Stolen capacity! What are the eight core bottlenecks for the future development of AI?
```
The race for AI infrastructure has entered a new stage where "physical constraints outweigh algorithmic innovation." Victory no longer depends on who has the superior model architecture or more complete software stack, but on who can secure the most scarce "resource locks" in the global supply chain ahead of others. TSMC CoWoS capacity, HBM memory chips, 3nm/2nm advanced process, ABF substrates, high-end CCL copper-clad laminates, ultra-thin electronic cloth, power transformers, gas turbines—these eight core production nodes have all, without exception, been preemptively locked down by tech giants like NVIDIA, Apple, Microsoft, and Google through 2027 and even into 2028.
Why have these production capacities been preemptively seized?
I. What happened? Industry transformation
The race for AI infrastructure has entered a new stage where "physical constraints outweigh algorithmic innovation." Victory no longer depends on who has the superior model architecture or more complete software stack, but on who can secure the most scarce "resource locks" in the global supply chain ahead of others. TSMC CoWoS capacity, HBM memory chips, 3nm/2nm advanced process, ABF substrates, high-end CCL copper-clad laminates, ultra-thin electronic cloth, power transformers, gas turbines—these eight core production nodes have all been locked down ahead of time by NVIDIA, Apple, Microsoft, Google, and other tech giants extending to 2027 and even 2028.
Why have these production capacities been preemptively seized?
The fundamental reason is that explosive growth in AI computing demand far exceeds the physical expansion capacity of the global semiconductor supply chain. On the demand side, all major AI accelerators—NVIDIA Rubin, Google TPU v7/v8, AWS Trainium 3, AMD MI350X—will simultaneously migrate to the most advanced nodes in 2026; on the supply side, capacity expansion in advanced packaging, HBM, ABF substrates, etc., is limited by equipment lead times, material bottlenecks, and yield ramp-up. Even with the most aggressive capital spending, it falls far short of the exponential growth in demand. The final result: these "locked" segments stand out in the short term—capacity is sold out, lead times extend to 6-12 months or even over 30 weeks, prices increase by 10%-30% per quarter, and supply chain power is shifting totally from design to backend production capacity.
The structural scarcity in AI computing capacity is not accidental, but driven by three long-term factors. First, the physical limits of the supply chain: 2nm wafer foundry pricing is as high as $30,000, the area of intermediary layers in advanced packaging is evolving from 2,800mm² to even larger, and yield remains an uncertain variable; second, strategic restructuring of customer relations: AI chips have gone from "mass commodity" to "strategic resource," with cloud giants like Microsoft and Google flying directly to Korea to lock in HBM capacity, willing to make 30% prepayments for 3-5 year long-term contracts; third, concentrated technical iteration: upgrades from CCL M8 to M9 and M10, the leap from HBM3E to HBM4, the change from traditional to vertical power delivery architectures, with all technological upgrades happening in the same window. The value distribution upstream in the industry chain is undergoing profound reshaping—ABF substrate prices keep rising, CCL prices see weekly increases of 10%-20%, upstream raw materials account for over 60% of PCB cost structure, and price increases at each link are passed down layer by layer.
Taking into account factors like capital expenditure, technical barriers, and order visibility, we believe 2026-2028 will be the "settling years" for the AI supply chain. Industry leaders with large-scale capacity and technological advantage will reap sustained excess returns in this phase of scarcity, securing victory.
This is a significant paradigm shift: the AI boom's competition dimension has shifted completely from "algorithm superiority" to "grabbing underlying physical production capacity." Under the pressure of NVIDIA Blackwell Ultra and Rubin architecture evolution, global tech giants’ procurement strategy has shifted from "ordering as needed" to "strategic prepayment."
The essence of capacity plundering: This is a multidimensional blockade around "signal integrity (PCB/CCL)," "energy density (supercapacitor/power)," "logistics channels (shipping space)," and "compute futures (NeoCloud)."
II. Why is it important? The eight core production capacities surface
We break down these eight capacity bottlenecks to outline the full picture of the current supply chain crisis.
① TSMC CoWoS: The AI chip packaging dilemma
CoWoS (Chip-on-Wafer-on-Substrate) is TSMC’s advanced packaging technology and one of the most serious bottlenecks in current AI chip production. Its essence lies in heterogeneous integration—connecting several small chips (compute dies, HBM, I/O dies, etc.) through silicon intermediary layers with ultra-high density, integrated in one package, overcoming the size limit for a single chip.
