Blog · Analysis · Last reviewed June 24, 2026

The Subsea Cable Becomes the AI Border

AI infrastructure does not only live in data centers and chip supply chains. It also crosses oceans through privately built cable routes, landing stations, repair regimes, and national-security chokepoints.

The cable border is not a line on a map. It is a control surface where data movement, cloud access, network repair, ownership, terminal equipment, failover, and lawful interception risk become part of AI governance.

The practical question is not whether a public map can reveal every packet path. It is whether institutions can document which cable dependencies matter, what happens when they fail, and who has authority to change the route.

The Border Under the Sea

The public story of artificial intelligence usually points upward or inward: the cloud, the model, the chip, the prompt, the data center, the user. Submarine cables point downward. They show that the supposedly weightless system has a route across the ocean floor.

That route is not ornamental. The International Telecommunication Union says submarine cables carry about 99 percent of the world's internet traffic and enable cloud computing, financial transactions, government communications, and other critical services. The UN's 2026 reporting on ITU cable resilience described more than 500 commercial submarine cables stretching roughly 1.7 million kilometers. The internet feels wireless because the last hop is often wireless. The world system underneath it is still fiber.

This matters for AI governance because model-mediated reality depends on global data movement. Training data, model updates, enterprise workloads, inference traffic, backups, safety telemetry, content delivery, cloud regions, developer tools, and user prompts all depend on communication networks that can be routed, congested, damaged, licensed, repaired, monitored, or denied.

In this essay, a subsea-cable AI border means the technical and legal boundary formed by cable routes, landing stations, submarine line terminal equipment, terrestrial backhaul, cloud peering, repair rights, capacity leases, ownership, and national-security review. It is not a border in the ordinary passport sense. It is a boundary where compute governance, AI data residency, sovereignty, security, commerce, and communication meet. A model may answer from "the cloud," but the cloud has landing points.

The useful unit is a dependency claim: which AI service, public workflow, research system, or cloud region depends on which route class, landing jurisdiction, control plane, supplier, repair path, and failover option. The cable border is not proven by drawing a line on a public map. It is proven by records that show what the institution can still do when a route, landing station, terminal-equipment operator, or repair permission fails.

This is a dependency claim, not a packet-level map. AI traffic does not always cross the same route, and operators can shift paths through private backbone, internet exchange, content delivery, terrestrial fiber, and backup systems. The governance question is therefore not "which cable carried this prompt?" in isolation. It is who controls the reachable routes, which paths are allowed during ordinary operation and emergency failover, what logs can prove that a promise was kept, and which state or company can impose conditions when the route comes ashore.

Cloud Has a Shoreline

Cloud language hides geography. A customer buys a region, a service level, an API, a managed model, or an enterprise agreement. The provider sees routes, latency, peering, colocation, landing stations, terrestrial backhaul, redundancy, and jurisdictional risk. That is why the cable layer belongs beside AI data centers, AI compute, AI inference providers, and the compute border. A data-residency promise can fail at this layer if routing, support access, backups, monitoring, or failover cross the wrong legal or operational boundary.

A serious residency claim should therefore name more than where the primary database sits. It should define the control plane, identity service, key management, observability stream, support access, backup geography, inter-region replication, emergency failover, and audit trail that make the promise testable. Otherwise "regional AI service" can mean local compute with offshore control, local storage with foreign support access, or local inference whose outage plan silently depends on another jurisdiction.

Google's July 2025 announcement of the Sol transatlantic cable made the AI connection explicit. Sol is planned to connect the United States, Bermuda, the Azores, and Spain, with Palm Coast, Florida anchoring the U.S. landing and Santander anchoring the Spanish side. Google said the system would strengthen capacity and reliability for its global network of Google Cloud regions and help meet demand for Google Cloud and AI services across the United States, Europe, and beyond.

Meta's Project Waterworth announcement was even more direct. Meta described a planned 50,000 kilometer system reaching five major continents, serving the United States, India, Brazil, South Africa, and other regions, and opening new oceanic corridors with high-speed connectivity for AI innovation. The company also tied cable design to resilience: deeper routing, enhanced burial techniques, and high-capacity 24-fiber-pair systems.

