Infrastructure Dependency

In the AI Story

The training infrastructure comes first. A frontier language model requires clusters of thousands of specialized processors — GPUs or TPUs — connected by high-bandwidth networks and powered by dedicated electrical substations. Training a single frontier model consumes electrical power equivalent to tens of thousands of households running continuously for weeks. The hardware represents manufacturing concentration that makes the smartphone supply chain look diversified: the most advanced training chips are fabricated almost exclusively by Taiwan Semiconductor Manufacturing Company (TSMC) using lithography equipment from a single Dutch company, ASML. The entire AI training infrastructure of the Western world depends on a supply chain passing through a single island in the western Pacific.

Inference infrastructure is equally consequential and, in aggregate, more expensive. Training is a one-time cost; serving a model to millions of users requires continuous computation, power, cooling, and bandwidth. The International Energy Agency has flagged AI data centers as a significant and growing fraction of global electricity demand. In several regions, data center construction has been delayed or blocked by insufficient grid capacity. The magic is bumping against the physics of power generation and distribution.

The democratization Edo Segal celebrates operates at the interface level, not the infrastructure level. The developer in Lagos accesses Claude Code through a subscription; the subscription gives her access to a capability, not to the infrastructure that produces it. She cannot train her own model, run her own inference cluster, or choose an alternative provider if the one she depends on changes pricing, terms of service, or content policies. She is, in the most literal sense, a tenant on someone else's infrastructure.

The concentration is measurable. As of 2026, the three largest cloud providers — Amazon Web Services, Microsoft Azure, and Google Cloud — control approximately two-thirds of the global cloud infrastructure market. The AI capabilities users access through subscriptions are, in almost every case, running on infrastructure owned by one of these three (or, in the case of Anthropic's Claude, hosted on one of them). Moore's semiconductor history provides the precedent: fab costs escalated from millions in the 1970s to tens of billions today, driving industry consolidation to three companies worldwide — TSMC, Samsung, Intel — with geopolitical significance. The same consolidation dynamic is now visible in AI, compressed onto a faster timeline.

Origin

The framework of infrastructure dependency draws on decades of analysis of the semiconductor supply chain, particularly Chris Miller's Chip War (2022), and on emerging analysis of AI infrastructure concentration by researchers at Epoch AI, the AI Now Institute, and the Center for Security and Emerging Technology. The specific framing — that interface democratization masks infrastructure concentration — is articulated in this volume as a synthesis of Moore's industrial history with contemporary observations of AI industry structure.

Key Ideas

The interface is not the infrastructure. Democratization at the user level — a hundred-dollar subscription that unlocks AI-augmented building — does not democratize the underlying infrastructure, which remains concentrated in a handful of entities.

Training is capital; inference is metabolism. Training costs are one-time capital expenses; inference costs are ongoing metabolic expenses that compound with usage.

Supply chain concentration is geopolitical. The fabrication of advanced AI training chips depends on a supply chain passing through Taiwan, creating vulnerabilities no amount of software innovation can compensate for.

Dependency relationships are tenant-to-landlord. Users of AI infrastructure cannot choose alternative providers, cannot replicate the infrastructure, and have limited leverage to influence its terms.

Opacity compounds dependency. Training data, model architectures, inference optimizations, and cost structures are proprietary, leaving users without visibility into the foundation their work rests on.