Contents
- Opening
- Executive Summary
- 1. Scale of the Phenomenon: Hyperscaler Investment
- 2. Why an AI Server Uses Several Times More Material
- 3. Supplier Policy: How AI Takes Priority
- 4. Map of Bottlenecks in the Supply Chain
- 5. Active Components: Memory, MLCC, Microcontrollers
- 6. Consequences for Non-AI Industries
- 7. Scenarios for 2026-2028
- 8. Operational Recommendations
- 9. Sources
Opening
Picture a scene familiar to any purchasing director or production manager at an electronics-assembly firm: an order confirmed a quarter ago, with a firm delivery date, suddenly dissolves. A phone call from the laminate supplier: “We are sorry — the current allocation does not cover your line items. Please reconfirm your forecasts for the next twelve months.” In the same weeks, identical material is loaded onto trucks bound for addresses controlled by the world’s largest technology firms. Not because they ordered earlier. Because their demand is now weighed differently.
This is neither anecdote nor vendor malice. It is the new reality of the global electronics supply chain, in which the artificial-intelligence revolution and the massive build-out of data-center infrastructure are absorbing materials and components faster than anyone can produce them. In 2026 the four largest AI-infrastructure operators plan to invest roughly USD 610 billion — 70% more than a year earlier. That money becomes servers of extreme density, and each such server uses three to five times as much material as traditional equipment: multilayer boards, specialty glass fabrics, ultra-smooth copper foil, hundreds of capacitors and high-performance memory. The rest of the market — automotive, industrial, consumer electronics, medical — is standing in a queue whose length nobody can currently estimate.
This report exists to understand that queue. Step by step we show where the AI industry’s suction effect on materials comes from, why suppliers — sometimes openly — favor AI customers, and what it means in practice for firms outside the circle of the largest technology buyers. We analyze bottlenecks layer by layer: from glass-fabric furnaces and copper smelters, through laminates and substrates, to memory and the smallest capacitors. We present supplier policy, the effects on specific sectors, and three scenarios for 2026-2028 with the indicators that allow real-time assessment of which one is materializing.
Most importantly, the report closes with concrete strategies to secure supplies, shorten lead times and slow the rise of costs — even when you are not a tier-one customer. The causes of this crisis sit on the supply side and will not disappear overnight, but the way you respond to them in the coming months will decide whether your company is a player or a bystander in this arrangement.
Executive Summary
Key Findings
- The global electronics supply chain entered 2026 in allocation mode. AI-segment customers receive priority service; every other buyer operates under quotas and extended lead times. What we see in our own purchasing is not a single-vendor anomaly — it is a systemic behavior confirmed in the quarterly disclosures of listed companies.
- The four largest AI-infrastructure operators have lifted their combined capital expenditure to roughly USD 610 bn in 2026 (+70% y/y). Goldman Sachs estimates cumulative capex for these firms in 2026-2028 at USD 5.3 trillion. This capital flow translates directly into demand for high-end laminates, HVLP copper foil, T-glass fabrics, HBM memory and AI-grade MLCCs.
- Supply of high-end materials is growing far slower than demand. According to Goldman Sachs projections, the effective-capacity gap for HVLP-3+ copper foil will reach 28% in 2026, 39% in 2027 and 38% in 2028. Specialty glass fabrics will remain in shortage until at least the second half of 2027. New furnaces and production lines come on line over an 18-30 month horizon.
- Material costs have already risen sharply. FR-4 laminate prices in China more than doubled in a year, Korean CCL import prices reached USD 20,728/ton (+74.5% y/y), the BOM cost of an AI-server board has risen 25-40% versus the second half of 2025. The scale of change requires recalculation of quotations in live projects.
- TTM Technologies, the second-largest US PCB manufacturer, stated in its Q2 2026 report that laminate suppliers are moving capacity from lower-complexity materials toward higher-complexity classes; the effect is more pronounced on the lower-complexity side — automotive and industrial. Our purchasing difficulties emanate from behavior that the management of a large public company is openly reporting.