Production capacity data is persuasive. By end of 2024, TSMC’s CoWoS capacity will be about 35,000 wafers per month; by end of 2025, 80,000 wafers, more than doubling; end-2026 target is 115,000-130,000 wafers; 2027 is expected at 145,000 wafers. Overall, monthly CoWoS capacity in 2025 is about 65-70,000 wafers, and in 2026, 120-130,000 wafers; even with rapid expansion, it falls short of demand. Morgan Stanley estimates TSMC CoWoS monthly capacity will reach at least 120-130,000 wafers, and TSMC CEO C.C. Wei says: “Our CoWoS capacity is very tight and will remain sold out through 2025 and 2026.”
Even at such aggressive speeds, capacity still lags demand. NVIDIA alone has locked in over 60% of TSMC's CoWoS capacity for 2025-2026, for Blackwell and the upcoming Rubin architecture, with capacity sold out to the end of 2026, and advance booking into 2027. Google’s TPU capacity failed to meet expectations due to CoWoS shortages. “Packaging has become the narrowest bottleneck for AI computing”—CoWoS intermediary layer area has grown from about 800mm² to over 2,800mm², and technical difficulty is rising exponentially. TSMC is building new packaging facilities (AP7 and AP8) in Chiayi and Tainan and planning CoWoS capacity in Arizona, but ramping new capacity takes time, and the tension in 2026-2027 will be very hard to ease.
② TSMC wafer capacity: The N3 and N2 contest
If CoWoS is the packaging bottleneck, TSMC’s N3 (3nm) and N2 (2nm) process is the bottleneck in wafer manufacturing. In 2026, AI-related demand will take up close to 60% of N3 capacity, leaving 40% for smartphones and CPUs. All major AI accelerators will migrate to N3 in 2026: NVIDIA Rubin, Google TPU v7/v8, AWS Trainium 3, AMD MI350X. This simultaneous migration creates unprecedented production pressure. Major HPC customers have pre-booked N3 and N2 capacity through 2027, and Broadcom executives confirm TSMC’s advanced process capacity is pre-booked through 2028.
More critically, TSMC has pushed effective utilization rate to the limit through process optimization. TSMC says demand for advanced node wafers is now “about three times the company’s available capacity.” In 2027, things will be even more extreme: AI demand is expected to occupy 86% of N3 wafer output, almost completely squeezing out smartphone and CPU orders, forcing some smartphone lines to switch to N2 early.
N2 process is not optimistic either. TSMC starts volume production in Q4 2025, with initial capacity at 90,000-100,000 wafers per month. N2 uses GAA (gate-all-around) nanowire transistor architecture, offering 10-15% speed boosts or 25-30% power reductions compared to 3nm. However, Apple has locked in over 50% of early N2 capacity for 2026-2027 for its A20/A20 Pro chips; AMD, MediaTek, and Qualcomm flagships will compete for the rest. 2nm wafer foundry pricing has hit $30,000, making it unaffordable for many players.
③ HBM high-bandwidth memory: Sold out through end-2026
HBM (High-Bandwidth Memory) is the most performance-sensitive and supply-constrained component in AI accelerators. The severe tension in the supply chain is most evident in the HBM segment.
SK Hynix CFO said: “We have sold out all HBM supply for 2026.”
Micron CEO confirmed: “Our 2026 HBM capacity is fully pre-booked.”
Samsung responded with price hikes, raising HBM prices by double-digit percentages in 2026 contracts.
The three main suppliers—SK Hynix, Samsung, Micron—have all sold out their 2026 HBM capacity, largely via long-term contracts. SK Hynix said: “It will be very difficult to make meaningful adjustments to HBM and standard DRAM production lines in 2026,” and new orders are queued into Q1 2027. Samsung raised NAND flash supply prices by 100% in Q1 2026.
The structural causes of HBM shortage are multiple. First, HBM uses more process steps than standard DRAM, and HBM3E has a longer verification cycle. Second, the technology leap from HBM3E to HBM4—HBM4 uses a 2,048 bit interface, needs 12 and 16-layer stacking, with much higher thermal density and height control difficulty. Third, cloud giants are locking in allocations of HBM3E and next-gen HBM4 for multiple years—Microsoft and Google flew to Korea to negotiate with SK Hynix, made 30% deposits, signed 3-year contracts with price floor clauses. SK Hynix is using next-gen HBM supply as bargaining chips, hoping to extend current relationships for another 2 years.
④ ABF substrates: T-Glass shortage ignites the supply chain
ABF substrate is the core base material for packaging high-compute chips in AI servers. The root of current supply tension is serious shortage of upstream key material—T-Glass fiberglass cloth.
