The stack nobody talks about

Published: 27 Jul 2026

When most people think about artificial intelligence, they picture software. A model. A chatbot. Something that lives in the cloud and answers questions.

But that picture is radically incomplete.

The actual economic reality behind a single AI response involves mining operations on multiple continents, chemicals purified to eleven nines, machines that cost hundreds of millions of dollars each, and enough electricity to power a small city. The conversation you are having with an AI assistant sits at the very top of a stack that goes a very long way down.

Stack of industries #

The instinct is to call it a supply chain, but that implies something linear. One company hands a part to the next, which hands it to the next. Neat and sequential.

What the AI industry actually looks like is more like this:

People and businesses buying AI products
            ↓
AI applications and model providers
            ↓
Cloud-computing companies
            ↓
Data centres, electricity and networks
            ↓
AI servers and rack-scale systems
            ↓
GPUs, CPUs, memory, networking and storage
            ↓
Chip design, fabrication and packaging
            ↓
Semiconductor manufacturing equipment
            ↓
Specialised chemicals, components and materials
            ↓
Mining, refining, energy and industrial machinery

Each level is its own industry. Each has its own companies, economics, bottlenecks and risks. A problem at any level can ripple upward and stop everything above it — including the model you just asked to summarise a document.

Let's walk through each level, from the top to the bottom.

Level 1 - Where the money enters #

At the top are customers. Businesses buying API access, consumers paying for subscriptions, enterprises deploying AI agents, advertisers running recommendation systems. The companies collecting that money — OpenAI, Anthropic, Google, Microsoft, Meta, Amazon — convert it into purchases further down the stack.

The fundamental economic question at this level is simple and not yet fully answered: does the revenue generated by AI actually exceed the enormous cost of building and running it?

NVIDIA's fiscal 2026 results^[1] give you a sense of what "enormous" means here. Its Data Center segment alone generated $62.3 billion in revenue in a single quarter. That is not NVIDIA's profit — it is what companies are paying NVIDIA for the hardware to run AI. Every dollar of that eventually traces back to a customer paying for a subscription, an API call, or a cloud service.

Level 2 - Cloud platforms #

Cloud providers sit between the model companies and the physical hardware. They sell compute by the hour: GPU-hours, reserved clusters, managed AI services, storage, networking. AWS, Azure, Google Cloud and Oracle are the big names, but specialist GPU clouds are growing rapidly.

Their business model has an interesting tension. Hardware becomes an accounting expense spread across several years. Customers pay continuously for usage. That means a cloud operator is essentially a bank that lends out compute — and, like a bank, it is destroyed by idle assets.

An expensive GPU that sits unused is a disaster. The depreciation clock keeps ticking whether the chip is running models or sitting dark.

Level 3 - The data centre #

A data centre is a lot more than a warehouse full of computers. It is a small industrial city.

The site needs land, structural steel, concrete, roads, fire suppression, fibre connections, water access and planning approval. Getting permits for a large AI facility can take years.

The electrical infrastructure is its own engineering project. Utility grid connections, substations, transformers, high-voltage switchgear, power distribution units, uninterruptible power supplies, backup batteries, diesel generators, copper busbars and cabling running throughout. For many AI projects, securing grid power is harder and slower than securing the land.

Cooling is now a first-class problem. Traditional server rooms used air. Dense AI racks generate so much heat that air cannot remove it fast enough. Modern AI data centres increasingly use direct liquid cooling: cold plates attached to the chips themselves, coolant distribution units, pumps, heat exchangers, cooling towers, specialised fluids and the pipework to connect it all. AWS has built direct-to-chip liquid cooling systems with recirculating coolant specifically because the heat density of clustered AI chips makes air cooling impractical.

The International Energy Agency reported that data-centre electricity consumption rose 17% in 2025, with AI-focused facilities growing even faster.^[2] It also estimates that natural gas and coal together may meet more than 40% of the additional electricity demand from data centres through 2030.^[3]

The energy system, in other words, is now part of the AI supply chain.

Level 4 - The server rack #

Inside the data centre are racks of AI servers. Not ordinary servers — systems built specifically for AI workloads, running platforms like NVIDIA's GB or Rubin architecture, AMD Instinct, Google TPUs, or AWS Trainium.