Recommendations
- Secure supply via long-term agreements. Without a 12-18 month volume forecast, suppliers will not reserve capacity. Priority service goes to clients that commit to frame contracts.
- Qualify second sources for every critical material class. A parallel laminate vendor (EMC, Iteq, Shengyi, Kingboard, Nan Ya), HVLP/RTF foil alternatives, and a second PCB supplier in another geographic region.
- Check projects against available material classes. Dropping from ultra-low-loss (M6-M10) to High-Tg FR-4 cuts delivery from 6 months to 6-10 weeks and reduces material cost by 5x-20x.
- Introduce strategic buffering for critical materials. With lead times stretching to six months on high-end classes, holding minimum safety stocks is justified. Just-in-time stops working in an allocation regime.
- Monitor leading indicators on a quarterly cycle. LME copper price, pricing announcements from Kingboard, Resonac and Mitsubishi Gas Chemical, MLCC lead times at distributors, and TTM and Amphenol quarterly reports are leading signals for procurement-policy adjustment.
Scale of the Phenomenon: Hyperscaler Investment
The pace of investment growth in AI infrastructure in 2026 has no precedent in the recent history of the electronics industry. Amazon forecasts capex of USD 200 bn in 2026. Alphabet estimates spending in the USD 175-185 bn range. Meta plans USD 115-135 bn, and Microsoft USD 110-120 bn. The combined spending of the four largest cloud-infrastructure operators will reach approximately USD 610 bn — 70% more than a year earlier.
Including Oracle and smaller players lifts the total to roughly USD 690-760 bn, depending on methodology. Goldman Sachs estimates cumulative 2026-2028 capex of the four hyperscalers at USD 5.3 trillion — a sum with no equivalent in any previous electronics investment cycle.
The trajectory is asymmetric. Meta moves from USD 72 bn in 2025 to USD 115-135 bn in 2026. Microsoft lifts its spending from approximately USD 90 bn to USD 110-120 bn. Investment programs executed in 2026 were planned long before the current supply situation, so they cannot be expected to reverse quickly in response to material price hikes.
Every dollar of capex multiplies across complex electronics supply chains: servers, mainboards, components, laminates, copper foil, glass fabric, memory, capacitors, power, enclosures. The chain in 2026 is more heavily loaded than ever before.
| Hyperscaler | 2026 capex (USD bn) | Year-over-year dynamics and commentary |
|---|---|---|
| Amazon | ~200 | Largest in the group |
| Alphabet | 175–185 | Upper bound lifted during the year |
| Meta | 115–135 | Up from USD 72 bn in 2025 |
| Microsoft | 110–120 | Up from ~USD 90 bn in 2025 |
| Big-four total | ~610 | +70% y/y |
| Including smaller players | ~690–760 | Varies by methodology |
Source: company filings, Futurum Group, Statista, Goldman Sachs via Yahoo Finance.
What it means for us
The pace of hyperscaler investment sets a floor on demand for high-end materials over the next 24 months. Forecasts expecting an AI-demand slowdown turned out to be premature. High prices and extended lead times should be treated as the base state, not as a passing condition.
Why an AI Server Uses Several Times More Material
The very construction of an AI server explains why materials disappear from the market so quickly. The 2022 H100 platform had 16-20 PCB layers, with a per-rack PCB value of approximately USD 5,000. The GB300 platform introduced in 2025 already requires 26 or more layers, and per-rack PCB value reaches USD 35,000. The Rubin platform, whose mass production began in the second half of 2025, pushes the layer count to 32-40 and lifts per-rack PCB value to USD 116,000.
Year on year the average PCB layer count in AI servers has risen from 18 in 2023 to 32 in 2025 — a 78% increase over two years. The Ultra architecture planned for 2027 may push layer counts to 78. The density scale pursued by the AI industry has no equivalent in any other electronics sector.
Beyond boards, the AI server consumes several times more T-glass and quartz glass fabric (3-5x compared with a standard server), requires HVLP-2/3/4 copper foil instead of standard HTE (processing cost over 10x higher), and uses low-loss M6-M10 laminates whose unit price is 6 to 20 times that of FR-4. High-CV and high-voltage MLCC capacitors go into an AI server by the thousand, not by the hundreds as in traditional equipment.
