Fabless

A fabless company designs and sells semiconductor chips but owns no factory ("fab"). It outsources manufacturing to a contract foundry such as TSMC, letting it focus capital on design and IP. Nvidia, Broadcom, AMD, Qualcomm and Marvell are fabless, and the model underpins most AI and data-center silicon.

What "fabless" means

A **fabless** semiconductor company designs chips and sells them under its own brand, but does **not** own a fabrication plant (a "fab") to manufacture them. Instead, it hands its finished designs to a separate **foundry** that runs the multi-billion-dollar factories and prints the silicon. The fabless firm keeps the high-value work — circuit design, intellectual property (IP), software and customer relationships — and avoids the enormous fixed cost of building and continuously upgrading a fab. The opposite model is the **IDM** (integrated device manufacturer), such as Intel or Samsung, which both designs and manufactures in-house.

How the model works

The split rests on a simple economic logic: a leading-edge fab can cost well over $20 billion and must be rebuilt for every new process generation. Few companies can fund that, and even fewer can keep it full. A fabless company designs a chip using licensed IP blocks and electronic design automation (EDA) tools, then sends the design files to a foundry, which manufactures the wafers, after which separate firms handle packaging and testing. This is why the industry is often described in three layers — **fabless designers, pure-play foundries, and IDMs** — with specialist IP and EDA vendors (Arm, Synopsys, Cadence) feeding the designers.

Why TSMC made it possible

The fabless model only scaled because someone built a factory that would manufacture *anyone's* design without competing with them. **TSMC** (Taiwan Semiconductor Manufacturing Company), founded by Morris Chang in 1987, pioneered this "pure-play foundry" idea. By selling manufacturing as a neutral service, TSMC let design-only startups exist at all. Today TSMC controls roughly 60% of the global foundry market and the lion's share of the most advanced nodes — and Apple, Nvidia, AMD, Broadcom and Qualcomm all rely on it for their leading-edge chips.

Why it matters for AI and data centers

Almost every chip driving the AI boom is fabless-designed and foundry-built. **Nvidia** and **Broadcom** design AI accelerators and networking silicon but outsource fabrication to TSMC's leading-edge processes; the same is true for AMD. Demand is now concentrated at the bleeding edge: TSMC reports around 15 customers for its 2-nanometer (N2) node, with roughly ten of them in high-performance computing rather than mobile — a sign that AI and data-center silicon, not phones, is pulling the industry forward. The fabless model lets these companies pour capital into design and ship new architectures yearly instead of tying it up in factories.

Where it sits in the chip and photonics supply chain

Moving data between AI chips is now as hard as the compute itself, and the fabless model has spread into **optics**. As copper links hit bandwidth and power limits, the industry is shifting to **silicon photonics** and **co-packaged optics (CPO)** — putting the optical engine in the same package as the switch or accelerator. Broadcom and Marvell are central here. Marvell's February 2026 acquisition of **Celestial AI** and its "Photonic Fabric" added optical scale-up interconnect to its portfolio, illustrating how fabless designers are extending from logic chips into the photonic interconnect that stitches AI clusters together.

Key players and what's changing now

The largest fabless names include **Nvidia (NVDA)**, **Broadcom (AVGO)**, **AMD**, **Qualcomm**, **MediaTek** and **Marvell (MRVL)**. Two shifts stand out. First, **custom silicon (ASICs)**: hyperscalers increasingly co-design bespoke AI chips with Broadcom and Marvell rather than buying only off-the-shelf parts — Marvell's data-center revenue reached about 75% of its business in fiscal 2025. Second, **foundry concentration**: as nodes advance to 2nm and beyond, fewer foundries can compete, deepening fabless reliance on TSMC and raising geopolitical and supply-chain stakes.

Frequently asked

What is the difference between fabless and a foundry?

A fabless company designs chips but owns no factory; a foundry (like TSMC) owns the factories and manufactures chips on contract for others. They are complementary halves of the same supply chain — designers send blueprints, foundries print the silicon.

Is Nvidia a fabless company?

Yes. Nvidia designs its GPUs and AI accelerators but outsources manufacturing, primarily to TSMC. Broadcom, AMD, Qualcomm, MediaTek and Marvell are also fabless, while Intel and Samsung are IDMs that manufacture in-house.

Why don't fabless companies build their own factories?

A leading-edge fab can cost well over $20 billion and must be rebuilt for every new process generation. Outsourcing to a shared foundry lets a designer avoid that fixed cost and concentrate capital on design, IP and software.

What is the opposite of fabless?

An integrated device manufacturer (IDM), which both designs and manufactures its own chips. Intel and Samsung are the best-known examples, though both now also use external foundries for some products.

How does fabless relate to AI and data centers?

Most AI accelerators, networking chips and increasingly the optical interconnects in data centers are designed by fabless companies and built by foundries. The model lets firms like Nvidia, Broadcom and Marvell iterate on designs quickly while TSMC supplies the advanced manufacturing.

Related companies