A single AI server is already a complex assembly:

An accelerator package (the GPU or custom chip) with its own memory, interposer, and substrate
A host CPU, conventional RAM, storage and network interface cards
Power supplies, fans, pumps and cold plates
Motherboard and printed circuit boards

And then the rack-level infrastructure: high-speed switches, optical transceivers, copper and fibre cables, power distribution, cooling distribution.

Here is something that gets overlooked: a single accelerator is not useful on its own. Large AI workloads require thousands of accelerators communicating at enormous speeds. That makes the network — the switches, cables and interconnects — nearly as important as the chips themselves. An incredibly powerful GPU connected by a slow network is like a Formula 1 car with a dirt road to drive on.

Level 5 - The chip package #

A modern AI processor is a small system assembled inside one package.

The logic die does the actual computation — matrix multiplications and other AI operations, billions of times per second. Companies like NVIDIA, AMD, Google, and Amazon design these. Most of them are fabless, meaning they design the chip but pay someone else to manufacture it.

High-bandwidth memory (HBM) sits right next to the processor. It is not conventional RAM. HBM stacks multiple memory dies vertically on top of each other, uses tiny connections called through-silicon vias to link them, and delivers data to the processor much faster than any memory module you would find in a normal computer. Producing HBM requires its own complex manufacturing process: thinning wafers, through-silicon vias, stacking dies, bonding and extensive testing.

The interposer connects the processor to the HBM stacks across extremely dense electrical pathways. TSMC's packaging technologies — CoWoS, InFO, SoIC — allow multiple dies to communicate as if they were one chip while keeping power consumption under control.

This matters for understanding shortages. When people say "GPU shortage," they often mean a shortage of the whole assembled package. An apparent GPU shortage may actually be a shortage of HBM, or interposers, or package substrates, or packaging capacity. The chip itself may be fine. The surrounding system may be the constraint.

Level 6 - Chip design #

Before anything is manufactured, engineers spend years designing it.

Designing a modern AI chip means placing and connecting billions of transistors. Nobody does this manually. They use electronic design automation (EDA) software for circuit design, physical layout, timing analysis, power analysis, simulation, verification, manufacturing preparation and thermal modelling.

Three companies dominate EDA: Synopsys, Cadence and Siemens EDA. If you have never heard of them, you might assume they are minor suppliers. They are not. Every major AI chip company depends on their software. A new chip cannot realistically be designed without it.

Chip designers also license building-block intellectual property: CPU cores, high-speed interface circuits, memory controllers, PCIe logic and other reusable components. Arm, for example, provides processor IP that appears inside chips from dozens of companies. This layer consumes almost no physical material, but it captures significant economic value precisely because it is indispensable and has high switching costs.

Level 7 - Fabrication #

The finished design goes to a foundry. TSMC in Taiwan is the dominant player. Samsung and Intel Foundry are the significant alternatives.

Inside the foundry, a silicon wafer passes through hundreds of processing steps, many repeated multiple times:

Deposition — adding microscopic layers of material
Photoresist coating — applying a light-sensitive chemical
Lithography — projecting a circuit pattern using light
Etching — removing selected material with extreme precision
Ion implantation — altering the electrical properties of specific regions
Cleaning — removing particles and chemical residues
Chemical-mechanical polishing — flattening the surface to atomic smoothness
Metrology and inspection — measuring dimensions and detecting defects
Testing — identifying which dies on the wafer work correctly

Building a leading-edge fabrication plant costs tens of billions of dollars including the building, tools, and supporting infrastructure. TSMC reported a 59.9% gross margin in 2025^[4] — a number that reflects just how much value accrues to a company that can do something essentially nobody else can at the same scale.

Level 8 - Chips manufacturing machines #

To manufacture chips, the foundry needs machines. And those machines are their own specialised industry.

Lithography is the most critical step. ASML, a Dutch company, is the only manufacturer of the EUV (extreme ultraviolet) lithography systems required for the most advanced chips. An EUV machine fires a high-powered laser at droplets of molten tin, generating a plasma that emits extreme ultraviolet light. That light is shaped by ultra-precision mirrors — not lenses, because EUV light would simply be absorbed by any glass — and projected onto the wafer to define circuit patterns at scales measured in nanometres.

ASML reported €32.7 billion in revenue in 2025 at a 52.8% gross margin.^[5] For a machine that makes other machines, those are extraordinary economics.