The numerical consequence: the share of material cost in PCB value has crossed 50%, while historically it stayed below 20%. What the AI server purchases is not a single board but a system of extreme density, in which materials — not processing — carry the bulk of value.
| Parameter | Standard server | AI server (Rubin / GB300) |
|---|---|---|
| PCB layer count | 8 to 16 | 26 to 40 (up to 78 in Ultra arch., 2027) |
| Average layer count (y/y) | 18 (2023) | 32 (2025), +78% over 2 years |
| PCB value per rack | ~USD 1–5 thousand | USD 35 k (GB300) – USD 116 k (Rubin) |
| Glass fabric | Standard-grade | 3–5x more; low-Dk / T-glass |
| Copper foil | HTE (standard) | HVLP-2/3/4 ultra-smooth (cost over 10x) |
| Laminates | FR-4 | Low-loss M6-M10 (price 6–20x FR-4) |
| MLCC capacitors | Hundreds | Thousands; high-CV, high-voltage |
Source: TPCA via PCBSHOP (17.06.2026); NVIDIA platform disclosures.
What it means for us
Even a moderate increase in AI-server production volume drives several times more demand for high-end materials than a comparable increase in any other segment. From our position, every new generation of an AI hyperscaler’s platform translates into a fresh wave of price hikes and extended lead times across our purchasing book.
Supplier Policy: How AI Takes Priority
Priority service for AI customers is not a salesperson’s anecdote. It is a systemic behavior that the management of large listed companies openly discloses. TTM Technologies in its Q2 2026 call described the situation as follows: laminate suppliers are shifting production from lower-complexity materials toward higher-complexity products, and the shift effect is more pronounced on the lower-complexity side.
The shift is logical from a margin perspective. High-end laminates give producers higher per-unit margin. Producers therefore raise the share of high-end classes in their product mix, even when it means lower availability for standard classes. From the perspective of an end customer who needs standard classes, the effect is the same as a physical shortage of goods.
Allocation mechanisms operate in parallel at several levels. Kingboard, the world’s largest laminate producer, has implemented 8 to 10 rounds of price hikes between February 2025 and mid-2026. Resonac raised prices by 30% from March 2026. Mitsubishi Gas Chemical applied a 30% increase from 1 April 2026. These pricing decisions are correlated, which points to coordinated action rather than independent calls.
AI customers secure material access through long-term agreements. ABF substrates and memory are blocked by AI customers over 2-3-year horizons, before the material is even available at full scale in the market. Large groups have placed orders with EMC/TUC for laminates worth more than KRW 10 bn (5x monthly consumption) months in advance. Customers outside this orbit operate in the residual pool, which by definition is smaller and more volatile.
In the shorter links of the chain we observe the same pattern. AI-grade MLCCs at Murata, Samsung Electro-Mechanics and TDK are formally under allocation; lead times have risen from 8 to 20-40 weeks, and price hikes from July 2026 reach 10-40%. MCU producers prioritize the highest-margin customers: TI raises prices 15-85% from April 2026, Infineon reports 20-30 weeks lead time, and Nexperia has frozen deliveries in selected product lines.
| Allocation mechanism | How it works in practice |
|---|---|
| Producer mix shift | CCL makers prefer high-end classes (higher margin); lower classes lose capacity. TTM Q2 2026: “effect more pronounced on lower-class materials”. |
| Quota systems | You receive whatever is granted, not what you ordered. Applies to M6-M10 laminates and AI-grade MLCC. |
| Long-term agreements (LTA) | AI clients reserve capacity 1–2+ years ahead. Applies to ABF substrates and HBM memory. |
| Correlated price hikes | Kingboard 8–10 rounds 2025–2026; Resonac and MGC +30% from March/April 2026. Decisions communicated near-simultaneously. |
| Highest-margin-client priority | Higher AI-class prices absorb capacity; commodity-class availability stays stable in price but shrinks in volume. |
| Delivery freezes | Extreme form of allocation in selected lines. Nexperia: deliveries effectively frozen in some positions. |
| Forward purchasing | Pre-buying as a buyer’s leverage; used by PCB makers for confirmed orders. |
Source: TTM and Amphenol quarterly reports; CCL producer price releases; industry publications (Refs 1, 3, 4, 5, 6, 12, 13, 14).
