Semiconductors — Buyers Fund the Factory: Glass Cloth Binds

Home / The Stack / Semiconductors

Semiconductors turn AI designs into chips that ship: design, foundry wafers, HBM memory stacks, advanced packaging, substrates, the materials inside them, optics, test, and the tools behind each step. This Stack layer tracks who gets paid in that chain, and which step is holding the cash.

Topic hubs: HBM · CoWoS · Company models: TSMC · SK hynix · Micron · ASML · Applied Materials · Lam Research · KLA

Status: Tightening · Updated: 2026-10-26

Investor read: On 22 October this page said the buyer had prepaid past the wafer and that substrates bind to 2028. This week the buyers went further and paid for the substrate factory itself. On 22 October Korea's Maeil Business Newspaper reported that Nvidia and Tesla will put about ₩3.7 trillion, roughly $2.8 billion, into Samsung Electro-Mechanics' new substrate lines at Sejong. Two days later Taiwan's Economic Daily News reported industry talk that Intel, Nvidia and AMD have paid more than NT$100 billion in deposits to Taiwan's substrate makers, which have announced a record NT$161.2 billion of new plants this year. Money for buildings and tools is no longer the problem, and the makers say so. At the Taipei board show Kinsus called its tool risk relatively controllable and named materials as the real risk: the low-expansion glass cloth, or T-glass, in the core of every large AI substrate, where it sees a 5% to 15% gap that widens in 2027. Thin cloth and laminate are expected to cost 10% to 15% more from November, and the shortage is spreading to memory substrates. The buyers can fund the factory. They cannot fund cloth that has not been woven. Buyers fund the factory. Glass cloth binds.

On this page

• Numbers that matter
• What this layer does / who gets paid
• Stuck step
• Role companies
• What limits it
• Cross-links
• Practical takeaways
• FAQ
• How we check
• What would change the view
• References

Numbers that matter

Read the table in three blocks: buyers paying for substrate plants that start in 2028; what those plants still have to buy, glass cloth and laminate first; and the steps around the substrate, each paced by something other than money.

What this layer does / who gets paid

An AI accelerator gets paid for in this order:

1. Tools — toolmakers sell lithography, etch, deposition and inspection machines for each new line (see who makes wafer fab equipment and what KLA does).
2. Wafers — the foundry makes the processor die and is paid per wafer (see what TSMC does).
3. Memory — memory makers stack DRAM into HBM and sell it on long-term contracts (see HBM explained and who makes HBM).
4. Interposer, stacking and packaging — the processor and its HBM are joined on an interposer or stacked (see what an interposer is, CoWoS explained, SoIC explained and who makes advanced packaging).
5. Glass cloth, laminate and film — woven glass cloth is soaked in resin and pressed with copper into the laminate core; build-up film goes on top. This is today's stuck step.
6. Substrate — the package is mounted on that layered board, the ABF substrate, which carries power and signals out to the server (see IC substrates and ABF explained).
7. Board, test and optics — the part is tested, mounted on a server or switch board, and the next networking parts add optical engines beside the chip (see silicon photonics explained).
8. Ship and bill — the chip designer sells the finished package into racks (see who makes AI chips and who makes custom AI chips).

This layer sits upstream of Hyperscalers and Neoclouds, which buy the finished systems, and of AI Data Centers, which house them. Chips, memory, packaging, substrates and their materials, optics, test and tools stay here.

Who gets paid first now: makers of the scarcest inputs — T-glass, thin cloth and laminate — who are sold out and expect higher prices from November; substrate makers, paid by the buyer before the plant is built; and board makers whose orders run well ahead of shipments. Who waits: funded substrate lines that reach mass production only in 2028; buyers who did not prepay; and bonder and tool makers whose machines ship only when the customer's line is ready. Who carries the risk: the buyer that prepaid, if the plant arrives late or its demand falls; and the substrate maker, if a funded line cannot get enough qualified cloth to run full.