T-Glass has a low thermal expansion coefficient and low signal loss, making it key to ABF and BT substrate production. Over 80% of the global T-Glass market is supplied by Japan’s Nitto and US's PPG, and when AI demand exploded in 2025, supply tightness emerged immediately. Foreign surveys show a 25% supply gap, lead times extending from 8-10 weeks to over 30 weeks, and Taiwan substrate factory inventory at only about 2 months. Prices jumped about 30% in a few months.
Ibiden, as the main ABF supplier for NVIDIA AI servers, decided to speed up expansion due to chip orders, and its new factory will start with 25% capacity in Q4 2025, reaching 50% capacity by March 2026. But Ibiden says new capacity may still fail to fully meet demand.
Foreign institutions have deeper forecasts on ABF substrate supply and demand. US reports say supply structure turned at end-2025, tension intensifies month by month; shortfall rate will reach 10% in H2 2026, 21% in 2027, and 42% in 2028, similar to the 2020 supply shortage. At the time, prices rose 20-30% annually, and price increases are likely in coming quarters.
⑤ High-end CCL copper-clad laminates: The M7→M8→M9→M10 upgrade wave
Copper-clad laminates (CCL) are the core material for PCB manufacturing and the most prominent link in the current price surge. The M-series numbers directly indicate technical generations—the higher the grade, the better the material performance and the smaller the loss. The AI boom is driving use of M2-M8 series high-speed CCL; NVIDIA Blackwell platform has upgraded to M8, Rubin platform may introduce M9/M10.
M9 material is a high-frequency, high-speed CCL developed for the next-gen Rubin architecture AI servers by NVIDIA, its core components include specialty resin, quartz cloth (Q cloth), and high-end copper foil (HVLP4/HVLP5). With the expected launch of the Rubin platform in 2026, the M9 CCL market may see demand surge. M10 uses hydrocarbon resin and electronics-grade quartz cloth composite; NVIDIA is testing M10, aiming for mass production at Rubin Ultra and Feynman platforms in 2027.
Price signals start from international giants. Japan’s Resonac raised copper foil base and adhesive prices by over 30%, Mitsubishi Gas Chemical raised high-end PCB material prices by 30% from April 1. Kingboard recently issued a price increase notice: recent chemical product price surge and tight supply caused CCL costs to rise sharply, so all board materials and PP (prepreg) prices were raised by 10%.
From December 2025, Kingboard, Nan Ya New Material, and others frequently released price adjustment notices; CCL weekly price surges reached 10%-20%. Price surges are supported by real supply and demand—single AI server’s PCB usage grows 3-5 times over traditional servers, value grows 8-12 times. Upstream raw materials account for 60% of PCB costs; CCL, copper foil, and prepreg price fluctuations directly determine the industry chain’s profitability.
In CCL cost structure, copper foil makes up about 42%, resin 26%, fiberglass cloth 20%. There’s a serious supply gap in high-end fiberglass cloth (first, second generation, and Q cloth), with Apple, Qualcomm, and AI server suppliers competing for limited supply.
⑥ Ultra-thin electronic cloth: Structural shortage due to AI capacity squeeze
Electronic cloth is the core reinforcement material in CCL, and this round of shortage shows clear structural characteristics. AI compute demand is pushing fiber producers to shift capacity towards LDK fiberglass cloth, quartz cloth, and other low-CTE high-end types, causing structural shortages in ordinary 1080, 2116, and 7628 standard cloth.
From H2 2025, a clear price surge cycle began. In Oct/Dec 2025, Jan/Feb 2026, ordinary electronic cloth saw four price hikes: thick 7628 cloth up 1-1.2 yuan/meter, thin cloth even more. Nitto raised fiber prices 20% from August 2025, with strong expectation of catch-up increases. More importantly, ultra-thin cloth (1080 type) supply is limited by imported loom bottlenecks, domestic equipment can’t meet process precision needs, so shortages likely all year.
⑦ Transformers and vertical power delivery modules: From power demand to physical constraints
As GPU power consumption climbs from 700W to over 1,400W, AI server power systems are transforming from “horizontal delivery” to “vertical power delivery” (VPD). Traditional horizontal delivery, when GPU current exceeds 850-1,000A, loses over 100W in the network. VPD architecture delivers power vertically through PCB layers directly to the processor, cutting total resistance from 90-140μΩ to 10-15μΩ.
Suppliers like Infineon have evolved vertical power modules, with the third generation reaching 24A/mm² current density. More importantly, widespread use of VPD requires redesign of server architecture and higher capacity for supporting components like transformers, inductors, capacitors.
