What Is a Semiconductor Foundry? Fabless, IDM and OSAT

In the news: TSMC and Rapidus

"TSMC's fab in Kumamoto, Japan, has started volume production." "Rapidus has made 2 nm prototypes in Hokkaido, Japan." Headlines like these often use the word foundry. Both companies make chips for other companies.

Your phone is a good example. Apple designs the chips in the iPhone itself. But Apple has no factory that makes chips. In an August 2025 announcement, Apple said it is the "largest customer" of TSMC's factory in Arizona.

In other words, the company that draws a chip's blueprint and the company that builds it are different. This article looks at how that happened, and at what the building side actually sells.

Three jobs that turn a design into a chip

Making a chip involves three broad jobs. The first is design. Engineers decide what the circuit should do and create its blueprint on a computer.

The second is the front end. Here the circuit from the blueprint is built into a round slice of silicon called a wafer. Patterns are printed with light, films are added, and material is etched away, again and again, dozens of times. This stage needs the largest and most costly equipment.

The third is the back end. The wafer is cut into individual chips. Each chip goes into a protective case with terminals that connect it to the outside, called a package, and is then tested. "How Are Chips Made?" walks through the steps inside the factory.

The four types of semiconductor company

In the industry, a chip factory is called a "fab." The word is short for fabrication, which means manufacturing. Knowing this makes the names of the company types easier to follow.

A company that only designs and sells chips, with no factory of its own, is called fabless. The name simply means "without a fab." NVIDIA and Qualcomm are typical examples, and they sell chips under their own names.

On the other side, a company that takes other companies' blueprints and does the front-end manufacturing is a foundry. The English word originally meant a workshop where metal is melted and poured into molds. A foundry does not sell its own products. It sells the work of making them.

A company that designs, manufactures and sells all by itself is called an IDM, short for integrated device manufacturer. "Integrated" here means one company covers the whole chain, from design through manufacturing to sales. Intel and memory maker Samsung Electronics are typical examples.

A company that takes on back-end assembly and testing is called an OSAT. A common split is that the foundry makes the wafers and the OSAT turns them into finished chips.

A company does not always fit just one type, though. Samsung Electronics is an IDM that makes its own memory. At the same time, it runs a foundry business that makes chips for other companies.

Why design and manufacturing split

In the past, most chip companies designed their own chips and made them in their own factories. In other words, the IDM was the norm. A big reason the split spread is that factories became extremely expensive.

A factory pays off best when it runs nonstop after it is built. The share of a factory's capacity that is actually in use is called the utilization rate. If one company's products are the only thing in the factory, utilization drops whenever sales fall. Costly equipment then sits idle.

A foundry fills its factory with orders from many companies. If one company orders less, orders from others can more easily fill the gap. This way, the huge cost of equipment is shared among many customers.

The pioneer of this model is TSMC, founded in Taiwan in 1987. The company says it "pioneered the business model of manufacturing customers' designs only." It does not design or sell chips under its own brand, so it does not compete with its customers. Customers can hand over their blueprints with confidence.

Once there was a partner to take on manufacturing, even companies without factories could make leading-edge chips. This is how fabless companies, which compete on design, grew large.

What does a foundry sell?

What a foundry sells to its customers falls into three broad groups. To keep things simple, this section focuses on front-end manufacturing.

1. Manufacturing technology (process)

Which materials and steps are used, and how fine the circuits can be. This full recipe is called a "process." Besides the newest generation, there are processes for each use: low power, high voltage, sensors and more. TSMC reports that in 2025 it used 305 technologies to make 12,682 products for 534 customers.

2. Production capacity

Being able to make a set amount by a set date is also a key product. Capacity is measured by how many wafers a fab can process per month or per year. Slots for popular generations fill up, so customers try to book them early.

3. Design tools and support

A design company must draw its blueprint to fit the foundry's process. It needs rules, such as the minimum width of a line and the minimum gap between lines. It also needs data that lets a computer simulate how the transistors behave. The foundry hands all of this over as one package, called a PDK (process design kit).