What it means for us
As a customer outside the AI segment we compete for material with large firms that enjoy higher price elasticity and are ready to sign LTAs months in advance. Without an analogous offer from our side we will not obtain priority service, even when our forecasts are correct.
Map of Bottlenecks in the Supply Chain
The earliest and deepest bottleneck is specialty glass fabric: T-glass and quartz. An AI server uses 3 to 5 times more glass fabric than a standard server, and new furnaces to produce it are built on multi-year timelines. According to TPCA the supply-demand gap for these fabrics will exceed 60% by 2027. New capacity comes on line at the earliest in mid-2027, with production stabilization not before around 2028.
The second bottleneck is HVLP copper foil. It requires ultra-smooth surfaces and processing significantly more expensive than standard HTE foil. According to Goldman Sachs, the effective HVLP-3+ capacity gap will stand at 28% in 2026, 39% in 2027 and 38% in 2028. Goldman also projects HVLP-foil demand growth at 97% CAGR through 2029, from roughly 679 metric tons per month today to over 5,000 metric tons monthly.
The third link is CCL laminates. Korean import prices reached USD 20,728/ton (+74.5% y/y, the first time above USD 20,000 since 2000). Kingboard raised FR-4 from RMB 70 per sheet in July 2025 to RMB 260 in June 2026 — a rise above 270%. Prices in individual classes rose 20-45% in the first half of 2026. ABF substrates are the longest-delayed link: AI customers block multi-year contracts, and lead time is estimated at 52-78 weeks.
A fourth, less spectacular but significant link is production equipment. mSAP lasers are booked until 2027. Toyota looms produce 2,000-2,500 units per year against demand of 6,000-7,000. Order visibility for CCL production equipment extends into 2028. Even if capital is available, physical expansion capacity is constrained.
In April 2026, a non-AI disruption occurred. An attack on the Jubail petrochemical complex in Saudi Arabia (operator: SABIC) interrupted about 70% of global PPE resin production. Resin prices rose roughly 400% (from RMB 120,000/ton to RMB 600,000/ton) and PCB prices jumped 40% within a month. This is one of the reasons high-end laminates rose so sharply in the first half of 2026.
| Supply-chain link | Status, mid-2026 | Key figures |
|---|---|---|
| Glass fabric (T-glass, quartz) | Earliest and deepest bottleneck; furnaces built on multi-year cycles | 3–5x consumption; >60% gap by 2027; new capacity from H2 2027 |
| HVLP copper foil | Effective-capacity gap widening | 28/39/38% gap (2026/2027/2028, Goldman); 97% CAGR demand through 2029 |
| PPE resins | Geopolitical disruption (Jubail/SABIC, ~70% of capacity) | +400% resin price; +40% PCB price within a month (April 2026) |
| CCL laminates | Global tension; +20–45% prices | FR-4: 4–8 wk lead time; High-Tg: 6–10 wk; M6-M10: up to 6 mo; Korea USD 20,728/ton, +74.5% y/y |
| ABF substrates | Longest queues; AI customers block LTAs | Lead time 52–78 weeks (estimate); capacity +37%, AI demand above 50% y/y |
| Production equipment | New bottleneck on capacity expansion | mSAP lasers booked to 2027; looms 2,000–2,500 units/yr against 6,000–7,000 demand; visibility to 2028 |
Source: TPCA, Goldman Sachs, TrendForce, NextPCB, Highleap; Utmel quoted for ABF substrates (industry estimate, flagged).
What it means for us
Each of these links is a bottleneck for our specific project portfolio. Projects using high-end classes should expect extended lead times even into 2027. Projects on FR-4 are in a better position, but the Jubail incident shows that even apparently stable classes can be disrupted from independent sources.
