Governance Arbitrage

Rules for technology are not the same everywhere. What is forbidden in the European Union may be allowed in Singapore or Texas. Governance arbitrage describes the strategy of exploiting exactly these differences. A company picks the country with the loosest regulations for a given activity and carries it out there. The word arbitrage comes from finance, where it means buying the same item cheaply and selling it on at a higher price. Applied to rules, it means: you choose the place where an action costs the least – not in money, but in obligations.

When rules end at national borders, but software does not

Laws apply within a state or a union of states. An AI model, by contrast, runs on servers somewhere in the world and can be accessed everywhere via the internet. This gap between national law and global technology is the ground on which governance arbitrage grows. Whoever regulates strictly risks companies simply working elsewhere.

For states, this creates an uncomfortable trade-off. Strict rules protect citizens from surveillance, discrimination by software, or unsafe products. At the same time, they can drive away investment and jobs. Experts call the risk that all countries undercut one another a race to the bottom. Whether this race actually exists is disputed – but it is documented that large corporations also choose locations based on the legal situation.

It is important to distinguish this from breaking the law. Governance arbitrage is normally legal. No one forges documents or ignores a court ruling. One merely takes advantage of the fact that a different legal jurisdiction assesses the matter differently. This is exactly what makes it so politically difficult to combat: there is often no one to sue.

The levers: location, legal form, timing

The simplest form is relocation. A model is trained wherever copyright law permits the use of large volumes of text. The finished model is then offered worldwide. Training – the expensive learning process in which the model derives patterns from example data – thus takes place under one legal system, while sales take place under another.

A second lever is corporate structure. Corporations set up subsidiaries in various countries and distribute tasks among them. The research department is located in one place, server operations in a second, sales in a third. A different set of rules then applies to each task. This has worked similarly with taxes for decades.

A third lever is timing. New laws rarely take effect immediately, but only after transition periods of two or three years. Some providers deliberately release a product beforehand, because milder rules still apply at that point. A common misconception, by the way, is that governance arbitrage concerns only taxes. It affects every set of rules: data protection, liability, environmental requirements for data centers, labor law for the people who review training data.

The term in news about AI laws

The term appears most often in connection with the AI Act, the European Union’s AI law. It classifies applications by risk and imposes high requirements on risky systems. Providers regularly announce that they will not initially roll out a new feature in Europe. Whether this is a genuine technical hurdle or pressure on policymakers is often impossible to determine from the outside.

The effect is also visible with data centers. They require enormous amounts of electricity and cooling water. Regions with cheap electricity and lenient environmental regulations therefore attract construction projects. For investors, this is an important metric, because energy costs place a lasting burden on the operation of AI services.

As a countermeasure, regulators rely on two things. First, on market-location principles: whoever has customers in Europe must comply with European rules, regardless of where the servers are located. Second, on international coordination, for instance within the G7 or OECD. So far, both approaches only work partially, which is why the term remains firmly anchored in the debate over AI regulation.