Blog · Book Review · Last reviewed June 25, 2026

The Undersea Network and the Ocean Floor of the Internet

Nicole Starosielski's The Undersea Network is a book about the part of the internet that refuses to become metaphor. The cloud is not in the air. Wireless life is carried by fiber, beaches, cable stations, ships, permits, repair regimes, island histories, military routes, finance, and ocean ecologies. Its AI-era lesson is direct: every model call travels through infrastructure that has owners, chokepoints, jurisdictions, maintenance labor, environmental conditions, and political histories.

The Book

The Undersea Network was published by Duke University Press in March 2015 in the Sign, Storage, Transmission series. Duke lists the book at 312 pages, with 55 illustrations, and gives the paper ISBN as 978-0-8223-5755-1, the hardcover ISBN as 978-0-8223-5740-7, and the electronic ISBN as 978-0-8223-7622-4. Google Books gives the same publisher, 2015 date, and 312-page length.

The book follows submarine cable systems from deep ocean routes to landing zones and cable stations, especially across the South Pacific. Starosielski's method is not only technical history. It is media archaeology, infrastructure studies, fieldwork, and environmental attention joined together. The book asks readers to stop imagining communication as pure flow and start looking at the places that make flow possible.

Duke also points to Surfacing, the interactive companion project by Nicole Starosielski, Erik Loyer, and Shane Brennan, with additional writing by Jessica Feldman and Anne Pasek. That companion matters because the book is not only about hidden cables. It is about a method for making hidden infrastructure narratable without flattening it into a single global map.

Current Context

As of June 25, 2026, the book's infrastructure method has become a live governance map. ITU's submarine-cable resilience backgrounder, last updated in April 2026, identifies accidental human activity such as fishing and anchoring, natural hazards, aging infrastructure, and complex regulation as key resilience problems; it also reports more than 170 cable repairs worldwide in 2025. ITU's 2026 Porto Summit release described about 500 submarine cables extending more than 1.7 million kilometers, and said more than 99 percent of international data traffic travels by subsea cables.

That does not mean every model call crosses an ocean or that one cable failure threatens the whole internet. It means the ordinary language of the cloud hides a global repair, permitting, routing, landing, and redundancy system. The governance question is not whether cables are important in the abstract. It is which services depend on which route classes, landing jurisdictions, repair capacities, suppliers, and failover promises.

The policy record now names those dependencies directly. The FCC's August 7, 2025 submarine-cable action tied secure cable buildout to AI and next-generation technologies while adopting measures around licensing, foreign-adversary control, covered equipment, cybersecurity, physical security, and capacity leasing. In Europe, the Commission's cable-security materials frame submarine communication and electricity cables as critical infrastructure, with prevention, detection, response and repair, and deterrence as organizing priorities. The EU's June 24, 2025 cable-security update also describes EU-wide mapping and coordinated risk assessment for submarine cable infrastructure.

The current context sharpens Starosielski's point: cable systems are not passive pipes under a neutral internet. They are governed corridors. They can be licensed, delayed, secured, damaged, repaired, rerouted, monopolized, overclaimed, or made invisible. That is exactly why AI infrastructure analysis has to move from the model interface down to the shore.

Against Flow

The book's first major intervention is conceptual. The internet is usually described through movement words: flow, stream, traffic, cloud, feed, link, route. Those words can be useful, but they can also make infrastructure feel weightless. Starosielski pushes in the other direction. A cable route is not a line on an abstract graph. It is a negotiated path through ocean space, coastal property, regulation, repair capacity, corporate investment, military history, and local conflict.

That shift from topology to topography is the book's core discipline. Topology asks how nodes connect. Topography asks where the connection runs, what it crosses, what it avoids, who maintains it, what histories it reuses, and what happens when it lands. The difference matters because politics often appears at the landing point, not in the network diagram.

Read this way, the cable is neither a neutral pipe nor a romantic relic. It is a working arrangement. Its path embodies old routes, new incentives, technical constraints, environmental permissions, institutional risk calculations, and public invisibility. The internet becomes less like a placeless cloud and more like a set of maintained corridors through the world.

Landing Stations

Starosielski is especially good at the shore. Cable systems spend much of their public life unseen, but they have to come ashore somewhere. A landing station is where the planetary network becomes local infrastructure: a building, a beach, a fence, a permit, a security practice, a neighbor, a labor process, and a vulnerability.

