Legal infrastructure and the Six Layer Cake

The Missing Layer

Jensen Huang's March 2026 essay lays out AI as a five-layer stack: energy, chips, infrastructure, models, applications. He argues that each layer reaches down through the ones beneath it, and that the whole thing is really an industrial transformation rather than a software story. For the hardware and the economics, this is probably a useful way to describe the world with AI.

What it leaves out is law.

This is probably less of an oversight than a feature of how a well-functioning legal system is largely invisible to the people relying on it.  This is also why law tends to get described as context rather than infrastructure. It surfaces when a contract turns out not to bind the person it was supposed to bind, or when an asset that has been treated as owned is held by a court to belong to someone else. The rest of the time it is just the medium everything else moves through.

That arrangement worked when law could afford to sit next to the stack and intervene at the edges, because the edges were where the interesting legal events happened: formation, breach, dispute. Everything in between was execution, and execution did not raise novel questions.

Huang's central claim is that this world is over. Intelligence is now generated in real time, every response is newly created, every answer depends on the context provided. If that is true of the computation, it has to be true of the legal substrate the computation operates in, and yet to date it is not.

AI needs Techno-Legal Infrastructure

Consider agentic procurement. One organisation's AI agent will negotiate terms with another organisation's AI agent, settling price and delivery and penalty clauses in the time it takes a human to read the email announcing the deal. The same pattern will show up in logistics, in financial services, in any domain where transactions can be decomposed into structured decisions and handed off.

None of the legal questions this raises have clean answers yet. Whose agent bound whom, and to what. What authority, if any, the agent actually had at the moment it acted, as opposed to the authority its principal had, or thought the agent had. What law governs an exchange where the two agents are running in different jurisdictions and their principals are incorporated in two more. Who owns what gets produced, and who is on the hook when something goes wrong. Whether what just happened is even a transaction in the sense a court would recognise, or whether it is a sequence of machine events that a court will later have to reconstruct into one.

These questions cannot be answered by adding an analogue legal review step to the application layer. That review would have to happen at the speed of the agent, and human legal review does not happen at that speed. This is the legal attribution gap, and it is not going away by itself.

Law as infrastructure

The response cannot be more regulation written in the old register. Statutes and case law were built for a world of prerecorded software and human decision-makers, and they will keep being built for that world unless something changes in how legal obligations are expressed and enforced. The legal layer needs to be rebuilt on the same first principles Huang applies to the rest of the stack, which means operating in real time, at machine speed, as infrastructure rather than as an overlay consulted when something breaks.

In practice that means legal identity for AI systems that other machines can actually resolve, rights and obligations that are structural components of the transaction rather than clauses filed away for later, and attribution established at the moment of action rather than reconstructed afterwards through discovery. Nooriam is building this under the Techno-Legal Infrastructure Framework.

The sixth layer

So the cake needs another layer. It can be drawn above or below energy - let the scientists and lawyers argue that one! Either way, it is a foundational layer that determines whether the chips can be financed, whether the infrastructure can be insured, whether the models can be trained on the data they are being trained on, whether the applications can transact at all, and whether anyone can be held responsible when something eventually goes wrong.

Huang is right that this is the largest infrastructure buildout in human history. What his essay does not quite reach is that an infrastructure buildout of this kind needs a legal infrastructure built alongside it, not bolted on afterwards. Law working well is invisible. Law working at the speed of agents has to be designed.