Stuck step

On 18 October this page put the stuck step at scanner slots for 2027 and 2028. On 22 October it moved past the wafer, to substrates, with the buyer prepaying for capacity. This week showed what that money buys and what it cannot. It buys buildings and tools. It does not buy the glass cloth those lines run on.

1. The buyer funds the factory — PRIMARY (new 2026-10-26). On 22 October the Maeil Business Newspaper reported, citing industry sources, that Tesla will invest about ₩1.1 trillion and Nvidia about ₩2.6 trillion to build FC-BGA substrate lines at Samsung Electro-Mechanics' Sejong plant. Reuters carried it at $2.8 billion. Tesla gets five years of supply from 2029. The building is due by the end of 2027, with tools and trial runs through the first half of 2028 and mass production from the second half. Samsung Electro-Mechanics filed a ₩4.27 trillion Sejong budget in September. On our arithmetic the two sums equal about 87% of it, an upper bound because Nvidia's money also covers Busan. Neither buyer has published the amounts. See the Wire on the $2.8B Sejong investment.

2. Taiwan's makers get deposits and spend a record — SECOND PRINT. On 24 October Economic Daily News reported industry talk that Intel, Nvidia and AMD have paid more than NT$100 billion, about $5 billion, in deposits to reserve long-term capacity at Taiwan's high-end substrate makers. The split is an analyst estimate: about $3 billion for Intel and $1 billion each for Nvidia and AMD. The makers do not comment on single customers. Their spending is on the record: Unimicron, Kinsus and Nan Ya PCB have announced a record NT$161.2 billion of plants this year, and Unimicron has lifted 2026 capital spending to NT$53.7 billion. On our arithmetic the deposits equal at least about 62% of the plants announced, though one figure is an estimate and the two cover different years. Either way, the buyers carry much of the cost of the building. See IC substrates and ABF explained and the Catalyst on Nvidia's 2.5x substrate plan.

3. The substrate maker says the risk is the cloth, not the tools — THIRD PRINT. At the Taipei board show Kinsus set out a plan to double its ABF substrate output from 40 million units a month in 2026 to 80 million in 2029, with customers paying for the equipment; on our arithmetic about 26% more each year. It called the equipment risk relatively controllable, because its new capacity lands from 2028 to 2030. The real risk, it said, is materials. Its gap in T-glass, the low-expansion cloth that keeps a large substrate flat when the chip heats up, swings between 5% and 15% and should widen again in 2027 as customers launch new platforms. It is asking customers to move some designs to ordinary E-glass, proposing hybrid cores and qualifying new suppliers. Industry estimates put the 2027 T-glass gap at 20% to 30% even after Nittobo, the leading maker, adds capacity, in line with Unimicron's 20 October warning that some key materials could be about 20% short. Glass and ceramic cores are not expected in volume before 2029. A deposit can buy a plant. It cannot buy a loom that is already booked. See the Taiwan note on the glass cloth squeeze and the Analysis: if substrates stay short through 2028.

4. The thin cloth and the laminate reprice — FOURTH PRINT. On 22 October TechNews reported that the supply chain expects glass cloth and copper-clad laminate prices to rise about 10% to 15%, as early as November. The thin weaves known as 2116, 1080 and 1078 are tightest, because weavers moved looms to the special AI grades, and much of the new capacity due in 2027 is thick cloth. Relief is not expected before the second half of 2027, possibly 2028. These are expectations, not price lists. The same day TTM, a US board maker whose data center and networking sales rose 115%, said laminate and T-glass lead times run about 12 weeks, sometimes 24 or longer, and are going up, while its commercial orders ran at 2.5 times shipments. See the Catalyst on TTM's board orders.

