The physical infrastructure behind AI

A ten-layer map of what AI is actually built on.

The AI infrastructure map

The infrastructure atlas needs JavaScript. The ten layers are summarised below.

Three things at once, which is why it is not drawn as a stack. On the left, the industrial supply base that everything is made of. In the middle and on the right, the two physical enclosures where AI actually runs: the data centre, and the machine at the edge. Layers 4, 7 and 8 appear inside both, because the same capability class runs at two scales — that repetition is deliberate. Layer 9 is the only path between them, and the physical world above is not a layer at all but the ground the whole structure stands on.

The two enclosures divide the work between them. The data centre runs the central learning loop — stored experience becomes curated datasets, datasets become training and evaluation, and evaluation becomes a validated update sent back out. The machine at the edge does none of that: it senses, decides and acts in real time. Layer 9 carries the traffic between them, which is exactly why safety-critical control must never depend on it. A robot that stops working when the link drops is not autonomous.

The loop

Read as a chain, the infrastructure stack is a value chain. Close it — actuators change the world, sensors observe the change — and it becomes a flywheel. That is the road to AGI, and it is what makes the thesis compound rather than merely grow.

Each turn generates data that could not have existed before that turn. A model trained only on text learns what people have written about physics, not what physics does. Sensors close that gap by returning an error signal from reality itself — which is why a decade of fleet driving data cannot be bought, scraped or licensed by a competitor.

The commercial consequence sits in the two corners nobody funds. Energy gates how fast the loop can turn; the actuator supply chain gates whether it can reach the physical world at all. Everything in between is where the capital has gone.

This report draws its diagrams and tables with JavaScript. Here is the substance without it.

Layer 1 · Energy, power and utilities
Generation, transmission, substations, electrical distribution gear and thermal management. An AI accelerator is a machine for converting electricity into inference — that is the whole business.
Binding constraint. The chokepoint is long-lead-time electrical equipment and turbine hot-section castings, not generation. Global heavy-duty turbine capacity runs 60–70 GW a year against roughly 110 GW of outstanding orders, and the constraint is specialised foundries and skilled welders laid off during the lean 2010s.
Layer 2 · Raw and processed materials
Ore, brine, gas and by-product streams, and the refining, separation and transformation steps that turn them into something a fab, a transformer or a magnet can actually use.
Binding constraint. For most of these materials the chokepoint is refining, separation and qualification rather than reserves in the ground — but not for all of them: high-purity quartz is geologically concentrated, and by-product metals like gallium and germanium cannot be produced to order at any price. Mining is geographically diverse; refining is not. China holds 86–90% of rare earth separation, 48% of copper smelting against 8% of copper mining, and two mines in one North Carolina district supply roughly 80% of the world’s high-purity quartz.
Layer 3 · Semiconductor production ecosystem
Design software and IP, wafer fabrication equipment, materials and chemicals, the foundry itself, and advanced packaging and test. Everything required to turn sand into a working advanced logic or memory die.
Binding constraint. The bottleneck moved. For years it was lithography; it is now advanced packaging — bonding a logic die to twelve stacks of HBM on a silicon interposer. TSMC controls over 90% of global CoWoS output, and CoWoS allocation is effectively the revenue ceiling for every accelerator vendor.
Layer 4 · Compute, memory and networking
The accelerators, the memory that feeds them, and the networking that binds them into a single training fabric. Highest revenue growth in the infrastructure stack, and the shortest moat half-life.
Binding constraint. Three separate chokepoints, and memory is currently the tightest. The most profitable chip company in history is guiding margins down because its memory suppliers raised prices — value transferring up the supply chain against the buyer’s wishes is what a genuine bottleneck looks like.
Layer 5 · Data centres, cloud and edge
The physical buildings and the operating businesses that sell capacity: hyperscale cloud, colocation and GPU-rental neoclouds. Four genuinely different business models that are routinely conflated.
Binding constraint. This is no longer a technology sector. It is a leveraged infrastructure sector wearing technology multiples. Once the build-out is financed with debt, SPVs and depreciating collateral rather than operating cash flow, the right analytical toolkit is project finance, not software.
Layer 6 · Data and knowledge infrastructure
The data assets and the systems that govern them: ingestion, storage, cataloguing, cleaning, annotation, provenance and curation. It supports evaluation, retrieval and monitoring as much as training — and it was previously compressed into a single card inside the model layer.
Binding constraint. The binding constraint here is legal rather than physical. Clear, transferable rights to train on third-party content are contested across several jurisdictions simultaneously, and lineage sufficient to evidence those rights cannot be retrofitted once data is already mixed into a corpus. That is a chokepoint with no lead time to shorten.
Layer 7 · AI models and inference
Frontier training, post-training and alignment, inference serving, and the data pipelines that feed all of it. Where capital expenditure is converted into capability.
Binding constraint. There may not be one. Model weights are a depreciating asset; a fab is not. The durable assets at this layer are proprietary data and distribution, not the weights themselves — which is why the vertically integrated players look structurally stronger than the pure-play labs.
Layer 8 · Agentic software and applications
Orchestration frameworks, agent runtimes, vector and operational databases, observability, security, and the application software customers actually pay for.
Binding constraint. Data gravity and systems of record. The scarce input is context, not intelligence — model capability is available to everyone at collapsing prices, but a company’s customer records, permission structures and process knowledge are not. Whoever owns the context owns the deployment.
Layer 9 · Connectivity and communications
The communications layer that links racks, buildings, continents and machines. It was missing from the earlier map: optics sat inside compute silicon, carrier networks appeared nowhere, and the path between the data centre and the edge was drawn as an arrow.
Binding constraint. The chokepoints are geographic rather than manufactured. Metro conduit sits on rights of way that would not be granted again, and three firms outside China can lay a transoceanic system. Neither constraint is relieved by capital expenditure, which is precisely what makes them durable.
Layer 10 · Embodied AI and autonomous systems
Robots, vehicles, drones and devices, plus the actuators, reducers, sensors and batteries beneath them. Three industries at three completely different stages of maturity, routinely conflated.
Binding constraint. The geography inverts here. Layers 3 and 4 concentrate in the Netherlands, Taiwan, Japan and Korea. Layer 10 concentrates in China, which holds 63% of humanoid component supply and over 70% of rare earth magnet production. Adding this layer to a semiconductor-heavy portfolio adds a second, oppositely-directed geopolitical exposure — it does not diversify the first.