Active Components: Memory, MLCC, Microcontrollers
Active-component constraints are as severe as those on base materials. IDC projects that in 2026 data centers will absorb roughly 70% of global memory output. That means the remaining 30% must serve smartphones, PCs, automotive and industrial. HBM memory is sold out through the end of 2026. PC DRAM recorded a 105-110% quarter-on-quarter rise in Q1 2026.
MLCC capacitors are formally allocated by the three largest producers — Murata, Samsung Electro-Mechanics and TDK. Lead times for AI-server classes reach 20 to 40 weeks, compared with around 8 weeks in normal conditions. Price hikes from July 2026 range from 10 to 40% depending on specification. AI-server and automotive-premium classes pay up to 3x market price. Commodity classes stay price-stable but tighten on availability.
In the microcontroller and mature-node segment the pressure has a different shape. Semiconductor foundries prioritize advanced lithography and advanced packaging for AI accelerators. The consequence is shortages of 8-bit and 32-bit MCUs. Texas Instruments has lifted prices 15-85% from 1 April 2026. Infineon reports 20-30 week lead times. Nexperia has frozen deliveries of selected positions.
Power devices (PMIC, VRM, SiC components) are another link in which the bottleneck has moved from logic to power. Power-IC shortages will persist through 2026. The effect cuts across every device category — from 5G base stations to electric vehicles.
Active components behave differently from base materials. Base materials can be buffered, with effort. Active components cannot be buffered effectively, because their specifications and firmware versions change quickly. The buffer must rotate, which generates costs from obsolete-stock disposal. A coherent active-component strategy requires a different approach from base-material buffering.
| Component category | Status, mid-2026 | Key figures |
|---|---|---|
| DRAM and HBM memory | Allocation: data centers ~70% of memory output in 2026 | HBM sold out through end-2026; PC DRAM +105–110% QoQ in Q1 2026 |
| MLCC capacitors | Allocation at Murata, Samsung EM, TDK; AI and AEC-Q200 classes | 20–40+ week lead time; +10–40% price from July 2026; up to 3x market for AI |
| Microcontrollers (MCU) | Foundry priority for AI; 8-bit and 32-bit MCU shortages | TI +15–85% from 1.04.2026; Infineon 20–30 weeks; Nexperia froze deliveries |
| Power devices (PMIC, VRM, SiC) | Bottleneck moved from logic to power | Power-IC shortages present through 2026 |
Source: IDC, TrendForce, Murata, TI, Infineon; industry allocation data (Refs 6, 11, 12, 14).
What it means for us
Projects that require specific premium MCUs or MLCCs must plan deliveries at least a quarter ahead and run active-component stock-rotation as a separate process from base-material buffering. Without that separation, obsolescence drives disposal costs that erode margin.
Consequences for Non-AI Industries
Cascade effects are not confined to PCB manufacturers. They are visible across every electronics sector. In automotive, TTM Technologies forecasts a sales decline of mid-single-digit percent for 2026, with the shortage of lower-class CCL materials cited as one direct cause. EE Times confirms that automotive shortages stem from foundry priorities.
In consumer electronics IDC describes a “memory crisis” that reshapes smartphone and PC markets. Memory price increases, changes in device specifications (less RAM in new phones) and supply-availability risks are already observable. Specification cuts driven by cost are also appearing across product lines.
In industrial and automotive electronics MCU prices are rising at double-digit rates and lead times are multiples of historical norms. For firms whose products depend on specific MCU models the situation forces redesign or acceptance of higher costs — often both at the same time.
In medical and defense the situation is relatively better — but more expensive. Amphenol reports +24% y/y in defense and +21% in commercial aviation. TTM has invested in Syracuse for Ultra-HDI production serving Aerospace & Defense. Acquisitions of STG and ILFA grow its medical exposure. The reason is straightforward: these customers pay higher margins and carry regulatory priority.
