The chain

The whole AI value chain, part by part.

80 parts across 5 layers — from the power that runs it to the apps on top. Open any part to see what it is and who’s exposed.

Browse by layer

The AI energy layer The AI chips layer The AI infrastructure layer The AI models layer The AI applications layer

Energy

14

The amount of electricity the local power grid can actually deliver to a data center.

Transformer availability

Whether the large electrical transformers needed to connect a site to the grid can actually be bought in time.

Using liquid instead of air to carry heat away from hot AI chips.

Power density per rack

How much power, and therefore heat, is packed into a single equipment rack.

Substation capacity

The capacity of the local substation that steps grid power down for a data center.

Generation co-location

Building a data center right next to a power plant so it can take electricity directly, without waiting on the public grid.

Power purchase agreement

A long-term contract that locks in where a data center's electricity comes from and what it will cost.

Power delivery chain

All the equipment that carries power from the grid to the servers without interruption.

The price paid for each unit of electricity used by the data center.

Water availability

The water that fabs and some data centers need to operate and cool.

Backup generation

Generators and batteries that keep a data center running during grid outages.

A score for how much of a data center’s power actually reaches the computers.

Grid battery storage

Big battery installations that store electricity and smooth out the grid.

The traditional way of cooling servers by blowing cold air over them.

Chips

24

The main chip that does the heavy math for training and running AI models.

Super-fast memory stacked right next to AI chips so they are not starved for data.

Packaging capacity

The ability to assemble chips and memory together into a finished AI processor.

Lithography equipment

The machines that print chip patterns onto silicon wafers. One Dutch company, ASML, is the only maker of the most advanced kind (EUV) — nobody else in the world can build it.

Foundry capacity

The factory space available to actually manufacture advanced chips.

Hybrid bonding / 3D stacking

A way of stacking chips by welding their copper wiring directly together, with no solder bumps in between — so stacked dies sit closer and talk faster.

CoWoS / advanced packaging

A specific advanced method for placing chips and memory on a shared base.

Deposition & etch equipment

The machines that build a chip layer by layer — laying thin films down (deposition) and carving them away (etch). Three companies dominate the market.

EDA & design IP

The software and reusable building blocks engineers use to design chips before anything is manufactured.

Silicon wafer supply

The polished silicon discs every chip is printed on.

Metrology & inspection

The measurement and inspection machines that find microscopic defects while chips are being made — the quality-control eyes of every fab.

Process chemicals & gases

The ultra-pure chemicals and gases fabs consume every day to print and clean chips.

A chip custom-built for one job, like a specific AI workload.

The high-end circuit board inside a chip package that connects the silicon to the outside world.

Photomask & reticle

The master 'stencil' that a chip's circuit pattern is printed from — every wafer is exposed through it, so one flawed mask ruins everything printed after it.

The chip inside network switches that routes data between servers.

Chiplet interconnect (die-to-die)

The standardized 'plug' that lets separate chip tiles (chiplets) talk to each other inside one package — like USB, but between chips on the same substrate.

Test & burn-in capacity

The machines and time needed to verify every chip actually works before it ships.

Tiny circuits that send data very fast over a single wire or fiber.

NAND & AI storage

The flash memory and hard drives that store the mountains of data AI trains on and produces.

The share of manufactured chips that come out good and usable.

Glass-core substrates

A new kind of package base made from glass instead of the plastic-like film used today — flatter and sturdier, so bigger, denser chip packages can be built.

The regular fast memory used broadly across computers and servers.

Host compute (server CPUs)

The "normal" processors that sit next to AI chips in every server, feeding them data and running everything the accelerator does not.

Infrastructure

24

Optical interconnect

Using light through fiber, instead of electricity through copper, to move data between machines.

Scale-out fabric

The network that connects many machines across a data center into a giant cluster.

Scale-up fabric

The very fast links that join many chips into one big tightly-coupled machine.

A whole data center designed and run as one big AI machine.

The plug-in modules that turn electrical signals into light and back again.

The next generation of optical modules, doubling speed to 1.6 terabits per second.

Current high-speed optical modules running at 800 gigabits per second.

Indium phosphide

A special crystal material used to make the lasers inside optical modules.

Silicon photonics

Building optical parts using the same silicon chip-making methods.

AI buildout risk

The chance that planned AI data centers are delayed or cannot be finished.

The smart network card at the edge of every AI server that plugs it into the fabric and handles networking, storage and security so the main chips can focus on compute.

Deployment lead time

How long it takes to go from deciding to build to actually running AI workloads.

Putting the optical parts right next to the switch chip to save power.

Data-center construction

The specialized builders, electricians and equipment that physically construct AI data centers.

Photonic components

The building-block optical parts — lasers, detectors, modulators — inside optics gear.

High-density racks

Racks built to hold far more powerful, hotter equipment than before.

Optical circuit switching

Switches that steer beams of light directly, so data centers can rewire their network on the fly without turning light back into electricity.

Subsea bandwidth

The undersea fiber-optic cables that carry almost all internet traffic between continents.

Cluster scheduler

Software that decides which jobs run on which chips so nothing sits idle.

Sovereign compute programs

Governments buying and building their own AI supercomputers and national AI clouds.

Spine-leaf network

A common way of wiring a data center so every machine can reach every other quickly.

A simpler, lower-power optics approach that removes some power-hungry chips.

Structured cabling plant

The physical bundle of fiber-optic cables, connectors and patch panels that actually wires every machine in a data center together.

An earlier generation of optical modules running at 400 gigabits per second.

Models

12

Inference serving

Actually running a trained AI model to answer users, fast and at low cost.

How much it costs to produce each chunk of AI output.

Teaching a model general skills by training it on huge amounts of data.

Post-training & RLHF

The step after basic training where a model is taught to follow instructions, be helpful, and behave safely using human feedback.

The software layer that lets AI use tools and take multi-step actions.

Reasoning models

AI models that "think" longer to solve harder problems step by step.

Synthetic training data

AI-generated data used to train AI, created because the supply of high-quality human-written text is running low.

Model distillation

Training a small, cheap model to copy a big, expensive one so it can do similar work at a fraction of the cost.

Multimodal models

Models that understand and produce text, images, audio and video together.

Guardrails & safety layer

The safety filters around an AI that block harmful outputs, resist jailbreaks and keep private data from leaking.

Evaluation harness

The tests used to measure how good and reliable an AI model is.

How much text a model can consider at once.

Applications

6

An AI that carries out a whole multi-step task for you, not just one answer.

Enterprise knowledge assistant

An AI that answers questions using a company’s own documents and data.

Autonomous workflow engine

A system that runs entire processes on its own using AI.

An AI assistant that helps you write, code or work inside an app.

Robotics control

AI that controls robots to act in the physical world.

AI built into the tools programmers use to write software.