Technology

TSMC's CoWoS Queue Is the Real AI Capacity Constraint

TSMC is tripling CoWoS packaging capacity, yet the queue is still booked a year out. Here is what actually sets AI chip supply.

Anthony Dami

Anthony Dami

Technology & Culture

Nov 21, 20269 min read
TSMC's CoWoS Queue Is the Real AI Capacity Constraint

A single Nvidia Blackwell package carries two GPU dies and eight HBM3E memory stacks on a silicon interposer roughly three times the area of a lithography reticle. The finished module is a few centimetres on a side and costs more than a family car. None of it is fabricated in the ordinary sense of that word. It is assembled, layer by layer, and the assembling has become the hardest step in the AI supply chain.

TSMC calls the process CoWoS, short for chip-on-wafer-on-substrate. It is what turns working silicon into something a data centre can bolt into a rack, and it is the step almost nobody modelled when the current buildout was planned. The wafer fabs get the headlines and the capex announcements. The packaging halls scattered across Taiwan set the actual limit.

The financial evidence sits in TSMC's October 15 results: third quarter revenue of NT$1.494 trillion, up 41.2% year on year, net income of NT$763.06 billion, up 68.7%, and fourth quarter guidance of $52.3 billion to $54.3 billion. On the July 16 earnings call, chairman and CEO C.C. Wei described the constraint without decoration: "packaging capacity is so tight that now it's limiting my customers' growth." That is the language of rationing, and it has not loosened since.

The Chip Isn't Finished When the Wafer Is Done

High bandwidth memory earns its name from distance. Every additional millimetre of copper between a GPU die and an HBM stack adds resistance and capacitance, and the signal integrity budget for a stack running at several terabytes per second is measured in fractions. Routing that traffic across an ordinary circuit board is not possible at the densities modern accelerators require, which is why the logic dies and the memory stacks sit side by side on an interposer carrying thousands of very short traces.

TSMC ships two variants. CoWoS-S uses a full silicon interposer, which is elegant and expensive and increasingly hard to build as packages outgrow a single reticle. CoWoS-L replaces most of that silicon with a redistribution layer plus tiny local silicon bridges placed only where the wiring density demands them, which keeps cost and warpage inside bounds. Blackwell and the Rubin generation that follows depend on the L variant. Beneath the module sits an organic substrate built on Ajinomoto Build-up Film, the laminate that carries the package to the board.

Assembly runs in a fixed order. Dies and memory stacks are bonded to the interposer while it is still part of a whole wafer, the populated wafer is thinned and diced, each module is mounted on its substrate, then underfilled, moulded and capped with a lid. Once that stack is bonded, a failed die or a cracked interconnect cannot be reworked. A package worth thousands of dollars is scrapped whole. That is why test cycles have stretched: reporting by Tom's Hardware put Blackwell test times at three to four times Hopper's, with Rubin possibly seven times longer.

The closest everyday parallel is a switchboard, not a highway. You can pour concrete for more lanes all day, but if the panels are few and only a handful of technicians can terminate the wiring, the traffic still waits, and the waiting is what costs money.

The Capacity Numbers, and What They Hide

None of the figures below come from TSMC. The company does not disclose CoWoS capacity and does not publish lead times, and it has not confirmed the price increases reported in Taiwanese trade press. What follows is the range of third-party estimates that analysts work from.

PeriodReported CoWoS capacityBasis
Late 2024~35,000 wafers/monthtrade estimates, before the expansion
End 2025~70,000 to 80,000 wafers/monthanalyst estimates
End 2026~120,000 to 130,000 wafers/monthcompany target as reported by analysts
2027~200,000 wafers/monthanalyst forecast
End 2028~260,000 wafers/monthcompany plan as reported
2026 non-TSMC overflow~80,000 wafers/monthASE and Amkor, simpler flows

A tripling of capacity in twenty-four months ought to end a shortage. It hasn't, and the reason is that three queues were stacked on top of each other. The first is demand, which grew faster than any forecast filed in 2024. The second is tooling: equipment lead times have stretched from three to six months to six to eight months as suppliers struggle to add production, according to reporting on the CoWoS expansion. The third is materials, where interposer wafers and ABF substrate are made by a small set of suppliers who did not plan for this either.

What that produces is a capacity curve that looks impressive on a slide and disappointing in a purchase order. Every wafer of new CoWoS output has already been spoken for before it exists, which means the announcement of a new monthly figure tells a buyer nothing about when they can actually receive product.

Why the Queue Outlives the Capacity

The metric that gates AI accelerator supply is the booking window, and almost nobody publishes it as tracked data. Trade reports put CoWoS-S and CoWoS-L lead times somewhere in the range of 52 to 78 weeks, which describes a customer ordering today and expecting delivery in the third quarter of 2027. TSMC will not confirm those numbers, but its own filings tell a compatible story: customer payments made to retain the company's capacity, running beyond twelve months, more than doubled in the first half of 2026. Nvidia's supply and capacity commitments rose from $119 billion to $279 billion in a single quarter, by one analyst tally.

