The metaphor and who is selling it
The image is memorable: energy at the bottom, then chips and computing infrastructure, then cloud data centres, then the models, and on top the applications a bank or a hospital actually buys. The argument that followed was that every layer has to be built by someone, that this is the largest infrastructure build-out in history, and that the economic benefit will land at the top.
It is worth noting who was speaking. The company sells the second layer, so it has an interest in everyone agreeing the other four need building. And none of the labour figures quoted on stage came with a source attached. But set the salesmanship aside, because the metaphor is genuinely useful — for what it leaves out.
Count the layers you control
Walk the cake from the enterprise’s side of the table. Energy: you buy it. Chips: you will never see them. Data centres: rented, from one of a handful of providers. Models: licensed, and the leading ones change every few months. Applications: purchased, or built on top of the layer below.
An enterprise controls none of the bottom four layers and, for the most part, merely selects the fifth. If that were the whole picture, AI adoption would be a procurement exercise, and every organisation with a budget would get the same result.
They do not. The organisations getting value from AI and the ones accumulating pilots are buying from the same vendors. The difference lies in a layer that is not in the metaphor at all.
The layer that is not on the menu
That layer is the organisation’s own information: the warehouse, the systems of record, the documents, the workflows, and the decades of accumulated structure that say what a customer is, what an order is, which rule applies when. It is where the applications at the top of the cake have to plug in, and it is the only layer nobody can buy from a vendor, because it is specific to the organisation.
It is also the layer most AI programmes underinvest in, precisely because it is not new. The budget goes to the model and the application; the data platform is assumed to be ready. It rarely is. Access is improvised, classification is patchy, the systems of record expose no clean interface for an agent to act through, and the result is a capable model connected to nothing.
What the good examples have in common
The examples cited on stage are instructive, and not for the reason intended. Radiology has adopted AI faster than almost any field, and the number of radiologists has gone up, because reading a scan is a task rather than the purpose of the job. Nurses spend close to half their time charting; AI transcription gives that time back to patients.
Look at what those two have in common. Neither is a new application sitting beside the work. Both are AI integrated into an existing workflow, at a specific step, inside the systems the clinician already uses — the imaging platform, the electronic health record. The value came from the integration into the workflow and the record, not from the model, and not from a new layer on top.
That is the pattern we see everywhere it works: AI arriving inside the systems people already run, at the point where a decision or a document is produced, reading from data the organisation already governs. The cake’s top layer, in other words, is only valuable when it is wired into the sixth.
Where the investment should go
If you accept the metaphor, the strategic conclusion follows. The bottom four layers will be built whether or not any given enterprise participates, and their prices will be set by the market. The top layer is a buy-versus-build decision that will be revisited annually. The layer worth investing in, because it compounds and because no competitor can purchase it, is the one you own.
Practically: make the data platform genuinely ready — governed access, classification that travels, a catalogue that tells an AI system what a field means. Give the systems of record the interfaces an agent needs to act through safely. Choose the first AI use cases by where they land in an existing workflow, not by what the model can demonstrate.
The industry will keep talking about the build-out, because that is what it sells. The enterprise conversation should be about the layer that was never for sale.
