Power and dependencies¶

The Patrician has observed that Ankh-Morpork’s essential services have, through entirely natural market forces, concentrated among a small number of providers. The water supply, waste disposal, and food distribution are each controlled by guilds that achieved their positions through superior efficiency, aggressive acquisition of competitors, and the considerable advantage that comes from being the incumbent everyone already depends on. The concentration is excellent for the guilds. It is mostly adequate for the city. The Patrician maintains careful relationships with guild leaders because the alternative is the city discovering quite how dependent it has become on organisations whose interests and the city’s interests are related but not identical.

The internet’s infrastructure has concentrated similarly. Three companies provide the majority of cloud computing services that underpin enormous portions of global commerce. TSMC manufactures most advanced semiconductors. Nvidia supplies most AI accelerators. These companies are not monopolies in the legal sense, because alternatives exist. They are dependencies in the practical sense, because the alternatives are inferior, more expensive, or simply not available at the scales most customers need. The distinction is meaningful in antitrust proceedings and largely irrelevant to the people experiencing the dependency.

The cloud situation¶

Three companies control the majority of global cloud infrastructure, which is fewer than the fingers on one hand and considerably fewer than the number of organisations depending on that infrastructure. Amazon Web Services remains the largest. Microsoft Azure is growing fastest. Google Cloud is third but substantial. Together they host enormous fractions of the internet’s services.

The switching costs that accumulate as organisations integrate deeply with provider-specific services make leaving difficult even when dissatisfaction is genuine. An organisation using proprietary databases, serverless functions, and provider-specific management tools faces substantial reengineering to migrate to alternatives. The costs are not insurmountable. They are sufficient that most organisations only consider switching under duress, which is a comfortable position to be in if you are the incumbent.

Multi-cloud strategies provide psychological comfort and modest actual risk reduction. Maintaining genuine redundancy across multiple providers costs more and requires more operational expertise than single-provider strategies. Most organisations implement multi-cloud reactively after experiencing outages, and even then the implementation typically means different providers for different purposes rather than the seamless failover that the strategy implies. The Patrician does not consider this adequate preparation, but he notes it is the normal amount of preparation.

The semiconductor chokepoint¶

TSMC manufactures most of the world’s advanced semiconductors in a geopolitically contested location. This is not a secret. It is a widely acknowledged strategic vulnerability that everyone is aware of and nobody has quickly resolved because the resolution requires building competitive fabrication capacity elsewhere, which costs tens of billions of euros, takes years, and currently produces chips that are less capable than TSMC’s at higher cost. Progress is being made. It is being made slowly.

Nvidia has achieved a position where they supply the majority of AI accelerators, the software framework that runs on them, and increasingly the networking that connects them. This vertical integration means building large-scale AI infrastructure increasingly means buying Nvidia throughout the stack. The products work well together because they were designed as a system. This is genuine value. It is also, from the customer’s perspective, the sort of value that requires you to keep buying from the same vendor. Nvidia is sanguine about this.

Data as infrastructure¶

Companies that accumulate large datasets gain advantages that compound over time: more users provide more data, which enables better services, which attract more users. New entrants cannot compete without equivalent data, which they cannot obtain without equivalent users, which they cannot attract without competitive services that require the data. This circle is self-reinforcing and explains why certain markets remain concentrated despite the apparent absence of explicit barriers to entry.

The Patrician notes that data portability requirements, while well-intentioned, provide individual users the ability to export their own data rather than the aggregate behavioural data that makes recommendation systems work. The portability is real. Its practical effect on competition is modest.

The Patrician’s assessment¶

The current dependency structure is efficient. It is also fragile in ways that become apparent only when something fails, at which point the failure propagates through supply chains and technology stacks in ways that are difficult to predict and expensive to address.

Everyone knows the single points of failure. TSMC’s geography. The cloud oligopoly. Nvidia’s AI dominance. The open source maintenance gap. These are not mysteries. They are accepted risks because eliminating them would require investment and structural changes that reduce efficiency and that most participants prefer not to fund.

The Patrician observes that concentrated infrastructure works excellently until it doesn’t, that the transition between these states is rapid and unfortunate, and that preparing for the “doesn’t” possibility while assuming the “works excellently” phase continues indefinitely is the standard organisational approach. He has seen this before. He takes notes.