A language model trained on user data from India, hosted on servers in Virginia, whose intellectual property is registered under a subsidiary in Ireland, and which generates licensing revenue billed to a company in Mexico. Where the taxes on the value that model generates actually get paid has no clear answer. This fiscal ambiguity is the void that emerges because traditional frameworks assume value is created in an identifiable place.
Firms like Thomson Reuters have documented that DEMPE functions still lack a specific tax framework for artificial intelligence. Companies building these models can freely choose which jurisdiction hosts the development of the asset. And therefore where it gets taxed.
The regulatory void exists. The urgency narrative surrounding it deserves scrutiny. When people talk about the need for proactive governance on AI taxation, one question arises. Proactive for whom?
The companies that profit most from the ambiguity already had time to structure their assets. This resembles the Apple and Ireland case. Between 2004 and 2016, Apple used a structure known as the double Irish with a Dutch sandwich. The goal was to move profits from markets like Germany, France, or the United Kingdom into Irish subsidiaries that transferred royalties to the Netherlands and finally to jurisdictions with minimal corporate taxes. Google followed similar routes. The European Commission concluded in 2016 that Ireland had granted Apple illegal tax advantages worth up to thirteen billion euros. The litigation dragged on for years. By the time it was resolved, the structure had already served its purpose.
The coordinated response arrived late. The OECD's BEPS project formally began in 2013 following the G20's call. Countries watched digital multinationals pay single-digit effective tax rates while generating enormous revenue in places where they barely had physical presence. The framework, with its fifteen action items, took years to be substantially implemented. More than ten years passed between the emergence of these strategies and coordinated regulatory action. Enough time for the tactics to become entrenched and refined.
The AI industry is repeating this pattern. A language model requires no office, no employees, no permanent physical infrastructure. It's trained in one data center, deployed in another, and licensed from a third. None of those locations need to coincide with where the value is actually created. Intellectual property intangibles were already hard to pin down. AI models represent a further leap. They are numerical weights. Files that replicate with zero physical friction.
I've observed in systems architecture how this distributive capability is no accident. It's the design itself. Who really benefits from this framework's continued absence? AI companies, without a doubt. Also an entire tax advisory industry that charges to design the navigation routes. The more ambiguous the rule, the more valuable those services become. The same firms that warn about regulatory urgency often supply the very instruments to postpone it. This isn't conspiracy — aligned incentives produce convergent outcomes without anyone needing to orchestrate them.
The same pattern shows up in other corners of the ecosystem. Amazon warns about risks in Anthropic's models, a company in which it has invested billions. OpenAI and Anthropic fund the institute that evaluates the safety of their own products aimed at minors. Whoever creates the problem also designs the audit.
The losers are emerging economies. That's where user data gets generated, where labeling and training happen, and where consumer markets are growing. They lack the diplomatic weight to demand that value be taxed where it arises. BEPS already documented that these countries lost proportionally more tax base than advanced economies.
I don't believe this reflects some malicious central plan. Corporate incentives and the absence of regulatory friction produce similar results. That doesn't lessen the damage. It makes it harder to fight because there's no single lever to pull.
From a technical standpoint, this is almost elegant and also unsettling. AI models are built to be distributed and replicable. That property drives their capacity for growth. It becomes a challenge when tax systems keep looking for a fixed location. If a system is designed to locate itself wherever it pays the least tax, that's exactly where it ends up. It's not a failure. It's the system operating as designed.
This connects to themes I develop in Stones Don't Lie. Technology governance rarely regulates who decides the direction of development. It focuses on how the system behaves once built. European Union regulation classifies risks and demands transparency. It does not, however, address who controls the architecture that makes it possible to dodge the tax question from the outset.
I still don't have a clear solution, and it would be dishonest to pretend otherwise. The window between identifying the problem and reaching a coordinated response has historically been wide. In that interval, the value has already moved, already been structured, already become part of the normalized business model. While public forums debate ethics and safety, it's worth asking whether that conversation also serves as a smokescreen.
Will we ever align taxation with where value is actually generated, or will we keep letting technical architecture dictate the tax rules?
