Leading-Edge Foundry and CoWoS: The Capacity Gate for AI Chips

This link in one sentence

An analogy: an AI chip is like a fully fitted-out luxury apartment. NVIDIA is the architect who draws the blueprints (how to design the chip); but the ones who actually turn the blueprints into a building — and knock several units together into one sprawling open-plan home — are the leading-edge foundry and the advanced packaging house. This link is the AI chip's "construction and assembly shop." Many people assume that once a chip is designed, product exists — in fact, building and assembling it is the real bottleneck.

Getting a few terms straight

• GPU: a chip built for massively parallel computation, the main compute workhorse for AI training and inference. NVIDIA also designs its own companion CPU (a general-purpose processor that handles scheduling-type tasks) and DPU (a chip dedicated to networking, storage, and security that takes those chores off the GPU).
• Leading-edge process / leading-edge node: making chips with the newest, finest-featured process generation; the industry uses designations like N4, N3, N2 — the smaller the number, the more advanced and the harder to manufacture. Today virtually only TSMC can do this step well at scale.
• Packaging: after a chip "tapes out" (the first trial production run at the fab once design is complete), the main compute die, HBM (a high-speed memory that sits right next to the GPU — covered in detail in the next unit), and several other chips still have to be assembled at high density, wired together, and fitted with cooling to form a deliverable module — that step is advanced packaging, and it is far more than "putting on a case."
• CoWoS: TSMC's single most critical advanced packaging scheme. Simply put, the main GPU die and the HBM are "mounted" side by side on a silicon board packed with ultra-dense wiring (this board is the silicon interposer — essentially the expressway between chips), which then goes onto a high-end substrate (the ABF substrate). Because the chips are laid out flat and then stacked one more layer, the industry calls this "2.5D packaging." TSMC bundles the CoWoS, SoIC, and InFO packaging technologies into one family it calls 3DFabric.

Why this link matters from an investor's standpoint

One takeaway is enough: whether high-end AI accelerators can be built at all and delivered on time is often bottlenecked not by "design," but by this link's leading-edge process capacity + advanced packaging capacity. And it isn't decided by any single machine — it takes wafer fabrication, packaging, substrate, memory, and materials suppliers working as one coordinated set. This is a "systems integration" type of bottleneck that no one can break through at a single point. When reading a value chain, the key is not memorizing terms but seeing who delivers to whom, who gates whom, and who can charge whom.

Why the money stays here

This link earns high profit on the strength of three things others can't learn quickly. First, yield and capacity at leading-edge nodes are scarce in themselves (yield = the share of good chips; the yield ramp is slow). Second, the process experience of packaging a very large GPU together with HBM — ruining one piece means a huge loss. Third, customers will pay extra to get product sooner (the industry calls this time-to-volume, the speed from design to high-volume delivery). So leading-edge foundry doesn't earn a simple "foundry fee" — it earns the return that heavy investment, high yield thresholds, and strong customer lock-in buy together. Once CoWoS packaging capacity runs short, it flips from "backstage support" to "the chokepoint at the front of the stage," and profit concentrates here accordingly.

Three questions to size up its bargaining power

To judge whether any link can keep making money, ask the same three questions, applied here:

• Can customers do without it? — No: high-end AI chips today have virtually no route other than TSMC's leading-edge process + CoWoS.
• How hard and painful would switching away be? — Very hard: other fabs can't catch up on yield and capacity in the short term.
• Does expanding capacity take time, experience, and coordination up and down the chain? — Yes: packaging, silicon interposer, substrate, and test all have to expand in step; it can't be rushed.

If all three hold, the profit is more likely "structural" (durable); if orders are merely bunched together for now, it's more likely "cyclical" (it will fade).

Don't mistake "bigger revenue" for "better profit"

An AI volume ramp lifts revenue at almost every link in the chain, but rising revenue doesn't equal high-quality earnings. When judging, distinguish: does this link collect "platform rent" on irreplaceability, or does it merely haul high-value parts for others and earn a hard-work processing fee? Leading-edge foundry / CoWoS belongs to the former.

Which variables to watch

• The pace of process progression: how fast the move from N4 / N3 toward N2 / A16 (all TSMC process designations; the later, the more advanced) is going.
• Yield on large chips: the bigger the bare chip (the industry calls it a die), the harder it is to make and the more gets scrapped; yield directly determines cost.
• Packaging capacity: how many units CoWoS can process per month and how fast it expands; whether the supporting silicon interposer sizes, ABF substrates, high-end substrates, and test capacity keep up.
• Geographic distribution: leading-edge manufacturing and packaging are currently highly concentrated in Taiwan; how is overseas capacity progressing.
• Whether customers lock in capacity in advance: willingness to book capacity long term usually signals that demand is real.

The cycle side is straightforward: when AI orders exceed capacity, this link's profits are strong; once cloud providers cut budgets (capex deceleration) or packaging capacity comes online in bulk, the scarcity premium recedes.

Sort variables into "structural" and "cyclical"

This is a habit to keep for good: structural variables determine long-term bargaining power (process barriers, ecosystem lock-in, customer qualification, difficulty of substitution); cyclical variables determine current conditions (inventory, capacity expansion, order cadence, capex, prices). Mixing the two together is the easiest way to make the classic mistake of reading a "cyclical peak" as a "long-term moat."

Main risks

The biggest risk is a supply-demand mismatch: expand too slowly and NVIDIA's shipments get stuck; expand too fast and "shortage" may flip into "glut." Next is geopolitics and single-point concentration — both manufacturing and packaging are staked on Taiwan. Then there is technology substitution: new higher-integration packaging, different interposer approaches, or other foundries catching up would all shift where the profit sits.

How to judge this link

When judging this link, don't just ask "is CoWoS in shortage" — push one layer deeper: is this tightness genuinely underpinned by yield, process experience, and upstream-downstream coordination, or are orders simply crowding into the same window for a while? If customers can't switch away, the capacity ramp is slow, and yield thresholds are high, the profit leans structural; if the expansion playbook can be copied, demand is propped up by a single customer, and the queue clears quickly, the profit leans cyclical.

Boil the judgment down to one sentence

Does this link sit on a critical bottleneck of NVIDIA's AI factory, and is that bottleneck hard to remove in the short term through expansion, substitution, or standardization? If the answer leans "yes," it has a chance to earn durable, structural profit; if it leans "no," then however good the orders look right now, it is more like a cyclical stock amplified by demand.

What should change your mind

A good judgment starts by asking "what evidence would overturn it." For this link, if you later see any of the following — alternative packaging schemes maturing, customers' multi-supplier strategies working, expansion clearly outpacing demand, or downstream GPUs going underutilized — you should reassess whether it still holds this much bargaining power.