A blueprint that follows the PDK rules can be expected to be buildable in that fab and to work as predicted. Foundries also partner with design-software companies and with companies that sell ready-made designs for common circuits. Since 2008, TSMC has run this kind of partnership program under the name "OIP."

The flip side is that a blueprint is tied to one foundry's process. Moving to another foundry usually means reworking the design. This is one reason customers cannot easily switch.

Why are leading-edge fabs so expensive?

One reason is that the tools that print the finest patterns are very expensive, and a fab needs many of them. The main example is the EUV lithography tool, which uses light with a wavelength of 13.5 nm. A fab also needs rooms with almost no dust, and many tools to add films, etch them and measure them.

Let's look at some figures. In 2025, TSMC spent US$40.9 billion on factories and equipment, known as capital expenditure. That is up again from US$29.8 billion in 2024, according to the company's earnings call.

In March 2025, TSMC announced a plan to expand its investment in the United States to a total of US$165 billion. The plan includes six fabs, two advanced packaging plants and a research and development center. Note that this is a planned amount, not money already spent.

Only a few companies can earn back costs on this scale. In 2018, GlobalFoundries in the United States announced that it was halting development of its 7 nm generation, then the leading edge, indefinitely. It judged that too few customers needed the leading edge to justify the investment (IEEE Spectrum, 2018). Since then, it has focused on specialized processes, such as those for communications.

The main foundry companies

Here are the main companies and where each one stands. Numbers like market share change quickly, so the table lists only features that change slowly.

TSMC has also set up a subsidiary in Japan, called JASM, in Kumamoto Prefecture. Its first fab started volume production at the end of 2024, according to TSMC's annual report. Sony Semiconductor Solutions, DENSO and Toyota have also invested in it.

The first fab makes chips in generations older than the leading edge (12 to 28 nm). Many chips for cars, industrial equipment and home appliances are made in older generations, which are proven and cheap to produce. The race for the leading edge is not the only work a foundry does. For the second fab, now under construction, a plan to make the newer 3 nm generation was presented in 2026.

How to read foundry news

Market share: who counted what, and when

News such as "TSMC's share tops 70%" often quotes figures from TrendForce, a market research firm in Taiwan. According to TrendForce, foundry revenue shares for April–June 2026 were as follows. TSMC had 72.5%, Samsung Electronics 5.9%, and SMIC 5.4%.

These figures are based on revenue for each three-month quarter. Each research firm decides differently which companies, and which revenue, to count. So you cannot compare them directly with another firm's figures. When you see a number in a headline, check where it came from and what period it covers.

"X nm" is a generation name

Numbers such as "2 nm volume production" are names given to process generations. They do not mean that some part of the chip measures 2 nm. Between companies, the same number does not mean the same thing. "What Does 2nm or 3nm Mean?" explains this in detail.

Plan, prototype or volume production?

In news about fabs, notice which stage the story is about. A "plan" or "investment pledge" means the money has not yet been spent. A "prototype" means a small batch made to check that things work. "Volume production" means making large numbers as real products. Even for the same fab, the meaning changes a lot with the stage.

Common misconceptions

Key numbers

Of these, customer counts, investment amounts and shares change almost every year. Check company announcements and research firms' reports for the latest values.

Learn more

• Compare the company types in a table: "IDM vs Fabless vs Foundry" sets out strengths, weaknesses and typical companies.
• Front end vs back end: see "Front-End vs Back-End Semiconductor Process" and "How Are Chips Made?"
• From idea to blueprint: "What Is an Integrated Circuit?" explains the design flow and tape-out, the step that hands design data to the manufacturer.
• Sharing the cost of prototypes: an MPW puts designs from several companies on one wafer.
• Companies and words in the news: see the glossary for foundry, Rapidus and the silicon cycle, the industry's repeating booms and slumps.
• Progress beyond shrinking: "Beyond Scaling" covers stacking chips and other approaches in one place.