For every B2B segment the BOM increase is already measurable. AI-server board BOM rose 25-40% versus H2 2025; in other sectors the trend is slightly weaker but parallel. Any offer that assumes flat prices or margins anchored to historical material costs becomes impossible to sustain in 2026.
| Sector | Observed effect | Evidence and source |
|---|---|---|
| Automotive (vehicle electronics) | Volume decline; pressure on lower-class CCL availability | TTM Q2 2026: automotive mid-single digits y/y; EE Times on foundry priorities |
| Consumer electronics (PCs, smartphones) | Price rises, spec cuts, production constraints | IDC: memory crisis reshapes 2026 PC and smartphone markets |
| Industrial and automotive electronics (MCU) | Long lead times, double-digit price rises; redesign or higher-cost acceptance | TI +15–85%; Infineon 20–30 weeks; Nexperia froze deliveries |
| Medical and Aerospace & Defense | Relatively shielded (regulatory priority), but more expensive | Amphenol: Defense +24% y/y, Commercial Air +21%; TTM Syracuse Ultra-HDI A&D |
| All B2B sectors | BOM cost increase; 12-month+ forecast discipline required | AI-server BOM +25–40% vs H2 2025; laminate lead times up to 6 months |
Source: TTM and Amphenol quarterly reports; IDC; EE Times.
What it means for us
We are not alone in this situation. The effects are also visible to other end customers in automotive and consumer electronics. This is an opportunity to renegotiate commercial terms with our own customers. Those who have already done so are in a stronger position than those still waiting.
Scenarios for 2026-2028
The projections in this section are flagged as scenarios, not facts. Three scenarios are defined along the pace of new-capacity ramp-up, material-price dynamics and the behavior of non-AI sectors. Each has distinct consequences for purchasing, stock policy and quotation pricing.
The Base scenario assumes that new capacity (Iteq Jiangxi from September 2026, EMC/TUC from 2027, Doosan Thailand from H2 2028) comes on stream as planned, and that data-center build pace is partially slowed by power-supply limits. Laminate prices stabilize at an elevated level (FR-4 roughly 2x vs 2025), with general-purpose capacity appearing in H2 2027-2028. Availability for non-AI sectors is moderate and planable.
The Strained scenario (currently the most probable) materializes when the HVLP gap of 39% in 2027 and the glass-fabric gap above 60% translate into PCB-production limits, while AI demand holds in line with hyperscaler capex. High-end-class prices grow another 10-20% y/y, lead times run 20-26 weeks, and allocations are the norm. Availability for non-AI sectors is constrained, queues on the order of 6-12 months, and BOM-cost growth stays at 25-40% y/y.
The Crisis scenario materializes when geopolitical escalation (further Jubail-like resin incidents) or panic-buying is layered on top of the existing deficit. Possible step-change price increases of 50%+ on critical classes, lead times beyond six months, and a quota regime that reaches most classes. The time horizon is undefined; crisis episodes can persist indefinitely.
The indicator that allows you to judge which scenario is materializing is the quarterly data stream. LME copper price, the count of pricing-initiation rounds by Kingboard, Resonac and MGC versus the year before, MLCC and ABF lead times in distributor reports (Fusion, Avnet), and the language used in TTM and Amphenol quarterly statements about allocation and lead times are the strongest leading signals.
| Aspect | Base | Strained (most likely) | Crisis |
|---|---|---|---|
| Key assumptions | New capacity on plan; DC build partially slowed by energy | HVLP gap 39% and glass-fabric gap above 60% materialize; AI demand aligned with capex | Geopolitical escalation (Jubail-type), panic buying, delivery freezes |
| Time horizon | H2 2027–2028 | 2027 – early 2028 | 2026+ (no fixed horizon) |
| Laminate prices | Stabilization at elevated level (FR-4 ~2x vs 2025) | Further +10–20% y/y on high-end | Step +50%+ on critical classes |
| Lead times | FR-4: 4–8 weeks; high-end: 12–20 weeks | High-end: 20–26 weeks; allocation widespread | Over 6 months; quota regime widespread |
| Availability for non-AI sectors | Moderate, planable | Constrained; queues 6–12 months | Some classes unavailable; production stoppages |
| BOM-cost increase y/y | +10–20% | +25–40% (sustained) | +50% and above |
Source: own scenarios, built on Goldman Sachs, TrendForce and IDC data.