These announcements are not only engineering news. They are institutional signals. The companies building model platforms are also building the private routes through which model infrastructure reaches users, customers, and partner economies. The user sees a product. The provider builds a world map.

Hyperscalers Become Cable Builders

Submarine cables were once mainly the domain of telecom consortia. That world still exists, but content providers and hyperscalers have become central actors. TeleGeography's refreshed list of content-provider cable holdings tracks substantial publicly disclosed involvement by Meta, Google, Microsoft, and Amazon, and describes a wider content-provider category that includes platform and cloud firms. It lists many Google systems, including sole-owner routes such as Curie, Dunant, Equiano, Grace Hopper, Nuvem, Sol, Topaz, and Umoja, alongside part-owned systems. It also lists Meta as part owner or sole owner across multiple routes, including Project Waterworth.

The ownership detail matters because cables are not generic pipes once they become strategic infrastructure for model platforms. Ownership, capacity rights, maintenance arrangements, landing partnerships, supplier choices, and route diversity shape who can move data cheaply, reliably, and at low latency. A firm with its own cable portfolio has a different institutional position from a firm that must rent capacity from others.

This is the infrastructure version of platform power. The same company can operate social feeds, AI assistants, cloud services, data centers, developer tools, internal research clusters, and long-distance network routes. The stack is not only software. It is oceanic.

There is a public-benefit version of this story. New cables can increase capacity, reduce latency, diversify routes, improve reliability, and connect underserved regions. There is also a dependency version. If AI-era connectivity is increasingly built around the priorities of a few hyperscale firms, then public institutions may gain bandwidth while losing independent leverage over the maps, contracts, and operational memory that make the bandwidth real.

The Landing Station as Jurisdiction

A submarine cable is most politically visible where it comes ashore.

The landing station is where oceanic infrastructure enters domestic infrastructure. It connects to terrestrial fiber, data centers, carriers, cloud regions, and regulated communications systems. It is also where states can impose license conditions, ownership scrutiny, cybersecurity requirements, physical-security standards, vendor restrictions, and national-security review.

In the United States, Team Telecom is the informal name for the Committee for the Assessment of Foreign Participation in the U.S. Telecommunications Services Sector. The Department of Justice says Executive Order 13913 formalized the body in 2020 and that it reviews certain FCC applications and licenses for national-security and law-enforcement risks. It can advise the FCC that an application raises no risk, should be granted with mitigation conditions, or should be denied, revoked, or terminated. It can also re-review existing FCC licenses when new risks or mitigation failures appear.

The FCC's 2025 submarine-cable process made the AI link explicit from the regulator's side. In July 2025, Chairman Brendan Carr announced rules intended to accelerate submarine cable buildout for AI and next-generation technologies while securing cables against foreign adversaries. In August 2025, the FCC adopted a Report and Order and a Further Notice of Proposed Rulemaking. The adopted order streamlined parts of the licensing process while applying national-security measures such as presumptions of denial for certain foreign-adversary-controlled applicants, limits on capacity leasing, covered-equipment restrictions, and cybersecurity and physical-security requirements. The further notice proposed additional measures for submarine line terminal equipment, reporting, trusted technology, repair ships, and AI use in cable systems.

By June 24, 2026, the proceeding had moved further into the equipment and control-plane layer. On June 4, the FCC released a public draft for tentative consideration at its June 25 open meeting, and the June 18 meeting agenda said the Commission would consider a Second Report and Order and Second Further Notice of Proposed Rulemaking on submarine cables. The public draft would create a regulatory framework for owners and operators of submarine line terminal equipment, add routine conditions and certification requirements, and create a fast-track path for applications meeting national-security standards. The draft itself says it was not official Commission action and remained subject to change. That status matters: source discipline requires treating it as a pending proposal at the June 24 review date, not an adopted rule.

That is the cable border in legal form. The state may not inspect every packet. It may not understand every model running through a cloud region. But it can govern landing, ownership, equipment, leasing, repair, reporting, and security conditions at the point where the network becomes domestic infrastructure.

