Muskeology
Frontier tech, minus the hype

AI & Compute

Why Advanced Chip Making Has So Few Suppliers

Leading-edge semiconductor manufacturing concentrated into a handful of firms because the capital costs, yield learning and equipment supply all reward scale enormously.

Detailed image of a circuit board featuring capacitors and intricate pathways, ideal for tech-related content.
Detailed image of a circuit board featuring capacitors and intricate pathways, ideal for tech-related content. · Photo via Pexels

Advanced logic chips are made by very few companies, and the number has shrunk with each generation. The concentration is a consequence of how the economics of the process work.

Each generation costs more than the last

Building a leading-edge fabrication plant requires an enormous upfront commitment, and the figure has risen steadily as feature sizes shrank. The investment is made before a single chip is sold.

Only a manufacturer confident of filling that capacity can justify the outlay. Firms with smaller order books fall behind, and once behind they cannot fund the next step.

The result is a ratchet. Every node transition removes participants, and none of them return, because re-entry would require matching a decade of accumulated spending at once.

Yield improves with volume, not with time

A new process initially produces a large fraction of defective chips. Engineers improve it by examining failures, adjusting parameters and repeating, which requires running enormous quantities of wafers.

A manufacturer with more customers gathers that data faster. Better yields lower cost per working chip, which attracts more customers, which produces more data again.

This feedback loop means the leader tends to extend its advantage rather than see it eroded, which is unusual among manufacturing industries.

The equipment supply chain is narrow too

The machines that pattern the smallest features are made by a very small number of specialist suppliers, and some critical tools have effectively one source worldwide.

Those tools are built from components drawn from their own deep supply chains, including optics and precision stages produced by firms that serve almost no other market.

Concentration at the equipment layer transmits directly to the chip layer. A delay in tool delivery moves the schedule of every fabrication plant expecting it.

Design and manufacturing separated decades ago

Most chip companies today do not own factories. They design a product and hand the layout to a contract manufacturer, sharing the plant with many other customers.

That split lowered the barrier to designing chips while raising it for making them. Hundreds of design firms now depend on a very small number of production sites.

It also means a single manufacturing constraint appears simultaneously across products that appear unrelated, from phones to accelerators to industrial controllers.

Geography makes the concentration a policy question

The most advanced capacity sits in a small number of locations, which turns an industrial structure into a strategic one for governments that depend on the output.

Efforts to build capacity elsewhere face the same economics that produced the concentration. A new plant needs sustained demand and experienced staff, neither of which appears quickly.

Subsidies can cover capital costs but not the accumulated process knowledge, which is why replication timelines are usually measured in many years rather than months.

Tobias Nkemelu
AI & Compute, Muskeology

Tobias builds and breaks machine learning systems for a living, which makes him a difficult audience for benchmark announcements.

More from Tobias →

Also by Tobias Nkemelu

AI & Compute

What to actually worry about with AI

A great deal of the risk discussion is about scenarios, and a shorter list of problems is already causing measurable harm.

Tobias Nkemelu··3 min read

AI & Compute

Where AI systems actually fail in production

Not in the model. In the data pipeline, the distribution shift, the feedback loop and the assumption that the world stays still.

Tobias Nkemelu··3 min read

AI & Compute

Evaluating an AI product claim

A short checklist for reading announcements, which mostly consists of asking what was measured and against what.

Tobias Nkemelu··3 min read

Robotics

Autonomous mobile robots outside the warehouse

Hospitals, hotels, factories and pavements — where wheeled autonomy has spread, and the specific reasons each environment is harder than a warehouse.

Lena Brandt··3 min read