This is where the book is useful for AI infrastructure. Model systems are often discussed at the level of benchmark, parameter count, product feature, or policy document. But every service also has landing points: data centers, cloud regions, fiber routes, exchange points, power substations, water systems, workforces, content-moderation queues, and customer institutions. The political question is not only what the model can do. It is where the system touches the world and who lives with that touch.

The landing point also breaks the fantasy that resilience is only a cybersecurity problem. Cable resilience involves geography, maritime activity, repair ships, permitting, route diversity, physical security, spare parts, cross-border coordination, and the less glamorous work of maintenance. The same lesson applies to AI: reliability is not achieved by model evaluation alone. It depends on the material and institutional system around the model.

The Pacific Network

The Undersea Network keeps returning to Pacific sites because submarine cables do not only connect powerful centers. They also pass through islands, colonial histories, military routes, tourism economies, environmental zones, and communities whose relationship to the network cannot be summarized by the word "node."

That distinction matters. A node is a point in a system. A place is lived, governed, contested, remembered, repaired, and exposed. Technical maps often need nodes because systems require abstraction. Starosielski's warning is that abstraction becomes politically dangerous when the node replaces the place in public memory.

This is one reason the book belongs near AI governance. AI systems also convert places and people into nodes: users, accounts, workers, edge devices, datasets, regions, tenants, assets, and risk objects. Those abstractions are administratively useful. They become harmful when the abstraction is allowed to stand in for the full social, environmental, and historical reality it compresses.

The AI Infrastructure Reading

Read on June 25, 2026, Starosielski's book looks less like media-history recovery and more like an AI-infrastructure manual. ITU's 2025 Global Connectivity Report chapter said submarine cables are the hidden backbone carrying more than 99 percent of international data flows, that investment rose from USD 0.8 billion in 2015 to USD 9.7 billion in 2025, and that hyperscale technology companies now play a leading role in financing new infrastructure. That matters because cloud and model providers are not merely customers of the network. They increasingly shape its map.

The FCC made the AI connection explicit in its August 7, 2025 statement on submarine cable buildout and security, saying submarine cable systems are key to AI and next-generation technologies. That framing matters because it moves submarine cables from telecom background to AI governance foreground.

A prompt sent to a hosted model is not just a string entering a model. It is a packet path through devices, access networks, exchange points, terrestrial fiber, submarine systems, landing stations, cloud regions, GPUs, storage, logging systems, policy filters, and return routes. The visible interface may be conversational. The operating reality is infrastructural.

This changes the meaning of "cloud dependence." An institution using AI through a remote API is also depending on route diversity, cable repair capacity, coastal regulation, data-center siting, vendor contracts, grid load, water use, jurisdictional control, and the business decisions of companies that own or lease capacity. The system is not only a software subscription. It is a geography of dependency.

Governance and Resilience

The governance lesson is practical: resilience is not a slogan about redundancy. It is a set of documented powers over routes, landing points, suppliers, permissions, repair, traffic shifting, outage communication, and public-service continuity. A cable can be technically redundant while still politically fragile if all plausible alternatives depend on the same vendor, landing jurisdiction, repair bottleneck, cloud region, or opaque failover plan.

Starosielski also complicates resilience language. Resilience is not just redundancy for rich centers. It can involve underserved regions, island geographies, repair delays, permitting bottlenecks, and the uneven capacity to recover from disruption. ITU's 2026 Porto Summit guidance emphasized repair times, regulatory procedures, geographic diversity, redundancy, and risk mitigation. Those are governance topics, not background engineering details.

Security framing needs the same discipline. Sabotage, espionage, and foreign-adversary control matter, and the FCC and EU sources show governments taking them seriously. But ordinary cable faults, permitting delay, fishing and anchoring, terrestrial backhaul failure, supplier concentration, and unclear incident authority are also governance problems. A policy that speaks only in crisis language can miss the maintenance conditions that keep the network working most days.

The transparency problem is real. Public accountability does not require publishing exploit-ready route maps or landing-station security plans. It does require enough public, regulator, and procurement evidence to know who owns or controls critical routes, what dependencies matter, which failover assumptions are tested, which communities host the infrastructure, and what happened when a service failed.

The AI version is clear. If schools, hospitals, emergency services, courts, government offices, or workplaces build everyday operations around cloud AI systems, they need manual fallbacks, local continuity plans, vendor-exit rights, incident procedures, and public accountability for outages and degraded modes. A resilient AI institution is not one that assumes the model will always answer. It is one that can keep serving people when the network, vendor, model, or policy layer does not.