Deposition, etching and cleaning equipment comes from Applied Materials, Lam Research and Tokyo Electron. They build the tools that deposit atomic-scale films, create plasma, etch microscopic structures and prepare wafer surfaces.

Inspection and metrology is KLA's territory. Every defect that goes undetected can destroy an expensive die. KLA's systems measure dimensions and detect problems at scales that are simply invisible to everything else.

Level 9 - The machines behind the machines #

Here is where it gets deep. ASML's EUV machine is itself assembled from thousands of specialist components from other companies:

Ultra-precision mirrors and optical coatings (ZEISS, Germany)
High-power industrial lasers (TRUMPF, Germany)
Vacuum pumps, valves and chambers
Precision bearings and vibration-isolation systems
Ceramic components, sensors, power electronics
Robotic wafer handlers and motion-control systems

Each of those components comes from its own supply chain, which reaches further down into machine tools, industrial metals, ceramics, specialty chemicals and energy.

The machine that makes a chip was itself made by machines, which were made by machine tools, using metals, chemicals and energy extracted from the ground.

Level 10 - Materials #

Advanced chips require an extraordinary variety of materials at purity levels that do not exist in ordinary industrial contexts.

Inside the transistors: ultra-high-purity silicon, germanium, gallium compounds, hafnium, cobalt, tungsten, tantalum, copper.

For lithography and processing: photoresists, photomask materials, industrial gases (hydrogen, nitrogen, argon, helium, neon, fluorinated gases), sulphuric acid, hydrochloric acid, hydrogen peroxide, ultra-pure water.

For packaging: copper, tin solder, gold and silver for specialised connections, epoxy compounds, glass fibre, ABF substrate film, ceramics.

And then the data centre itself, which is not particularly exotic but enormous in scale:

Material	Where it goes
Concrete	Building structure, foundations, substations
Steel	Structural frames, racks, generators
Copper	Transformers, motors, busbars, cables
Aluminium	Power systems, heat exchangers
Lithium, nickel, graphite	Backup batteries
Rare-earth elements	Motors, generators, some electronics
Diesel or natural gas	Backup generation
Water	Cooling and semiconductor fabrication

The concrete and steel carry no intellectual property premium. But the facility cannot run without them.

The whole stack #

Physical mass increases as you move down. Profit per kilogram increases as you move up.

A data centre contains thousands of tonnes of concrete and steel. But a small quantity of advanced silicon, software and precision optics determines whether that entire investment is commercially useful.

NVIDIA reported a 71.1% gross margin in fiscal 2026.^[1] TSMC reported 59.9%.^[4] ASML reported 52.8%.^[5] These figures are measuring different things and are not directly comparable, but the pattern is consistent: the companies that control something scarce and technically irreplaceable capture far more value than those handling high volumes of commodity material.

The companies further down the stack — construction contractors, metals suppliers, industrial gas producers — often move more physical material and run more physical infrastructure. But they capture a much smaller share of the total economic value. The intelligence and the scarcity sit near the top.

Five economies #

The simplest way to hold this in your head:

1. Intelligence economy
   Models, software, data and applications

2. Compute economy
   Clouds, servers, networking and storage

3. Infrastructure economy
   Buildings, electricity, cooling and fibre

4. Semiconductor economy
   Chip design, memory, foundries and packaging

5. Industrial economy
   Equipment, chemicals, metals, mining and energy

The public conversation about AI is almost entirely about economy 1. The current AI investment boom is largely happening in economies 2 through 5.

Potential show stopper #

The stack has real constraints at multiple levels simultaneously:

Electricity and grid connections — enough generation may exist nationally but not where you need it
AI accelerators — advanced designs, limited production allocation
HBM — long manufacturing cycles and difficult capacity expansion
Advanced packaging — the processor and memory cannot be combined without it
Leading-edge foundry capacity — a very small number of companies can do this
Lithography and inspection equipment — specialists with very limited alternatives
Qualified materials — semiconductor-grade purity is categorically different from industrial-grade
Skilled people — electrical engineers, fab engineers, chip designers, construction specialists, data-centre operators
Permits and geopolitics — export controls, trade restrictions and geographic concentration

Any one of these can stop the stack above it. That is the key insight. AI is not just a software problem or a chip problem. It is a system of interdependent industries, each with its own capacity constraints, and a slowdown anywhere propagates upward.