Early-warning indicators (monitor quarterly)
| Indicator | What to watch | Data source |
|---|---|---|
| LME copper price | Record USD 13,300/ton in January 2026; y/y trend | LME / Reuters |
| CCL price announcements | Number of price-rounds y/y at Kingboard, Resonac and MGC | Producer announcements / Highleap / Vexos |
| MLCC / ABF lead times | Lengthening in AI and AEC-Q200 classes | Fusion WW / Avnet / Hotai |
| TTM and Amphenol reports | Data-center growth pace; language on allocation | investors.ttm.com / investors.amphenol.com |
| New-capacity launches | Iteq Jiangxi (IX 2026); Doosan Thailand (H2 2028); EMC/TUC | I-Connect007 / TrendForce |
| PPE geopolitical risk | PPE resin price; SABIC facility status | NextPCB / Highleap |
Source: own compilation.
What it means for us
Operational planning should anchor on the Strained scenario for budget purposes, with the Crisis scenario held as a contingency. Quarterly values of each indicator should be recorded in an internal report so we can adjust direction quickly when data points change.
Operational Recommendations
The recommendations below are split across four areas: sourcing, engineering/DFM, pricing policy and monitoring. Each recommendation has a deployment horizon and an expected effect. Roll-out should be coordinated across Purchasing, Engineering and Sales.
Long-term agreements (LTAs) with 12-18 month forecasts. What AI customers are doing is the model to emulate: reserving capacity with producers in advance and reporting forecasts consistently. Priority service is not guaranteed by declarations but by volume commitment. Deployment horizon: Q4 2026 – Q1 2027. Expected effect: stabilization of critical-material availability.
Qualification of secondary sources. For every critical material class we must have a qualified second supplier. For laminates the alternates are EMC, Iteq, Shengyi, Kingboard and Nan Ya. For copper foil — HVLP / RTF alternatives. For PCBs — a second manufacturer in a different geographic region (Malaysia, Taiwan, USA). Deployment horizon: 6-9 months for engineering qualification. Expected effect: emergency supplier in case of allocation at the primary source.
DFM for availability. New projects should start from the question of which material class is truly needed and whether we can step down a class. Moving from M6-M10 to High-Tg FR-4 cuts lead time from 6 months to 6-10 weeks. Qualifying RTF instead of HVLP where parameters allow reduces dependence on the foil bottleneck. Deployment horizon: ongoing process from Q4 2026. Expected effect: shorter lead times and lower material prices.
Strategic buffering. With lead times of up to six months on high-end classes, holding minimum safety stocks is justified. The same applies to active components — though their case requires stock-rotation rules, as components have shorter lifecycle than materials. Deployment horizon: rotation rules online by Q1 2027. Expected effect: 4-8 week production buffer for high-end materials.
Update of pricing policy. Quotations issued in 2026 should plan for material-cost increases of 25-40% versus end-2025. Long-term contracts with our own customers must include revaluation or renegotiation clauses for material prices, otherwise margin erodes during execution. Deployment horizon: Q4 2026 for all new quotations; renegotiation of selected existing contracts by Q2 2027. Expected effect: margin protection under cost volatility.
Geographic diversification. New capacity is being built in South-East Asia (Doosan Thailand, EMC/TUC) and in the USA (TTM Syracuse, Sterling). A European EMS provider should have qualified partners in at least two regions. Deployment horizon: capability audit by Q1 2027. Expected effect: reduced dependence on any single procurement region.
The full “solution” of the problem sits on the supply side. New glass furnaces, HVLP capacity and CCL plants are scheduled to come online in the 2027-2028 window. Local actions can mitigate effects but cannot remove the underlying cause. AI demand will outstrip high-end material supply for at least 12-18 months from September 2026.
Sources
A. Quarterly Reports of Listed Companies
- Ref. 1. TTM Technologies, Inc. — Q2 2026 quarterly report and earnings call: data-center segment +91% y/y; automotive mid-single-digit negative; CCL producer policy; Penang and Syracuse expansions; brownfield strategy (investors.ttm.com).