5. The shortage spreads to memory substrates — FIFTH PRINT. On 19 October Samsung Electro-Mechanics disclosed a ₩360.5 billion supply deal with a global memory maker. Kiwoom Securities read it as next year's volume of BT substrates, the cheaper boards memory chips sit on, and called an annual contract like this unusual. BT makers already run in the mid-80% range of capacity. Kiwoom expects ABF capacity to more than double by 2028 while BT grows only in the mid-teens percent, and a BT shortage in 2027 and 2028. That is a broker view. Both kinds of substrate use glass cloth in the core. See what Micron does and the Analysis on SK hynix's back end.

6. Around the substrate, machines wait on lines — SIXTH PRINT. On 22 October Besi, which makes chip-bonding machines, reported third-quarter orders of €342.6 million, up 96% on a year earlier, with hybrid-bonding orders so far this year more than double all of 2025. It guided fourth-quarter revenue flat because deliveries are timed to customers' lines for the first half of 2027, not because orders moved out. On our arithmetic orders ran about 1.21 times revenue. Upstream, Intel's chief technology officer Pushkar Ranade said the world has tools for probably twice today's leading-edge wafer output, not ten times, with about 18 months to deliver. The 18 October tool queue stands. See the Wire on Besi's orders and High-NA EUV explained.

Demand is not the bind. Buyers such as Amazon say their AI capacity is booked well ahead. They are short of time on a step money cannot hurry, not of cash. Our read, labelled as ours: substrate makers with secured cloth, or designs qualified on alternatives, fill funded lines first. See the Wire on AWS sold out into 2028.

So the stuck step, named plainly: the buyers are now paying for the substrate factories as well as the capacity. Those factories run from 2028, and the makers say the risk is the glass cloth inside every board, not the tools. The thin cloth and laminate reprice from November, and the shortage is spreading to memory substrates. Buyers fund the factory. Glass cloth binds.

Role companies

Roles only — not a leaderboard, not a buy list.

Always verify listing, ADR, and tax treatment for your own jurisdiction. Names here map roles. They are not recommendations.

What limits it

1. Glass cloth — low-expansion T-glass is the tightest input; Kinsus sees a 5–15% gap widening in 2027, and industry estimates run to 20–30%.
2. Looms and furnaces — new cloth needs new furnaces and weaving lines; much new capacity is thick cloth, not the thin weaves that are short.
3. Plant timing — funded substrate lines reach mass production from 2028 (Sejong) and 2028–2030 (Kinsus).
4. Board size — AI substrates keep growing, so each package uses more cloth and laminate, and the same square metres cover fewer chips.
5. Qualification — E-glass, hybrid cores or a new cloth supplier each need redesign and customer tests.
6. Tools upstream — Intel's CTO says tool supply supports about 2× leading-edge wafer output, with about 18-month lead times.

Cross-links

Explained

• IC substrates and ABF explained
• What is an interposer?
• What is SoIC? Why stacking chips still leaves HBM beside the GPU
• What is a chiplet?
• CoWoS explained
• CoWoS vs EMIB
• Who makes advanced packaging?
• Foundry and advanced packaging explained
• What is silicon photonics?
• Optics and networking in AI clusters
• Who makes wafer fab equipment?
• What is High-NA EUV?
• HBM explained
• Who makes HBM?
• Who makes AI chips?
• Who makes custom AI chips?

Company models

• What Does KLA Do? The process-control model
• What Does Lam Research Do? The etch-and-deposition model
• What Does ASML Do?
• What Does Applied Materials Do? The process-chamber model
• What Does TSMC Do? (CoWoS)
• What Does SK hynix Do?
• What Does Micron Do? The deposit-locked DRAM model

