We type a sentence into ChatGPT, Claude, Gemini or another model and, a few seconds later, an answer appears.

The experience feels almost weightless.

A text box.

A cloud.

A response.

But that interface hides something enormous.

Behind every query sits a physical chain that begins long before the data center: chip design, semiconductor fabs, lithography machines, high-bandwidth memory, power grids, transformers, copper, water, steel, land, logistics and billions of dollars of investment.

The best way to understand modern artificial intelligence is not to see it only as software.

It is better understood as a distributed industrial machine spanning much of the planet.

AI as a physical-industrial stack: model, chip, fab, lithography, memory, data center and energy.

The mistake of looking only at the model

When we talk about AI, we usually think about the visible layer:

ChatGPT
Claude
Gemini
Grok

But physically, a model is essentially a huge collection of parameters stored and executed on hardware.

By itself, it does nothing.

To produce a response, it needs an entire infrastructure stack.

A simplified version looks like this:

model
  ↓
accelerators / GPUs
  ↓
semiconductor manufacturing
  ↓
advanced lithography
  ↓
HBM memory
  ↓
servers and networking
  ↓
data center
  ↓
electricity + water + materials + land

The interface we see is only the final millimeter of a much longer chain.

Layer 1: the model

The model is the logical brain of the system.

It contains billions —and in some cases far more— parameters tuned during training.

When we send a prompt, those parameters must be loaded and processed to produce the next token, then the next one, and so on.

The appearance of intelligence depends on an enormous number of mathematical operations being executed very quickly.

That immediately leads to the next layer.

Layer 2: the chips

Modern models rely on specialized accelerators capable of performing matrix operations at massive scale.

Nvidia dominates much of the public imagination around this layer because its GPUs and software ecosystem have become central to the AI boom.

But Nvidia, AMD and other chip designers do not manufacture most advanced chips themselves.

Design and manufacturing are separate industries.

The designer defines the architecture.

The foundry turns that design into silicon.

And this is where one of the most important industrial concentrations in the world appears.

Layer 3: TSMC and advanced manufacturing

TSMC is the world’s largest contract semiconductor foundry and a critical supplier for advanced processors used in AI.

A large share of leading-edge manufacturing capacity remains concentrated in Taiwan, even as TSMC aggressively expands internationally, including massive investment in Arizona.

That makes geography a technology variable.

A revolution that feels digital depends on factories that are extraordinarily difficult to build, operate and replicate.

Having a chip design is not enough.

It must be manufactured at near-atomic tolerances with economically viable yield.

Layer 4: ASML, the machine behind the machine

The dependency becomes even more striking one layer deeper.

The most advanced chips require extreme ultraviolet lithography, or EUV.

In its 2025 annual report, ASML states that it is currently the world’s only manufacturer of EUV lithography systems.

That means a foundational part of the AI industry passes through a Dutch company that builds some of the most complex industrial machines ever produced.

The chain now looks like this:

AI model
   ↓
GPU / accelerator
   ↓
TSMC or another advanced foundry
   ↓
ASML EUV

A global industry worth trillions depends, at key points, on highly concentrated suppliers.

Layer 5: HBM memory

An accelerator can perform an extraordinary number of operations per second, but it needs data delivered at comparable speed.

If memory cannot feed the processor fast enough, the chip spends part of its time waiting.

That is why High Bandwidth Memory, or HBM, has become essential to AI systems.

Instead of treating memory as a conventional peripheral component, HBM stacks memory dies vertically and connects them through very wide interfaces.

Conceptually:

very fast GPU
    +
slow memory
    =
GPU waiting

very fast GPU
    +
HBM
    =
much higher data flow

But advanced stacked memory is difficult to manufacture, consumes industrial capacity and adds another critical supply chain to the system.

AI does not need only better processors.

Everything around the processor must improve at roughly the same pace.

Layer 6: the data center

Then thousands or tens of thousands of these components must be assembled in one place.

An AI data center is not simply a room full of servers.

It is an industrial facility with:

  • high-capacity electrical feeds;
  • cooling systems;
  • high-speed networking;
  • storage;
  • redundancy;
  • physical security;
  • transformers;
  • backup systems;
  • fiber connectivity;
  • enormous amounts of committed capital.