- Ref. 2. Amphenol Corporation — Q2 2026 quarterly report and earnings call: IT datacom +89% USD / +63% organic; no broad bottlenecks; capex 3-4% of sales (investors.amphenol.com).
B. Market Analyses and Data
- Ref. 3. TPCA via PCBSHOP — “AI Computing Demand Drives PCB Upstream Material Prices Into a Super Cycle”, 17.06.2026: NVIDIA platform layers and value, CCL market, HVLP-4, quartz fabric, PPE, equipment (pcbshop.org).
- Ref. 4. TrendForce — “AI Demand Tightens CCL Supply, Drives Advance Orders, Price Hikes and Expansion”, 06.05.2026: Korean CCL import price USD 20,728/ton, CCL cost structure, KRW 10 bn forward orders, EMC/TUC/Doosan expansions (trendforce.com).
- Ref. 5. NextPCB — “PCB Raw Material Shortage in 2026: Full Timeline, Data and How to Manage It”, August 2026: 140-day lead time, FR-4 RMB 230-260, Kingboard +50%, Resonac/MGC +30%, PPE/Jubail, copper USD 13,300/ton (nextpcb.com).
- Ref. 6. Highleap Electronics — “PCB Manufacturing Cost 2026: Full BOM Crisis Explained”: BOM +25-40%, quota systems, FR-4/High-Tg/M6-M10 lead times, TI/Infineon/Nexperia, DRAM +105-110% QoQ (hilelectronic.com).
- Ref. 7. Vexos — “How OEMs Can Navigate AI-Driven PCB & CCL Supply Constraints” and “Copper Foil: The Next Critical Bottleneck”: 18 to 32 layers; Goldman Sachs quote on HVLP-3+ gap 28/39/38% (vexos.com).
- Ref. 8. Yahoo Finance via Goldman Sachs — “Meta, Microsoft, Amazon, and Alphabet Are About to Spend…”: 2026 capex ~USD 610 bn, +70% y/y; cumulative 2026-2028 USD 5.3 trillion (finance.yahoo.com).
- Ref. 9. Futurum Group — “AI Capex 2026: The $690B Infrastructure Sprint”: Amazon USD 200 bn, Alphabet USD 175-185 bn, Meta USD 115-135 bn (futurumgroup.com).
- Ref. 10. Statista — “Big Tech’s AI Spending to Reach $760 Billion in 2026” (statista.com).
- Ref. 11. IDC — “Global Memory Shortage Crisis: Market Analysis and the Potential Impact on the Smartphone and PC Markets in 2026”: data centers ~70% of memory output in 2026 (idc.com).
C. Industry Publications (estimates flagged in the text)
- Ref. 12. Hotai — “2026 MLCC Shortage: Why Procurement Teams in the US/Europe…”: MLCC allocation confirmation at Murata/Samsung EM/TDK (hotai-int.com).
- Ref. 13. Utmel — “2026 Advanced Packaging Components Trends”: ABF lead times 52-78 weeks, industry estimate (utmel.com).
- Ref. 14. Passive Components EU / Murata announcements: MLCC price hikes 10-40% from July 2026 on AI-server and automotive classes (passive-components.eu).
- Ref. 15. I-Connect007 — “CCL Makers Stick to Capacity Expansion Plans”: new Iteq Jiangxi factory, +600,000 sheets/month (iconnect007.com).
- Ref. 16. Korea Herald via TrendForce: Doosan KRW 180 bn (USD 121.7 mn) investment in new CCL plant in Thailand, H2 2028 production (koreaherald.com).
- Ref. 17. Accuristech via Omdia: 30-50% of planned data centers may never be built due to power-supply limits, analytical estimate (accuristech.com).
Note: Figures flagged as industry estimates (ABF substrates, parts of the MLCC data) have not been confirmed in audited financial reports or with the original vendors. They should be treated with caution. All other values are drawn from class-A or class-B sources, verified as of 3 September 2026.