⑧ Gas turbines: “Energy infrastructure” of computing capacity
The scaling up of AI data centers is becoming an important incremental driver for natural gas power generation. Gas turbines, as core equipment for data center backup and distributed power, face capacity strain as well. This is directly linked to server power and cooling needs—NVIDIA GB200 NVL72 individual cabinet power consumption has reached 1,200W, with whole cabinet data center power demand advancing from MW to GW scale.
III. What’s next? Industry structure and technology evolution
This round of AI supply chain competition shows clear polarization.
The winner-take-all effect is most evident at CoWoS—NVIDIA alone has locked over 60% of TSMC’s 2026 advanced packaging capacity, forcing competitors AMD and Broadcom to fiercely compete for the remainder. In HBM, SK Hynix’s early entry into the NVIDIA supply chain gives it leading market share. This “Matthew effect” means firms that establish long-term supply relationships ahead of others will retain competitive advantage for years.
Capacity shortages ultimately manifest as delivery delays for AI servers and rising compute costs. Data center GPUs now have lead times of 6-12 months; high-end GPUs like B200 are almost “invisible” in China, and prices for mid/low-end GPUs like 5090, 4090 surge for reasons including raw material costs.
More importantly, downstream firms are preparing for long-term supply tension. Microsoft and SK Hynix are close to completing DDR5 long-term contract negotiations—total contract valued at trillions of won. Samsung is bundling advanced packaging with DRAM and foundry services, pushing “locked capacity” to new heights of integration.
We divide this round of AI production bottleneck-driven industry transformation into four phases:
Phase I (2024-2025): Initial supply-demand mismatch. AI compute demand begins to surge but upstream capacity expansion lags; supply-demand mismatch emerges. CoWoS and HBM are first to show supply tension. TSMC starts expansion, but distant water can’t solve immediate thirst.
Phase II (2025-2026): Acceleration of capacity shortage. All major AI accelerators migrate simultaneously to advanced nodes, all eight production capacities feel the strain. Downstream customers start locking capacity with long-term contracts and prepayments—fundamental change in contract models. We are currently at the core acceleration stage.
Phase III (2026-2028): Capacity release and structure consolidation. TSMC AP7/AP8, SK Hynix P&T7, etc., begin to ramp up new capacity, but AI demand grows in parallel. Industry leaders with first-mover advantage and scale consolidate market position, bargaining power keeps shifting upstream.
Phase IV (beyond 2028): Supply chain maturity and reshaping. Global advanced packaging and HBM capacity form a new pattern of “Asian manufacturing + global layout.” Geographic diversification is initially complete, but high-end segments remain dominated by a few oligopolies.
Key conclusions—
Conclusion 1: The essence of AI compute capacity bottlenecks is not cyclical supply-demand volatility, but structural mismatch between technology generation transitions and capacity building pace. 2026-2028 is the "settling years" for the AI supply chain; companies locking in capacity first will have a competitive advantage for years.
Conclusion 2: Of the eight bottlenecks, CoWoS and HBM are the "bottlenecks among bottlenecks." CoWoS capacity grows fast but still lags demand—NVIDIA locks over 60% capacity; HBM’s 2026 capacity is fully sold out, with Microsoft and Google making 30% prepayments for 3-5 year contracts. The tension in these two segments far exceeds market expectations and will persist through 2027 and beyond.
Conclusion 3: ABF substrates and high-end CCL are upstream material links with greatest elasticity under tight supply. The T-Glass supply gap is 25%, with lead times extended to over 30 weeks; new capacity won’t come online until end-2026. US analysts expect ABF shortage rates will reach 10% in H2 2026 and expand to 42% by 2028. CCL weekly price surges reach up to 10%-20%; upgrades to M9/M10 materials are very certain.
Conclusion 4: Domestic substitution is seeing a historic window in high-end CCL, ultra-thin electronic cloth, and packaging substrate links. In high-end CCL, Kingboard, Shengyi Technology, Nan Ya New Material, and other local companies are catching up rapidly; in ABF substrate, Xing Sen Technology, Shenzhen Circuit, etc., have made partial breakthroughs. In the current “shortage wave,” companies with high-end product capacity and customer certification are expected to enter the supply chain quickly.
Risk Warning & DisclaimerThe market has risks, investments need caution. This article does not constitute individual investment advice and does not take into account a user's specific investment objectives, financial situation, or needs. Users should consider whether any opinions, viewpoints, or conclusions are suitable for their situation. Investing based on this is at your own risk. ```