AI is everywhere in the news, but the supply chain that actually makes the chips behind it is not well understood. It is one of the longest and most concentrated networks in the world today. Following a single AI chip from raw silicon to a working data center is one of the clearest ways to see how a real supply chain behaves under pressure.
Here is the road it takes, and why the hardest part of it keeps moving.
From design to data center
An AI chip passes through several different hands before it does any work.
It starts as a design. Companies like NVIDIA and AMD design their chips but do not make them. That job goes to a foundry, mostly TSMC in Taiwan, which turns the design into silicon wafers using some of the most advanced manufacturing there is.
A finished wafer is not a usable chip yet. AI chips need a lot of memory bandwidth, so they are paired with high-bandwidth memory, or HBM, made by a small group of suppliers led by SK hynix, with Micron and Samsung ramping up. The chip and the memory then have to be joined together in a step called advanced packaging (CoWoS), which stacks them onto one substrate. Without this step, a wafer is just silicon that is not a product yet.
Only then is the chip put onto boards, built into servers, and shipped to the data centers where AI runs.
The bottleneck keeps moving
What makes this supply chain a good example is that the tightest point keeps changing.
From 2021 to 2024, the problem was raw capacity. There were not enough wafers, and getting GPUs was the whole game.
By 2025 and into 2026, the tight point moved to advanced packaging. TSMC’s CoWoS capacity was reported sold out through 2026, with NVIDIA taking most of it and the rest shared among a few other companies. Making more wafers did not help if they could not be packaged.
As packaging capacity grew, the tight point moved again, this time to memory. HBM supply for 2026 has been described as sold out, and it is expected to stay tight. Memory, not the chip itself, now sets the limit on how many AI chips can actually be built.
Even the companies trying to get around this, the big cloud players designing their own chips like Google, Amazon and Microsoft, do not escape it. Their custom chips still compete for the same foundry and the same memory as everyone else.
Why the shape of the chain matters
Two things make this supply chain fragile.
- The first is concentration. A big share of the world’s most advanced chips goes through a small number of factories and suppliers, many in one region. When so much depends on so few points, any problem, whether political, logistical or technical, has an outsized effect.
- The second is how connected it is. A shortage in one step does not stay there. When memory makers move production toward HBM for AI, normal memory gets tighter for everyone else, and industries with no direct link to AI, like automotive and consumer electronics, feel it too.
That is what makes this supply chain worth studying. It is not just long. It is a system where the tight point keeps moving, where fixing one thing reveals a new limit somewhere else, and where a local problem spreads worldwide.
Seeing it, not just reading about it
A supply chain like this is hard to picture from a description. How capacity, allocation, cost and disruption interact only becomes clear when you can watch it run over time.
That is exactly why it makes a good simulation. When you build the network, run it, and then cause a disruption in one region, you see the effects move through the rest of the chain, and you have to make real decisions in response.
This is the idea behind one of the two case studies we’re building for the fall Supply Chain Competition 26 (see Supply Chain Cup 26 – Spring edition), open to students from universities around the world. They’ll step into the role of the supply chain manager behind an AI chip network and run it under real constraints and disruptions. Reading that the tight point moved from packaging to memory is one thing. Running the chain yourself and watching it happen is another.