Current Status: Cable Policy Becomes AI Infrastructure

As of June 24, 2026, cable policy is no longer just a background telecom issue. The ITU-ICPC International Advisory Body on Submarine Cable Resilience, created in 2024, has moved into practical work on deployment and repair times, risk mitigation, redundancy, information sharing, and best practices. ITU described global AI, cloud services, and advanced industrial applications as drivers of digital traffic across major routes. The FCC has likewise tied submarine-cable buildout to AI infrastructure and next-generation technologies, while its June 2026 public draft shows that the regulatory question now reaches SLTE, routine security conditions, trusted technology, and application-review pathways.

The February 2026 Porto Summit made the institutional agenda clearer: non-binding guidance on international cooperation, data-driven risk prevention, repair and maintenance coordination, investment, and clearer authority during incidents. That is a modest instrument, but the modesty matters. Cable resilience usually improves through boring coordination before crisis: known focal points, repair permits, compatible reporting, route information, spare capacity, and operators who have already practiced what happens when a cable fails.

The current status is mixed. The basic internet still depends on largely private cable systems. Governments are increasing national-security scrutiny at landing points, equipment layers, and capacity arrangements. International bodies are trying to improve repair permitting and resilience coordination. Hyperscalers are expanding private or partly private cable portfolios. Public-interest users get better connectivity from that buildout, but they do not automatically get visibility into route planning, capacity allocation, resilience assumptions, supplier dependencies, or the operational logs that explain a disruption.

That makes the cable layer a governance layer for sovereign AI. A country can fund domestic models and data centers, but if its international routes, landing stations, cloud connectivity, or repair capacity remain externally controlled or fragile, sovereignty is conditional.

What Must Be Recorded

The governance object should not be a public cable map alone. Maps are useful, but they can overpromise precision, expose sensitive infrastructure, and still fail to answer the institutional question: which AI systems depend on which routes, contracts, landing jurisdictions, and recovery plans?

A better object is a cable dependency register. For public agencies, critical infrastructure operators, large research institutions, and regulated AI deployments, the register would connect an AI system inventory to the communications layer that keeps the system reachable. It should record primary and backup cloud regions, relevant landing jurisdictions, route-diversity assumptions, known single points of failure, control-plane geography, key-management and support-access paths, major capacity or managed-service dependencies, SLTE and landing-station supplier categories, and the owner responsible for updating the record.

The register should have public and restricted fields. The public should be able to see that a high-impact AI service depends on cross-border connectivity, route diversity, failover promises, and named accountability. Sensitive fields such as exact fiber routes, physical-security plans, exploitable chokepoints, and incident response playbooks can be restricted to regulators, auditors, operators, and cleared procurement officials. Secrecy should protect infrastructure, not erase the existence of dependency.

The register should also preserve negative evidence. If an outage forces emergency failover outside the promised region, if a repair delay exposes a hidden dependency, if a cloud service routes telemetry through a different jurisdiction, or if a supplier change alters the control plane, that event should update the AI register, procurement file, audit trail, and incident record where lawful. The point is not to publish sensitive route diagrams. It is to make resilience and residency claims reviewable after they matter.

This is where the cable border meets AI bills of materials, procurement, incident reporting, and public registers. A cable dependency that is absent from the record will be absent from risk assessment, contract negotiation, continuity planning, and public explanation.

Resilience Is Governance

Submarine cable governance is not only about espionage and adversaries. Most cable failures are not cinematic attacks. Published counts vary by source, scope, and counting method: the OECD's 2025 report on communication-network resilience says submarine cables suffer roughly 150 incidents annually, while UN reporting around the ITU summit described 150 to 200 cable incidents each year and said around 80 percent are caused by human activity such as anchors and fishing trawlers. The policy mistake is to build a security regime only for sabotage while leaving ordinary repair, permitting, route diversity, and single points of failure undergoverned.

That ordinary fragility is still politically serious. The UN reported that 2024 Red Sea cable incidents disrupted an estimated 25 percent of data traffic between Europe and Asia. It also noted that repair can become difficult not because engineers cannot find the fault, but because permits, licenses, overlapping jurisdictions, ship access, and operational coordination slow the response.

