Foundry
A foundry is the manufacturing layer of the semiconductor industry: it takes a customer's finished chip layout and turns it into working wafers through a controlled sequence of lithography, deposition, etch, implant, metrology, and yield-learning steps — without ever touching the chip's actual design.
The industry exists because design and manufacturing scaled apart. A fabless company can focus entirely on architecture, RTL, verification, software, and go-to-market, while a foundry absorbs the capital intensity of process technology and factory operations — a single leading-edge fab now costs $10–20 billion to build. That split made the modern chip startup possible. It also turned foundry capacity into one of the most strategically contested resources in technology: when demand outstrips capacity, access — not just price — becomes the constraint.
Who does what:
| Role | Owns | Examples |
|---|---|---|
| EDA & IP | Tools, verification, reusable silicon blocks | Synopsys, Cadence, Siemens EDA, Arm |
| Fabless designer | The chip's architecture and product | NVIDIA, AMD, Qualcomm, most startups |
| Foundry | Manufacturing wafers from a customer's layout | TSMC, Samsung Foundry, GlobalFoundries, UMC |
| OSAT | Packaging and testing finished die | ASE, Amkor, JCET |
It's a loop, not a one-way handoff. The diagram above shows the forward flow — IP into design, design into manufacturing, bare die into packaging — but real engagements run a feedback path too: wafer yield data, DRC/LVS violations, and process-window limits flow back from the foundry to the fabless design team, shaping the next design revision before it ever reaches silicon. A foundry that only manufactures without that feedback loop produces chips; one that closes the loop improves them generation over generation — which is exactly what keeps Moore's Law-era scaling alive even as physics gets harder.