A chip design can be copied into a manufacturing database in an instant. Making it work reliably across thousands of wafers is a different undertaking. Inside a foundry, tiny variations in materials, equipment and process control determine whether an expensive design becomes a profitable product. The customer is buying repeatability at extraordinary precision, delivered in sufficient volume and on a schedule that its own business can trust.
The investment question
The foundry investment question is whether technical leadership produces sustained returns on a continually expanding capital base. Strong demand helps, but the business must translate factory spending into qualified capacity, high utilisation and acceptable yields. A new process can attract premium prices while depressing near-term margins as equipment depreciation begins ahead of a mature production ramp.
Our view is that the most durable advantage combines manufacturing execution with a trusted customer ecosystem. A process roadmap alone is insufficient. Designers need usable tools, proven intellectual property, packaging options and confidence that manufacturing problems will be resolved without derailing their product launches. Those relationships create a barrier that is harder to reproduce than the purchase of individual machines.
How a foundry works
A foundry manufactures semiconductor designs for customers. Its process design kit tells engineers how a particular technology can be used, including device models and physical rules. Design libraries and verification flows connect those rules to EDA software. Once a design enters fabrication, wafers pass through repeated stages of deposition, lithography, etching, cleaning, implantation and measurement before electrical testing.
Process names such as 3nm or 2nm are technology-generation labels, not a universal measurement of every transistor feature. Comparisons must consider density, performance, power, design rules and the product actually being manufactured. TSMC’s N2 research description explains its move to nanosheet transistors. The significance is improved control and design capability, not simply a smaller number in a presentation.
Yield is the proportion of output meeting requirements, but there are multiple yields along the route: wafer processing, functional die, packaging and final system qualification. A large accelerator can be more economically sensitive to defects than a small chip. The useful question is how many qualified products a process delivers at a given cost and performance level.
Market structure and competitive advantage
| Foundry position | Illustrative participants | Primary competitive issue |
|---|---|---|
| Advanced logic manufacturing | TSMC, Samsung Foundry, Intel Foundry | Process execution and customer qualification |
| Specialised and established processes | GlobalFoundries, UMC and other specialists | Application expertise and asset utilisation |
| Regional manufacturing ecosystems | Suppliers serving local customer bases | Qualification, supply resilience and market access |
TSMC’s dedicated foundry model aligns its business around manufacturing for customers. Samsung combines foundry operations with memory and product businesses. Intel’s foundry segment includes substantial activity supporting Intel products, making the distinction between internal and external demand essential. Segment revenue is not automatically a measure of independent customers choosing a foundry in the open market.
Specialised foundries pursue a different equation. Radio-frequency performance, embedded features, analogue behaviour, reliability and long qualification cycles can matter more than minimum transistor dimensions. GlobalFoundries’ second-quarter 2026 release highlights silicon photonics and silicon germanium within optical networking. This illustrates how AI can create opportunities outside the leading-edge logic race.
Economics: yield, utilisation and depreciation
Foundry economics begin with qualified wafer volume, product mix and realised prices. Costs include depreciation, labour, electricity, materials, maintenance and the research required to sustain the technology roadmap. Utilisation matters because many factory costs do not fall proportionally when wafer starts decline. A modest revenue reduction can therefore produce a much larger profit change.
Yield and utilisation must be distinguished. A factory can run close to capacity while generating disappointing saleable output because a new process is still improving. Conversely, a well-qualified process can suffer poor returns if customer demand leaves equipment idle. Capacity announcements address neither issue by themselves.
Consider an illustrative fab with annual fixed costs of 60 and variable costs equal to 30% of revenue. At revenue of 100, operating profit is 10; at revenue of 120, it is 24, assuming unchanged fixed costs and mix. The example is deliberately simplified. It shows the operating leverage that makes a successful ramp powerful and an underutilised expansion painful.
Free cash flow can lag earnings during expansion because factories consume cash before their costs appear fully through depreciation. Customer prepayments or long-term commitments can improve financing and visibility, but contractual protections differ. Assess how much risk the foundry retains if a customer delays a product or reduces volume.
AI and hyperscalers: indirect customers, direct influence
Hyperscalers influence foundry demand even when they are not the immediate invoiced customer. A cloud operator’s accelerator roadmap determines requirements passed through merchant designers or custom silicon partners. Large programmes must coordinate logic wafers, high-bandwidth memory and packaging capacity. The limiting component can change as each part of the chain expands.
Custom silicon broadens the set of designs entering production, but does not necessarily diversify the foundry’s ultimate demand. Several nominal customers may depend on the same small group of cloud buyers. Investors should distinguish the number of direct customer relationships from the concentration of underlying capital expenditure decisions.
Packaging also changes the customer relationship. A foundry able to coordinate advanced logic and integration can simplify accountability for complex AI products. That strengthens its role in the customer’s roadmap, while increasing the capital and execution burden. The relevant moat is successful production across the combined system, rather than control of one process step in isolation.
Current market debates — September 2026
The current debate concerns the durability and financial quality of advanced-node expansion. TSMC’s 2025 annual report states that N2 entered high-volume manufacturing in the fourth quarter of 2025, while its second-quarter 2026 results provide the more recent demand and investment context. A completed production milestone should be distinguished from the utilisation and profitability achieved as the technology scales.
The constructive case is that AI and energy-efficient computing support several years of demand, allowing advanced processes to ramp quickly. The countercase is that customers reserve capacity aggressively, then encounter downstream constraints or changing product economics. Evidence of sustained wafer loading and repeat designs matters more than a single quarter of strong orders.
Competition must also be assessed through actual customer production. Intel’s June 2026 process update describes 18A production and subsequent development work. Technical milestones are relevant, but independent customer qualification, recurring external revenue and acceptable economics are separate tests. It is premature to equate an announced roadmap with a fully established alternative supply chain.
Structural debates: geography and integration
Geographic diversification presents a trade-off between resilience and manufacturing efficiency. Additional locations can reduce dependence on a single region and align production with customer priorities. They can also require duplicated supplier networks, workforce development and higher initial operating costs. Subsidies can improve project economics without guaranteeing competitive operating performance.
Another question is whether chiplets weaken leading-edge concentration. Moving input/output or analogue functions to other processes may broaden supplier participation. Yet the most demanding compute die and final package can remain concentrated. The distribution of wafer area can change more than the distribution of economic value.
The final structural issue is customer trust. Foundries need access to sensitive designs and long-range roadmaps. Customers evaluate technical capability alongside confidentiality, commercial neutrality and delivery discipline. This makes competitive displacement a multi-year process in which one successful product creates the opportunity for the next, rather than instantly replacing the incumbent ecosystem.
What to watch
Track utilisation by process generation, yield commentary, the timing of depreciation and capital expenditure, external customer production and advanced-packaging availability. Compare geographic expansion milestones with qualified output rather than construction progress alone. Separate management’s demand expectations from delivered customer products.
The strongest foundry franchise should convert the next technology transition into repeat customer commitments and attractive cash returns after the investment cycle. Manufacturing leadership has economic value only when it becomes dependable, saleable output.
Explore this sector
Semiconductors & chipmaking — sector overview
Related sectors: AI infrastructure & data centres · Data platforms & analytics