Resilience is therefore governance in a very practical sense. A resilient network needs redundancy, route diversity, supplier diversity, repair capacity, landing-station security, crisis coordination, public communication, and clear authority during disruption. The OECD frames redundancy and diversity as basic principles of communication-network resilience and notes that AI and machine learning can help detect problems and dynamically reconfigure networks, while also adding complexity.

For AI systems, resilience records should be tied to incident memory. A major service disruption should not end with a status-page post if it reveals that model access, safety monitoring, emergency communications, research infrastructure, or public-service workflows depended on a fragile cable path. The follow-up belongs in an incident report, not only a network operations ticket.

This creates a recursive infrastructure problem. AI depends on resilient networks. AI is also being added to network management. The system that carries model-mediated knowledge may itself become model-mediated: monitored, predicted, optimized, and rerouted by software systems that few outside the operator can inspect.

The AI Reading

For AI governance, subsea cables change the object of analysis in five ways.

First, they make AI sovereignty material. A country can announce a national AI strategy, but if its cloud access, model access, research collaboration, and data flows depend on foreign-owned routes and landing arrangements, its sovereignty is conditional. Domestic data centers do not solve this if the wider network layer remains externally controlled.

Second, they expose the infrastructure behind model availability. Latency, outage, redundancy, and route congestion can shape which AI services feel usable, which regions receive high-quality access, and which institutions can rely on remote models for critical work.

Third, they complicate export-control thinking. The compute border is not only chips and cloud accounts. Remote access to AI capability depends on the communications layer. A state trying to govern frontier AI through hardware, model weights, cloud use, or data localization eventually meets the cable system.

Fourth, they reveal private cartography. Hyperscalers do not merely sell services into the existing internet. They help decide where new capacity is built, which corridors become redundant, which landing points become valuable, and which regions enter the AI economy as customers, hosts, repair zones, or strategic nodes.

Fifth, they put AI inside network operations. Regulators and operators increasingly discuss AI for network planning, traffic routing, predictive maintenance, physical-threat detection, and cybersecurity detection. Those uses can improve resilience, but they also create new audit questions: whose models operate the network, what data they see, what failure modes they introduce, and whether foreign-adversary-owned or opaque AI systems become part of critical communications infrastructure.

The danger is not that private firms build cables. Private capital and technical expertise are part of how the global internet exists. The danger is that public language keeps calling this "connectivity" long after it has become a governance layer for AI infrastructure.

A Governance Standard

A serious AI-era cable policy should treat submarine infrastructure as part of compute governance, not only telecom policy.

First, landing licenses should disclose control clearly. Public records should identify owners, capacity-rights holders, landing partners, major suppliers, maintenance arrangements, security obligations, and material changes over time, subject to narrow security redactions.

Second, resilience should be tested, not merely promised. Regulators should require route-diversity analysis, repair-time assumptions, landing-station contingency plans, incident exercises, and public reporting on major outages where disclosure is safe.

Third, AI-critical routes should be mapped as critical infrastructure. If a route materially supports cloud regions, public services, frontier-model workloads, emergency systems, financial markets, or government communications, that fact should shape risk assessment.

Fourth, national-security review should not become monopoly protection. Foreign-adversary risks are real, but security rules can also entrench trusted incumbents. Cable policy should protect against hostile control without making public-interest connectivity dependent on a small club of approved hyperscalers.

Fifth, repair governance needs international discipline. Cable damage often crosses legal and physical boundaries. Faster permitting, clearer focal points, pre-arranged repair access, and shared incident protocols matter more than dramatic rhetoric after a cut.

Sixth, public AI strategies should include network access. Public compute, safety research, universities, civil society, and smaller firms need more than GPUs. They need reliable routes, cloud neutrality, affordable capacity, and enough transparency to understand where dependency enters.

Seventh, terminal equipment and supplier dependencies should be governed as part of the cable. The border includes SLTE, landing-station operators, network-management software, maintenance vendors, repair ships, terrestrial backhaul, and cloud peering. A license that names the route but misses the control plane is incomplete.