Here is where the headline allocation numbers mislead. Morgan Stanley's widely cited estimate put Nvidia at roughly 595,000 CoWoS wafers for 2026, about 60% of global supply, with another 80,000 coming from non-TSMC packagers. Divide that reservation by twelve and Nvidia averages close to 49,600 wafers a month. Against an exit rate of 130,000 wafers a month, that is roughly 38%, not 60%. A share figure computed against a full-year average flatters the exit rate by more than twenty percentage points, which means the squeeze is seasonal and front-loaded, and every buyer who arrived late is competing for the marginal wafer rather than the average one.

The financing side has shifted too. As the AI trade has moved from equity into credit, capacity reservations have become contractual obligations that lenders price against delivery dates. That raises the cost of missing a slot by more than the price of the package.

The Price Signal That Clears Anyway

Commercial Times and TrendForce have both reported that TSMC is weighing increases of 10% to 20% on CoWoS services, with the company declining to confirm. Run the math forward. Third-party cost models put CoWoS-S near $750 per package for an H100-class design and CoWoS-L at $1,000 to $1,100 for a B200-class module. A 20% increase on the L variant is roughly $200 to $220 per package.

That number sounds large and behaves small. On an accelerator whose memory and packaging together account for 60% to 70% of bill-of-materials cost, and whose finished price runs into tens of thousands of dollars, $200 does not change a purchasing decision. Delay does. This is why a price increase of that size can clear without resistance and why it lands almost entirely on TSMC's margin line rather than on the customer's budget conversation.

Counterpoint Research measured TSMC's pure-foundry share at 73% in 2026, up from 68% in the first quarter of 2025, with Samsung flat at 7%. Packaging is where that share gets converted into pricing power, because a customer who cannot get CoWoS cannot substitute a different foundry's wafer and expect the same product at the other end.

Second Sources Arrive Late, and Smaller Than the Headlines Suggest

The alternatives are real and they are slow. Amkor's CoWoS-class line in Peoria, Arizona is slated to start in 2028, which makes it a 2029 story in volume terms. TSMC's separate supply arrangement with GlobalFoundries covers the first US-based silicon interposer production, a meaningful hedge on materials but not on assembly. Intel's EMIB has become a genuine second source with external customers including AWS shipping silicon on it, and it is the most credible challenger to appear so far. ASE and Amkor are absorbing the simpler overflow flows, roughly 80,000 wafers a month in 2026, mostly for parts like Vera CPUs and automotive devices rather than flagship GPUs.

That division of labour matters. The overflow goes to the packages that are easier to build and easier to test, while the largest multi-die GPU modules stay inside TSMC. Outsourcing the low-yield-risk work protects the customer relationship without handing a competitor the process knowledge that makes CoWoS-L work.

The concentration risk is not subtle. Every leading-edge CoWoS line sits in Taiwan, on an island with a fixed power budget and a live debate about data centre electricity. The IEEE's heterogeneous integration roadmap put chiplet-based solutions at $100 billion to $110 billion in annual revenue in 2026 while flagging exactly this, that logic from Taiwan, memory from Korea, substrates from Japan and assembly spread across Asia creates exposure no single company can hedge alone.

What the Constraint Actually Tells You

Watch the wrong indicator and the shortage looks permanent. Watch the right ones and the picture sharpens. Prepayment balances, equipment lead times and interposer supply move ahead of any capacity announcement, and none of them are in a press release.

The moat for the rest of this decade rests on yield and interconnect density rather than on floor space. Adding packaging halls is a construction project any large manufacturer with capital can attempt. Getting 95% of the good ones out the door, on modules that now take seven times longer to test than their predecessors, is the part that compounds. That asymmetry explains why the reported price increases are being pushed now, before the next capacity wave lands, rather than after.

What nobody has published is whether the first US interposer line changes the queue at all, or simply adds a second address to the same waiting list.

Strategic Takeaways

  • ✓TSMC's CoWoS packaging halls, not its wafer fabs, now set the ceiling on how many AI accelerators can ship in any given quarter.
  • ✓Reported capacity roughly triples to about 130,000 wafers a month by the end of 2026, but allocation is booked more than a year forward, so the queue shortens only for buyers who reserved early.
  • ✓Every credible second source, from Amkor's Arizona line to Intel's EMIB, arrives small and late, which keeps TSMC's pricing power intact well past 2027.
Anthony Dami
About the Author

Anthony Dami

Technology & Culture writer at Largern, focusing on AI, consumer hardware, film, internet culture, and the routines and tools behind everyday life.

Comments (0)

Join the Discussion

No comments yet

Be the first reader to share your perspective on this story.

Related Coverage