Foundry

A semiconductor foundry is a factory that manufactures microchips designed by other companies. "Pure-play" foundries like TSMC build but never design competing products, so fabless firms (Nvidia, Apple, AMD) trust them with their designs. Foundries supply the AI chips, data-center silicon, and optical components behind today's compute boom.

What a foundry is, in plain terms

A **semiconductor foundry** is a factory that manufactures microchips that *other* companies design. The foundry doesn't usually sell chips under its own brand; it takes a customer's blueprint and turns it into physical silicon. This is the manufacturing half of the chip industry, split off from the design half. The key distinction is the **pure-play foundry** model, pioneered by Taiwan's TSMC in 1987. A pure-play foundry builds chips but never designs competing products of its own. That neutrality is the whole point: a "fabless" design company like Nvidia, Apple, AMD, or Qualcomm can hand over its most valuable secrets without fear the factory will copy them. Contrast this with an **IDM** (integrated device manufacturer) like Intel or Samsung, which both designs and builds its own chips.

How a foundry actually works

Building a leading-edge chip means stacking billions of transistors onto a fingernail-sized piece of silicon through hundreds of precise steps: depositing thin films, patterning them with light (lithography), etching away material, and doping it to control how current flows. A modern "fab" (fabrication plant) costs tens of billions of dollars and runs in cleaner-than-hospital cleanrooms. Process generations are named by "node" — historically a transistor size, now mostly a marketing label for a density/performance tier. In late 2025 TSMC began volume production of its **N2 (2nm-class)** node, its first to use **gate-all-around (GAA)** nanosheet transistors, where the gate wraps fully around the channel for better control and less leakage. TSMC quotes N2 at roughly 10-15% more performance at the same power, or 25-30% less power, versus the prior generation.

Why foundries matter for AI and data centers

Every AI accelerator — Nvidia's GPUs, custom chips from Google, Amazon, and Microsoft — is physically built by a foundry, overwhelmingly TSMC. The AI boom is, at the physical layer, a foundry boom: the pure-play foundry sector grew about 30% year-over-year in early 2026, driven largely by AI GPU and ASIC orders. It isn't only the logic chips. Modern AI training packs many chiplets together using **advanced packaging** (like TSMC's CoWoS), which foundries also provide and which has become a supply bottleneck. Increasingly, moving data *between* chips at data-center scale relies on **silicon photonics** and **co-packaged optics (CPO)** — optical components that foundries such as GlobalFoundries and Tower Semiconductor manufacture alongside traditional silicon.

Where foundries sit in the supply chain

The foundry is the central node in a long chain. *Upstream*, it depends on equipment makers (ASML's lithography machines, Applied Materials, Lam Research), on ultra-pure materials and gases, and on silicon wafers. *Downstream*, the bare wafers go to packaging and test (often the foundry's own advanced-packaging lines or specialist OSATs) before reaching the customer. Foundries also split by what they make. **Leading-edge logic** (the newest nodes for CPUs, GPUs, phones) is dominated by a handful of giants. **Specialty / mature-node foundries** make the analog, power, sensor, and mixed-signal chips that don't need the smallest transistors but are essential for cars, industry, and 5G — a different and very large market.

Who the key players are

**[TSMC](https://www.tsmc.com)** (NYSE: **TSM**) is the dominant pure-play leader, holding roughly 70%+ of the pure-play foundry market and effectively all leading-edge AI production. **Samsung Foundry** and **Intel Foundry** are the main challengers at the leading edge; both are racing to ship 2nm-class nodes (Samsung's GAA process and Intel's **18A**, which powers its 2026 "Panther Lake" chips), but as of early 2026 TSMC remained the only foundry shipping advanced GAA chips externally at scale. In specialty silicon, **GlobalFoundries** (**GFS**) is a major mature-node and silicon-photonics player; **Tower Semiconductor** (**TSEM**) leads in high-value analog, RF, silicon photonics and CPO; and Germany-based **X-FAB** (**XFAB**) specializes in analog/mixed-signal plus wide-bandgap power technologies (silicon carbide, gallium nitride) for automotive, industrial and medical uses.

What's changing now

Three shifts stand out. **Geography:** advanced capacity is spreading beyond Taiwan, with TSMC building N2/N3 fabs in Arizona, Intel in Ohio, and Samsung in Texas — driven by AI demand and supply-chain/geopolitical concerns. **Architecture:** the industry has crossed from FinFET to GAA transistors, with backside power delivery (TSMC's A16, Intel's PowerVia) arriving next. **Optics:** as AI clusters strain on electrical interconnects, foundries are rushing into silicon photonics and co-packaged optics. In late 2025 GlobalFoundries acquired photonics specialist AMF and launched a CPO module platform, while Tower opened a dedicated CPO foundry — signaling that the next foundry battleground is moving light, not just electrons, around the data center.

Frequently asked

What is the difference between a foundry and a fabless company?

A fabless company (like Nvidia, AMD, or Apple) designs chips but owns no factories. A foundry owns the factories and manufactures those designs. The two are partners: the fabless firm provides the blueprint, the foundry turns it into physical silicon. An IDM, by contrast, does both itself.

Why is TSMC so dominant?

TSMC invented the pure-play model and never competes with its customers in chip design, which built deep trust with fabless leaders. Decades of reinvestment gave it the best leading-edge yields, advanced packaging, and scale — so it holds roughly 70%+ of the pure-play market and builds nearly all cutting-edge AI chips.

What does 'node' (like 2nm or 5nm) actually mean?

Historically a node referred to a transistor feature size. Today the numbers (5nm, 3nm, 2nm) are mostly marketing labels for a performance-and-density tier rather than a literal measurement. A smaller-numbered node generally means more transistors packed in, faster speed, and lower power.

Do foundries only make the newest, smallest chips?

No. Leading-edge logic gets the headlines, but most chips use mature nodes. Specialty foundries like X-FAB, Tower, and parts of GlobalFoundries make analog, power, sensor, RF, and silicon-carbide chips that are essential for cars, industry, medical devices, and communications.

How do foundries connect to AI besides making the GPUs?

Beyond fabricating AI logic chips, foundries provide advanced packaging that bonds multiple chiplets and high-bandwidth memory together, and they increasingly make silicon-photonics and co-packaged-optics components that move data optically between chips in AI data centers.

Related companies