Modern AI racks concentrate more and more power into less space.

That rising power density turns the data center into an electrical and thermal engineering problem as much as a computing problem.

Layer 7: electrons, copper, water and steel

At the final layer, any illusion that we are discussing software alone disappears.

A model needs electricity to answer.

Transporting that electricity requires grids, substations, transformers and copper.

Building facilities requires steel, concrete, land and labor.

Cooling them requires thermal systems that, depending on design and region, may involve significant water use or alternative cooling approaches.

AI therefore reaches industries that only a few years ago seemed far removed from software:

semiconductors
energy
mining
construction
logistics
utilities
nuclear
power grids

That is one of the most important economic transformations of the current boom.

The cloud is becoming heavy again

For two decades, the dominant technology narrative was about dematerialization.

A startup could create enormous value with:

few people
code
cloud
software distributed globally

Instagram became a classic example of a tiny company reaching a gigantic valuation.

AI does not erase that model, but it adds something new.

The competitive frontier is now also measured in:

gigawatts
fabs
memory capacity
racks
transformers
land
permits
capital

Software needs heavy industry again.

The energy bottleneck is already visible

The International Energy Agency estimates that data centers consumed about 485 TWh of electricity in 2025 and could approach 950 TWh by 2030.

That is nearly a doubling in five years.

The IEA also says AI-focused data centers are growing much faster than the total and that their electricity use could triple over that period.

In the United States, data centers could account for close to half of electricity demand growth through 2030.

That changes the discussion completely.

The problem is no longer only:

Can we manufacture enough GPUs?

It is also:

Can we connect enough megawatts in time?

Power grids are built far more slowly than software.

Transformers, transmission lines, turbines, permits and new generation capacity operate on multi-year cycles.

A model can improve in months.

A grid cannot.

When Microsoft needs a nuclear plant

One example captures the shift perfectly.

Microsoft signed a 20-year power purchase agreement with Constellation supporting the restart of Three Mile Island Unit 1, now renamed the Crane Clean Energy Center.

The unit had closed in 2019 for economic reasons.

The project aims to return roughly 835 MW of capacity to the grid to support, among other things, electricity needs associated with data centers.

A decade ago, a software story and a story about restarting a nuclear plant belonged to different worlds.

Today they can be part of the same technology strategy.

AI turns efficiency into a system-wide race

There is an important consequence.

Improving one component is not enough.

A faster GPU can expose a memory bottleneck.

Better GPUs and memory can run into advanced packaging limits.

Servers can be available while power is not.

Power can be contracted while grid interconnection is delayed.

Grid capacity can exist while transformers remain on long lead times.

The full machine moves at the speed of its constraints.

model
  ↓
compute
  ↓
memory
  ↓
network
  ↓
energy
  ↓
infrastructure

AI competition therefore becomes a race to optimize the whole system.

This may be the most important part of the shift

Thinking about AI as software makes it easy to assume that progress depends mostly on better algorithms.

Thinking about it as infrastructure reveals a different reality.

Progress depends on many industries scaling at the same time:

  • more efficient models;
  • faster chips;
  • better manufacturing nodes;
  • more advanced lithography;
  • more HBM;
  • more sophisticated packaging;
  • faster interconnects;
  • denser data centers;
  • more electricity generation;
  • more transmission;
  • better cooling systems.

It is an extraordinary industrial coordination problem.

The great paradox of artificial intelligence

AI feels increasingly abstract.

We talk to machines through natural language.

We do not see servers.

We do not see chips.

We do not see cables.

We do not see mines.

Yet the more magical the interface feels, the more physical the supporting infrastructure becomes.

Modern AI is simultaneously two things:

the most abstract interface in computing
+
one of the most material infrastructures in technology history

That is the paradox.

And it may be the best way to understand what is happening.

AI does not simply live in the cloud.

The cloud is made of factories, machines, silicon, memory, concrete, water, copper and electrons.

Every prompt touches —indirectly— that entire machine.

Sources and further reading