Eighth, AI use inside cable operations should be auditable. If models assist with routing, incident detection, repair prioritization, traffic engineering, or cybersecurity, operators and regulators should define logging, fallback, vendor-risk, data-retention, and human-override requirements before the system becomes invisible infrastructure.

Ninth, residency and failover promises should leave evidence. Enterprises, public agencies, and safety researchers should be able to inspect contractual and technical records showing where control-plane events, backups, telemetry, key access, support actions, and emergency routing actually went. The promise is not credible if it disappears during outage.

Tenth, public reporting should distinguish outage cause, impact, and dependency. A cable cut, terrestrial backhaul failure, cloud-region outage, DNS incident, DDoS attack, and supplier failure can look identical to users. Public infrastructure policy needs incident records clear enough to show whether the cable layer was really the bottleneck.

Eleventh, cable dependencies should attach to AI system records. A public agency or critical operator should not certify a high-impact AI deployment without naming the network dependencies that support access, failover, monitoring, and recovery. The cable layer belongs in the inventory, not in an undocumented architecture diagram.

Twelfth, procurement should make continuity testable. Buyers should require route-diversity evidence, failover test results, support-access constraints, notification duties for material network changes, and incident cooperation. Without those terms, the buyer has no practical way to know whether a cable risk has become an AI service risk.

Thirteenth, transparency should be tiered. Cable governance should not force operators to publish exploit-ready maps, but it should also reject blanket secrecy. Public records can disclose ownership classes, landing jurisdictions, dependency categories, resilience tests, audit duties, and incident summaries while reserving precise routes and security procedures for trusted oversight.

What This Changes

The subsea cable is the hidden line beneath recursive reality.

A person asks a model for an answer. A company routes work through a cloud region. A government adopts a managed AI service. A student uses a tutor. A hospital sends data to a vendor. A coding agent calls a tool. The interface looks immediate, but the action may cross oceans, licenses, landing stations, company-owned routes, and security regimes before returning as ordinary speech.

This is why the cable belongs beside the data center, the chip, the benchmark, the AI register, and the public compute commons. It is part of the material order that makes synthetic authority possible. It decides where the model can be reached, how fast it answers, how resilient the connection is, who owns the route, and which state can impose conditions at the shore.

The myth of the cloud says intelligence is everywhere. The cable map says intelligence travels through places. Those places can be inspected, licensed, damaged, repaired, monopolized, secured, contested, and governed.

The model-mediated world will not be governed only at the prompt box. It will be governed at the coast.

Source Discipline

Cable governance needs careful sourcing because company announcements, regulator actions, industry maps, intelligence claims, outage reports, and draft rulemakings often blur together. Company posts are useful for route plans, capacity claims, and design intentions, but they are not independent audits of resilience or public benefit. TeleGeography is valuable industry research, but it should be distinguished from primary regulatory records. News and think-tank claims about sabotage or foreign-adversary behavior should not be treated as official findings without regulator, court, or government documentation.

For durable claims, the hierarchy should start with primary or institutional sources: ITU and ICPC resilience materials, FCC orders and Federal Register notices, DOJ Team Telecom materials, official company announcements for the company's own systems, OECD or UN reports for broad resilience context, and public outage disclosures when available. The safest article language separates what a regulator adopted, what a regulator merely proposed, what a public draft says before a vote, what a company announced, and what analysts infer.

Two evidentiary traps deserve special care. First, "99 percent of internet traffic" claims are usually global or international backbone claims, not proof that every AI interaction crosses an ocean. Second, a cable map does not prove data residency or lawful access. Those claims need contracts, route policy, telemetry boundaries, support-access controls, key-management records, and incident evidence. The cable is a governance layer because it creates dependencies and chokepoints, not because it lets outsiders reconstruct every packet path from a public map.

As of June 24, 2026, the FCC's June 4 public draft should be described as pending and subject to change, while FCC 25-49 should be described as adopted and released in August 2025. Review after the June 25, 2026 open meeting should check the final Commission record before any language is hardened.

Sources

Return to Blog