Cable Dependency Register

The practical artifact this review adds is a cable dependency register. For any high-impact AI deployment, public workflow, research system, or critical communication service, the register should name the model and vendor, hosting and inference regions, major network providers where known, cloud and data-center dependencies, landing jurisdictions that matter to cross-border service, terrestrial backhaul assumptions, fallback paths, outage behavior, data-residency claims, logging locations, incident contacts, repair assumptions, exit plans, and the public services that would fail if connectivity degraded.

The register should have public and restricted fields. Public fields can state that a service depends on cross-border connectivity, cloud regions, route diversity, vendor support, and tested fallback procedures. Restricted fields can hold exact routes, exploitable chokepoints, security plans, supplier details, and incident playbooks for regulators, auditors, operators, and authorized procurement officials. Secrecy should protect infrastructure, not erase the dependency.

For AI systems, the register belongs beside an AI system inventory, AI procurement file, data-residency assessment, audit trail, digital infrastructure register, and vendor governance record. The point is not to pretend every packet path is knowable from outside the operator. It is to keep institutions from certifying cloud AI systems while the route, repair, and failover layer remains an undocumented assumption.

The register also gives Starosielski's topography a governance form. Instead of saying "the internet is material" and stopping there, it asks where the material dependency enters the institution's own records. If the answer is nowhere, the institution has not governed the infrastructure. It has only used it.

Where the Book Needs Friction

The Undersea Network is not a contemporary AI book. It predates the current foundation-model boom, hyperscaler cable strategies at today's scale, agentic AI products, and the 2025-2026 regulatory fight over submarine cable security. Its strength is method, not direct policy prescription.

The book also has a focused geography. Its South Pacific emphasis is a virtue because it prevents the network from becoming abstract, but it means readers still need other work on the Atlantic, Arctic routes, African connectivity, cable ownership, national-security review, repair-ship availability, terrestrial backhaul, exchange points, and data-center geography.

There is also a tension in making hidden infrastructure visible. Public understanding supports accountability, but full visibility can create security concerns. The answer is not secrecy as a default or exposure as a reflex. It is layered accountability: enough public knowledge for democratic governance, enough operational confidentiality for security, and enough institutional documentation that affected communities are not asked to trust a system nobody will describe.

What This Changes

The practical lesson is to stop treating AI as an interface floating above infrastructure.

Every serious AI deployment should ask: where does the signal travel, who owns or controls the path, what happens at the landing points, what jurisdictions touch the route, who can repair it, who can interrupt it, what communities host the supporting infrastructure, what energy and water systems are implicated, and what institutional service fails if the path fails?

That is not anti-cloud nostalgia. It is source discipline applied to infrastructure. The more institutions outsource cognition, administration, search, writing, triage, education, and public interaction to remote model systems, the more they need to understand the physical network underneath that outsourcing.

Starosielski's enduring value is that she gives the internet a floor. Once the network has a floor, AI has a geography. Once AI has a geography, governance can ask material questions before the system hides inside convenience.

Source Discipline

This review separates four source layers. Book metadata and publication context come from Duke University Press and Google Books. The companion-project description comes from Duke and Surfacing. Current submarine-cable and AI-infrastructure context comes from ITU's April 2026 backgrounder, ITU's 2026 Porto Summit release, ITU's 2025 Global Connectivity Report chapter, the FCC's August 7, 2025 submarine-cable statement, and European Commission cable-security materials. Internal links supply local governance vocabulary, not evidence for external factual claims.

The analogy is limited. Starosielski did not write about foundation models, agent platforms, or cloud AI procurement. The narrower claim here is that her infrastructure method makes those systems legible: AI depends on physical routes, coastal sites, operators, jurisdictions, repair regimes, and places that should not disappear behind the language of cloud service.

Cable statistics need precise verbs. "More than 99 percent" claims generally refer to international or intercontinental data traffic, not every local internet interaction or every AI call. A cable map is not proof of data residency. A regulator action is not proof that a route is safe. A company or government resilience claim is not an outage test. Governance-grade claims should preserve the difference between a publisher record, an industry statistic, a regulator order, a policy plan, an operator announcement, and an audited incident record.

This page does not claim that any AI system is conscious, divine, or AGI. It treats model access as a material dependency that runs through communications infrastructure, cloud providers, contracts, and public-service continuity plans.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books