How the Foundry Business Works
A semiconductor foundry manufactures chips designed by other companies. The foundry doesn't design the chips — it provides the manufacturing process, the fabrication facilities, and the packaging services. The customer (the "fabless" company) provides the chip design in the form of a GDSII file, and the foundry turns that into physical silicon.
The business model is capital-intensive in the extreme. A single leading-edge fab costs $15-20 billion to build and equip. A complete process development program for a new node takes $2-4 billion and 3-5 years. Once operational, the fab needs to run at 85%+ utilization to generate acceptable returns. A foundry that can't fill its fab with customer orders bleeds money at an alarming rate.
Revenue comes from wafer sales, priced per wafer at a given technology node. A 300mm wafer at TSMC's N3 process costs roughly $16,000-20,000. At N5, it's about $13,000-16,000. At N7, maybe $9,000-12,000. Mature nodes (28nm and above) cost $2,000-4,000 per wafer. These prices are for high-volume production; NTO (new tape-out) customers or small-volume orders pay premiums.
TSMC: The Undisputed Leader
Taiwan Semiconductor Manufacturing Company controls about 57% of the global foundry market by revenue. For leading-edge nodes (7nm and below), their share exceeds 90%. That's not a dominant market position — it's practically a monopoly in advanced process technology.
TSMC's dominance rests on several interlocking advantages:
This connects to the ideas in Medical Device Semiconductors: Ultra-Low Power, Biocompatibi.
- Process technology leadership — TSMC has been first to volume production at every node since 7nm. Their N3E process is the most advanced in production, and N2 (with gate-all-around transistors) is on track for 2025.
- Manufacturing discipline — TSMC's defect density and yield ramp are consistently better than competitors. This reliability is what makes customers willing to pay their premium pricing.
- Customer trust — TSMC is a pure-play foundry. They don't design their own chips, so there's no conflict of interest. Apple, NVIDIA, AMD, and Qualcomm can share their most sensitive design data knowing TSMC won't use it to compete against them.
- Ecosystem effects — the EDA tools, IP libraries, and design flows are heavily optimized for TSMC processes. Switching foundries means re-qualifying the entire design methodology, not just porting the chip layout.
TSMC's gross margin sits at about 53-58%, among the highest in the semiconductor industry. They can command this premium because the alternative for their customers is worse performance, worse yield, or both.
Samsung Foundry: The Perpetual Number Two
Samsung's foundry business holds about 12% market share and has been trying to close the gap with TSMC for over a decade. They have aggressive technology roadmaps and the manufacturing infrastructure to execute — Samsung's total capex rivals TSMC's. But they consistently trail by 6-12 months on new node introductions and typically have worse yields at equivalent nodes.
Samsung's biggest structural challenge is the conflict of interest. Samsung designs and sells their own Exynos processors, Isocell image sensors, and DRAM/NAND products. A potential foundry customer like Qualcomm has to weigh whether Samsung might use insights from the manufacturing relationship to improve competing Samsung products. TSMC doesn't have this problem because they literally can't compete with their customers.
Samsung has scored some significant wins despite this — notably Qualcomm's Snapdragon 8 Gen 1 on Samsung 4nm and some Google Tensor chips. But the yield issues on Snapdragon 8 Gen 1 (which reportedly contributed to thermal throttling problems) pushed Qualcomm back to TSMC for the Snapdragon 8 Gen 2 and subsequent generations. That's the kind of quality stumble Samsung can't afford.
We covered a related topic in Semiconductor IP Licensing: ARM, Synopsys, and the IP Block .
Samsung's Differentiation Strategy
Samsung is betting on gate-all-around (GAA) transistors to leapfrog TSMC. Their 3nm GAA process was technically first to production in 2022, though initially only for crypto mining chips — not exactly flagship applications. The idea is that Samsung's early experience with GAA will give them an advantage as the technology matures. TSMC is transitioning to GAA (they call it nanosheet) with their N2 node, coming about 2-3 years after Samsung's GAA debut.
Whether the head start on GAA translates into competitive advantage depends entirely on yield. Being first with a technology that doesn't yield well enough for high-volume production isn't an advantage — it's expensive education.
Intel Foundry Services: The Comeback Attempt
Intel's foundry business is the most ambitious catch-up attempt in semiconductor history. After falling behind on process technology (their "10nm" problems that started around 2018), Intel under Pat Gelsinger launched Intel Foundry Services (IFS) with the goal of becoming a major third-party foundry by 2030.
Intel's plan centers on their Intel 18A process (roughly equivalent to competitors' 2nm). If 18A delivers competitive performance and yield, it would be the first time a US-based foundry offers modern manufacturing on American soil — a proposition with obvious appeal for customers concerned about Taiwan geopolitical risk.
See also: Silicon Wafer Supply Chain: From Sand to 300mm Wafers.
The challenges are enormous:
- Intel has never operated as a customer-facing foundry at scale. The culture of serving external customers with different design methodologies and IP needs is fundamentally different from designing and manufacturing your own products.
- The ecosystem gap — TSMC's PDK (process design kit) quality, IP library breadth, and EDA tool integration are the product of 35+ years of pure-play foundry operation. Intel's PDK maturity for external customers is years behind.
- Trust — Intel still designs its own processors. The conflict of interest issue that plagues Samsung applies doubly to Intel, which competes directly with potential foundry customers AMD and NVIDIA in multiple product categories.
- Financial pressure — Intel's foundry investment is happening while their core product business faces increased competition. The company reported its first quarterly loss in decades in Q3 2022.
Intel has announced design wins from the US Department of Defense and reportedly from Microsoft and Qualcomm for specific products. But volume production on Intel 18A for external customers isn't expected until 2026 at the earliest.
Specialty and Mature Node Foundries
Not all foundry business is about leading-edge nodes. The market for 28nm and above — often called "mature" or "trailing-edge" — is substantial and served by a broader set of foundries:
- GlobalFoundries — exited the leading-edge race in 2018 and focuses on differentiated mature nodes (22FDX, 12LP) for automotive, IoT, and RF applications
- UMC — Taiwanese foundry focused on 28nm and above, strong in display driver ICs and power management
- SMIC — China's largest foundry, officially limited to 7nm by US export controls (though they've demonstrated some 7nm-class production). Their bread and butter is 28-55nm.
- Tower Semiconductor — specialty foundry for analog, RF, and power management ICs
These foundries serve markets where the latest process node isn't necessary or even desirable. An automotive microcontroller or a power management IC works perfectly well on a 28nm or 40nm process, and the lower cost per wafer and proven reliability of mature nodes are advantages in those applications.
Pricing and the Capacity Game
Foundry pricing is a complex negotiation based on volume commitments, technology node, wafer start timing, and the customer's strategic importance. TSMC's largest customers — Apple, which accounts for roughly 25% of TSMC's revenue — get preferential pricing and priority capacity allocation. Smaller customers pay more and wait longer.
Long-term agreements (LTAs) became a major feature of the foundry business after the 2020-2022 chip shortage. Customers commit to specific wafer volumes years in advance, often with prepayments, in exchange for guaranteed capacity. TSMC secured over $30 billion in customer prepayments during 2021-2022. This shifts inventory risk from the foundry to the customer, which works until demand softens and customers are stuck paying for wafers they don't need — exactly what happened in the 2023 correction.