Wire, Catalyst & Analysis

• Wire: Nvidia and Tesla put $2.8B into Samsung Electro-Mechanics substrates
• Wire: Besi books €343m of orders as HBM still uses bumps
• Wire: Micron plans to stack HBM on top of AI chips from 2028
• Wire: NVIDIA pays $1.1bn for US test as Amkor deals hit $3bn
• Wire: Amazon says AWS is sold out into 2028 despite $220B of capex
• Catalyst: TTM data center boards up 115% as orders run 2.5x sales
• Catalyst: JCET lifts 2026 capex plan 65% to RMB18.2 billion
• Catalyst: Nvidia books 2.5x substrates and 300% more interposer capacity
• Analysis: If substrates stay short through 2028, where does cash stick?
• Analysis: What does KLA's price assume about the inspection stuck step?
• Analysis: What does ASE's price assume about the packaging stuck step?
• Analysis: SK hynix and the HBM4 back-end stuck step
• Analysis: Where does the stuck step move

Markets

• Taiwan: 20% short, 30 months out — who gets paid while Taiwan's AI boards wait on glass cloth
• Taiwan: Wafers ramp — who gets paid while Taiwan's CoWoS still binds
• US: 194 days of stock, tools running late — Texas Instruments

Sibling Stack layers

• The Stack hub
• Neoclouds
• Hyperscalers
• AI Data Centers

Practical takeaways

1. Follow the money: last week buyers prepaid capacity, this week they fund the plant. The next step a buyer funds is the one it fears most.
2. For every funded substrate line, ask two dates: when it starts mass production, and where its glass cloth comes from.
3. Separate what is on the record from what is reported: Samsung Electro-Mechanics' Sejong budget is filed; the Nvidia and Tesla amounts and the Taiwan deposits are reports and estimates.
4. Watch the November price rises in glass cloth and laminate, and whether they reach the substrate makers' own prices.
5. Watch memory substrates. If more BT makers sign full-year deals, the shortage has spread beyond AI chips.
6. Watch SK hynix and Samsung Electro-Mechanics (28 Oct), Samsung and Intel (29 Oct) for what they say about substrates, materials and prepayments.

FAQ

Is this a buy list?

No. This is a map of roles and constraints. Not investment advice. Do your own research.

Why is status Tightening on 2026-10-26?

Because buyers are now paying for substrate plants that do not reach mass production until 2028, the substrate makers say the glass cloth those plants need is short and the gap widens in 2027, cloth and laminate prices are expected to rise from November, and the shortage is spreading to memory substrates. We would ease if cloth makers bring new low-expansion capacity forward, or if substrate makers report the gap closing.

How did this advance from the 2026-10-22 Stack frame?

The 22 October desk led with Prepaid past the wafer / Substrates bind to 2028: Nvidia paying substrate, packaging and test makers in advance for capacity. This update shows buyers now funding the substrate factories themselves, and the makers saying the limit has moved inside the substrate, to the glass cloth it is built on. The headline becomes Buyers fund the factory / Glass cloth binds.

What is T-glass?

A low-expansion glass-fibre cloth. It barely grows when it heats, which keeps a large chip substrate flat so the connections do not crack. The bigger the AI package, the more of it the substrate core needs. A few makers, led by Japan's Nittobo, produce the top grade.

Why can't the buyers just pay for more glass cloth?

They can pay, but cloth needs glass furnaces, yarn and weaving lines that take years to add, and much of the new capacity coming is thick standard cloth, not the special or thin grades that are short. Switching a design to a different cloth also needs new testing.

Are the Nvidia and Tesla amounts confirmed?

No. The amounts come from a Maeil Business Newspaper report citing industry sources, carried by Reuters. Nvidia named the partnership but gave no figure, and the companies have not published the sums. The ₩4.27 trillion Sejong budget was filed by Samsung Electro-Mechanics in September.

How we check

1. Company filings and releases first: Samsung Electro-Mechanics' filed budget, Besi's and TTM's results, and earnings-call remarks on the record.
2. Reported amounts from named publications, labelled as reported, with the publication's own sourcing stated (industry sources, analyst estimates).
3. Supply-chain price and shortage expectations labelled as expectations, not price lists.
4. Broker views, such as Kiwoom's BT forecast, labelled as broker views; ratings and price targets are not used.
5. For every funded line: what pays for it, when it reaches mass production, and what input it still needs.
6. Our own arithmetic, labelled as ours, never as a reported figure.

