12,000-km detour that none of them signed up for.

Our probes are still watching.

Data for this article comes from our own RIPE Atlas measurements (anchor 6923 and probe 65653 in Saipan), cable_health_checks observations from our probes in Minsk, Tbilisi, Jerusalem, Sevastopol and Almaty, and the public RIPE Routing Information Service archive for BGP updates on prefixes 103.142.152.0/24 and 103.151.64.0/24. Storm best-track figures are from JTWC and JMA advisories. Explore the Mariana-Guam Cable dossier or use the route calculator to see live latency between any two cities.