How Long Does Subsea Cable Repair Take in 2026? Data Guide

A subsea cable can detect a fault within minutes. Restoring normal service is another matter.

In 2026, operators should plan for a difficult transpacific repair as a 30–90-day recovery event. Three to six months is possible when vessel availability, weather, permits, geopolitics, or a second fault complicates the job. That’s the practical benchmark—not the optimistic estimate in a network brochure.

Subsea cable repair time: what the clock actually measures

A cable failure is a sequence of separate events:

1. Fault detection and confirmation
2. Vessel and permit mobilization
3. Transit to the fault area
4. Grappling and cable recovery
5. Cutting, jointing, testing, and redeployment
6. Traffic restoration and capacity normalization

Optical monitoring systems can usually identify an abnormality within minutes or hours. Operators compare measurements from both ends of the system, often using optical time-domain reflectometry, to estimate the fault location. They must also establish whether the problem is a physical break, repeater failure, landing-station fault, or terrestrial backhaul issue.

That part is relatively quick. The vessel is usually the constraint.

A repair ship must be available, equipped for the cable type, legally permitted to work in the relevant waters, and able to reach the site. A vessel already assigned to another repair may be unavailable for weeks. It may also need specialized grapnels, replacement cable, jointing equipment, and spare repeaters.

These rows are not automatically additive. Weather or regulatory restrictions can delay vessel mobilization, transit, grappling, or the repair itself. The totals are scenario ranges, not a promise that every delay stacks neatly on top of the others.

Many operators restore customer service through rerouting long before the cable is physically repaired. That can keep traffic moving, but usually with higher latency, less spare capacity, and greater exposure to a second failure. Full capacity normalization may therefore come weeks after initial service restoration.

A transpacific route shows why the timetable expands quickly. A vessel may have to leave a distant repair base, spend several days reaching the fault, and then wait for a safe weather window in the North Pacific. If another cable fails during the same period, the operator may no longer have a repair ship or enough alternate capacity available.

Capacity is not the same as resilience

Meta’s proposed Project Waterworth cable has put a striking number into the market: approximately 50,000 kilometres of route length and a designed capacity of up to 220 Tbps, according to Meta’s February 2025 announcement. Other proposed systems, including the Petal concept cited in industry coverage, have targeted capacities approaching 1 Pbit/s.

Those figures matter, but headline capacity is not the same as immediately usable, globally accessible bandwidth. Actual throughput depends on terminal equipment, modulation, error margins, spectrum allocation, fiber-pair use, landing-station electronics, and the terrestrial networks connected at each end. A hyperscaler may also reserve much of a system for internal traffic rather than sell it as public capacity.

The bigger risk is concentration at a few physical points:

• Several cables may use one landing station.
• Inland fiber may converge on one rail corridor, bridge, or exchange.
• A landing site may depend on one substation.
• Routes may cross the same fishing grounds or anchorage approaches.
• A regional earthquake or storm may damage multiple systems at once.

Three 300-Tbps systems with separate landing stations and inland paths can be more resilient than one 1-Pbps system. The smaller systems provide less headline capacity but more failure options.

Operational test: If this cable, both landing stations, and the preferred inland route were unavailable for 90 days, how much traffic could still move?

For delay-tolerant AI training, traffic can be queued or shifted to another region. Financial services, emergency communications, industrial control, and real-time inference have less flexibility. They need physically diverse paths, automatic failover, and, where practical, local processing.

A 30–50% spare-path-capacity range can be a useful planning starting point for critical networks, but it is not a universal engineering standard. The right figure depends on traffic elasticity, route diversity, service-level targets, failover behavior, and the restoration contracts behind the network. Thirty percent may work when workloads can be slowed. Low-latency services may need fifty percent, an active second route, or both.

The same exposure appears in fabs and power projects

Semiconductor manufacturing makes the geographic problem easier to see. A fab depends on chemicals, ultrapure water, electricity, equipment maintenance, advanced packaging, logistics, and skilled labor. Comparing TSMC’s Arizona operations with its Kumamoto/JASM facilities by construction cost alone misses the point.

TSMC’s U.S. expansion has been associated with advanced N4-class production, while JASM’s first Kumamoto fab was oriented toward 22/28-nanometre specialty processes. TSMC described those roles in its 2024 annual reporting and related facility announcements. They serve different products and have different yield and qualification curves.

The practical metric is good wafers per month after 12–24 months of production—not announced investment.

Cost formula: Cost per good wafer = (depreciation + labor + utilities + materials) ÷ (gross wafers × yield)

A site with higher operating costs can still be competitive if its yield on a complex product is materially better. Conversely, a two-year yield delay can erase labor savings and subsidies.

Energy projects have a similar gap between announced capacity and usable capacity. The proposed North Plains Connector, described by its developers as roughly 422 miles of 525-kV HVDC transmission with 3,000 MW of bidirectional capacity, could improve power sharing across the Upper Midwest. But a planned corridor is not firm supply for a data center. The buyer needs an interconnection agreement, construction schedule, delivery rights, and defined operating obligations.

One proposed AI-campus plan has been described as combining a 1.8-GW load with 2 GW of gas generation and up to 2.6 GW of battery storage. If that battery has four hours of duration, it stores about 10.4 GWh. A continuous 1-GW load consumes roughly 8.76 TWh per year, so the battery offers hours of bridging—not long-term energy security. The campus would still need fuel logistics, transmission redundancy, cooling, reserves, and black-start capability.

“Renewably powered” does not automatically mean firm power at every hour.

Build the recovery plan before the failure

For a subsea network, model the loss of one major route for 90 days. Then model a second event: the preferred landing station loses grid power or inland backhaul. Measure latency, packet loss, congestion, cloud-region reachability, and the traffic that must be throttled.

For a fab, simulate a six-month interruption in one chemical, equipment, or packaging supply. A second qualified source may cost more during normal operations, but it can preserve production when the primary source fails.

For an energy-intensive campus, distinguish “connected” from “energized.” Ask who pays for temporary generation, how long fuel can be delivered, whether batteries can bridge a transmission outage, and whether the grid can absorb a sudden restart.

LNG provides a useful cost comparison. An extra seven to twelve sailing days for a 170,000-cubic-metre carrier can add fuel, vessel hire, boil-off, tolls, congestion, and financing costs. Under stressed conditions, an illustrative diversion can cost roughly $1.9–$6.8 million per cargo, according to project-style shipping assumptions. It looks wasteful until the primary route closes.

Subsea networks face the same trade-off. Paying for a second route, independent landing infrastructure, and spare capacity is inefficient while every system is healthy. During a 90-day outage, it can be cheaper than lost transactions, delayed model training, and emergency bandwidth bought at the worst possible time.

The decision is not whether redundancy costs money. It is whether the business can afford to operate without it while a repair vessel is still crossing the ocean.