What would change the view

• Nittobo or other cloth makers bringing new low-expansion capacity forward from 2027, or qualifying new suppliers faster.
• Substrate makers reporting their T-glass gap closing, or designs moving to E-glass and hybrid cores without yield loss.
• November cloth and laminate price rises failing to stick.
• Buyers cancelling or deferring substrate deposits, which would point to weaker demand rather than short supply.
• More full-year memory substrate deals, which would tighten the view further.

Not investment advice. Do your own research. The 87%, 62%, 26% a year and 1.21× figures are our arithmetic on the reported figures. Deposit splits are analyst estimates as reported by Economic Daily News; BT capacity growth is a Kiwoom Securities forecast; price rises are supply-chain expectations as reported by TechNews.
Last updated: 2026-10-26 · Second Order by TKN

References

1. Maeil Business Newspaper, "Samsung Electro-Mechanics to receive ₩3.7 trillion from Tesla and Nvidia for AI chip substrate lines" (Korean), 22 Oct 2026. Industry sources: Tesla about ₩1.1 trillion, Nvidia about ₩2.6 trillion for Sejong FC-BGA lines; mass production from 2H 2028; Tesla supply 2029–2033; Sejong budget ₩4.27 trillion.
2. Economic Daily News (Taiwan), "Three tech giants said to pay over NT$100 billion to lock substrate capacity" (Chinese), 24 Oct 2026. Reported deposits of more than NT$100 billion from Intel, Nvidia and AMD (analyst estimates); record NT$161.2 billion of plant investment; Unimicron capex NT$53.7 billion.
3. SETN, "Nan Ya catching Nittobo? Glass cloth leaders" (Chinese), 21 Oct 2026. Kinsus: ABF capacity 40 million to 80 million units a month (2026–2029); tool risk controllable, materials the real risk; T-glass gap 5–15%, widening in 2027; industry 2027 gap estimate 20–30%.
4. TechNews, "Glass cloth and CCL shortage pressure rises; price increases brewing for November" (Chinese), 22 Oct 2026. Glass cloth and laminate prices expected up about 10–15% from November; thin 2116/1080/1078 cloth tightest; relief 2H 2027 or 2028.
5. Edaily, "Samsung Electro-Mechanics' ₩360.5 billion contract is a signal" (Kiwoom Securities report), 20 Oct 2026. ₩360.5 billion memory substrate deal read as 2027 BT volume; BT utilisation mid-80%; ABF capacity more than doubling by 2028 vs BT mid-teens percent.
6. TTM Technologies, "TTM Technologies, Inc. Reports Third Quarter 2026 Results", 22 Oct 2026. Data center and networking sales $493.6 million, up 115%.
7. The Motley Fool, "TTM Technologies (TTMI) Q3 2026 Earnings Call Transcript", 22 Oct 2026. Commercial book-to-bill 2.5; laminate and T-glass lead times about 12 weeks, sometimes 24 or longer, and rising.
8. BE Semiconductor Industries, "BE Semiconductor Industries N.V. Announces Q3-2026 Results", 22 Oct 2026. Orders €342.6 million, up 96.1% y/y; hybrid-bonding orders more than doubled vs 2025; Q4 revenue flat ±5%, deliveries aligned to H1 2027.
9. The Next Web, "Intel CTO tells Dwarkesh Patel the world lacks tools to 10x chip output", 23 Oct 2026. Pushkar Ranade: tools for 'probably 2X' leading-edge output, 'not 10X'; about 18-month tool lead times.
10. Wikimedia Commons, "Fibreglass fabric 1100g per m2.jpg", image by Achim Hering, CC BY 3.0. Featured image base (high-resolution scan of woven glass-fibre cloth).