The agent supply chain
Every agent sits on top of an industrial supply chain. From raw materials and optics to fabs, networking, cloud interfaces, models and tools, the systems behind an agent extend far beyond the application in front of you. This network maps the dependencies that make agentic software possible.
Every layer between raw silicon and the agent you deploy
AI cost is not an abstract number. It is the outcome of a stack. The graph follows the chain from materials, lasers and fab equipment through EUV, foundries, accelerators, cloud infrastructure, foundation models and the tooling that reaches an end user. Each layer adds capability, cost and dependency.
Show more
Tracing those connections makes it easier to understand how operating costs accumulate, why a seemingly simple agent relies on extraordinary underlying sophistication, and where those costs eventually surface. It also makes concentration visible. Where a critical part of the stack is controlled by a small number of suppliers, dependency becomes a strategic risk. Use the network to ask where power sits, and what it means to build on a single provider.
What is the AI supply chain?
The chain of physical dependencies behind every AI agent. It runs from raw materials and optics, through chipmaking equipment, foundries, memory and accelerators, to networking, power and datacentres, and then up to clouds, models and the agent you use. Each layer depends on the one below it.
Why does EUV lithography matter for AI?
Because leading-edge AI chips cannot be made without it. Extreme-ultraviolet machines print the smallest features on a wafer, and ASML is the only company that builds them. That makes it the hardest link in the chain to route around.
Where are the biggest bottlenecks in the AI supply chain?
Three places. Advanced packaging and high-bandwidth memory, which limit how many accelerators can actually be assembled. EUV tooling at the base. Power and datacentre capacity at the top. A squeeze in any of them ripples through everything above.
Who makes the chips AI agents run on?
A short list of companies. Nvidia designs most of the accelerators used for AI, with AMD and the cloud providers' own silicon behind it. Almost all of it is manufactured by TSMC, and the high-bandwidth memory comes from SK hynix, Micron and Samsung. Chip design and chip manufacturing are separate businesses, which is why one company can dominate the first without owning a factory.
Why is there a shortage of AI chips?
Demand is only half of it. Capacity for advanced packaging and high-bandwidth memory is booked far ahead, and a new fab takes years and billions to bring online. Grid power for datacentres is now a queue of its own, so buying the chip is not the only wait.