After GPUs and ASICs, CPUs are now vying for production capacity: Why has ABF become a new bottleneck for AI?

After GPUs and ASICs, CPUs are now vying for production capacity: Why has ABF become a new bottleneck for AI?

GPUs and ASICs have been consuming high-end ABF substrate capacity for the past two years, and now CPUs have become the new demand variable. Agentic AI is driving a rapid upward revision of server CPU demand, while the increase in AI chip size, number of layers, and decline in yield are continuously reducing "effective capacity".

More noteworthy is that the delivery time for high-end lamination equipment orders has been extended to 2031, and the expansion of substrate production is subject to multiple constraints, including equipment, materials, certification, and yield.

When three types of high-performance chips begin to compete for the same batch of high-end production capacity, what new opportunities will emerge in the ABF industry chain?

I. What happened? — The CPU became the third demand curve for ABF.

1. In addition to GPUs and ASICs, CPUs have suddenly joined the competition for production capacity:

Over the past two years, the market's understanding of the ABF substrate market has been driven by two main variables: the continuous upgrades of NVIDIA GPUs and the accelerated production of ASICs such as Google TPUs and AWS Trainium. These two types of chips have jointly driven the expansion of advanced package sizes and the increase in the number of layers, leading to a gradual tightening of high-end ABF production capacity.

Now a third variable has emerged—CPU.

On September 21, US CPU supply chain stocks collectively surged, with AMD's share price rising by about 10%, and Intel and Arm also showing strength. The market is reassessing the demand for CPUs from AI inference and agents. Intel CEO Chen Liwu recently stated that the company can currently only meet about 50% of its customers' CPU needs; Intel's data center and AI business revenue grew by 59% year-on-year in the second quarter of this year, and the Xeon 6 became one of the fastest-growing server CPU products in the company's history.

The underlying change stems from AI workloads. In the training era, a large amount of capital expenditure was concentrated on GPUs; as we move into the inference and agent stages, CPUs take on tasks such as task scheduling, data preprocessing, database access, network and storage control. Intel previously revealed that the CPU to GPU ratio in AI data centers may converge from approximately 1:4 to 1:8 to 1:1, and the company has already shifted some of its wafer capacity towards Xeon.

This has a direct impact on ABF: the demand for GPUs and ASICs has not weakened, and CPUs are now entering the competition for high-end substrate capacity.

Industry research estimates that the current global annual demand for ABF (Alternating Current Fabric) substrates is approximately 35 million units, of which about 18 million units are for AI-related GPUs, CPUs, and TPUs, accounting for more than half; the demand for substrates for traditional PC CPUs has remained largely unchanged in the past two years. The growth focus of the ABF market has clearly shifted from PCs to data centers.

2. The increased demand for CPUs is still driven by a batch of high-value items:

This round of CPU demand surge differs from the traditional PC replacement cycle.

The price of a single ABF carrier board for ordinary PCs is only about 40-80 yuan, while the price of a mid-range server CPU carrier board has entered the range of 20-50 US dollars; the price of a complex AI server carrier board has reached several hundred US dollars. Industry research shows that the price of the GB300 carrier board has increased from about 220 US dollars to 250-260 US dollars, Rubin is about 350 US dollars, and the Rubin Ultra pre-research product even reaches about 450 US dollars.

The price difference reflects a significant increase in area, number of layers, and processing difficulty. Taking the AMD MI450 as an example, industry research indicates that its substrate size reaches approximately 100mm × 120mm and has 24 layers, approaching the size limit of current mass production processes, which places high demands on warpage control.

Therefore, the marginal change brought about by CPUs cannot be simply understood as an increase in chip orders. GPUs, ASICs, and CPUs, the three types of high-performance logic chips, are beginning to jointly consume the scarcest portion of large-size, high-layer-number ABF (Alternating Current Components) production capacity.

This has also changed ABF's research framework: in the past, the first thing to do was to predict how many GPUs and ASICs could be sold; now, it must also answer the question of how many qualified carrier boards the industry can actually deliver when these chips are mass-produced at the same time.

II. Why is this important? —ABF's dilemma is shifting from "demand growth" to "insufficient effective production capacity."

1. AI is making the production capacity of one square meter of ABF increasingly "unreliable":

The requirements for ABF carrier boards cannot be calculated solely based on the number of chips. A more reasonable framework is:

Demand intensity ≈ Number of chips × Carrier area × Number of layers × Manufacturing complexity.

In the AI era, these variables are all moving upwards simultaneously.

Ajinomoto previously disclosed that the area index of its high-performance AI semiconductor substrate can reach about 3.5 times that of traditional products. After increasing the number of ABF layers from 6 to 18, the amount of ABF material used in a single product can reach about 10 times that of traditional products.

Circuitry is also continuing to shrink. The density (Df) of high-end products is evolving towards 0.002 or even lower, and the linewidth and spacing are continuing to shrink from 5/5μm and 3/3μm. With larger sizes and more layers, the difficulty of warping, interlayer alignment, and micro-hole fabrication increases simultaneously.

This will produce a result that has often been overlooked in the past: even if the factory's "capacity per square meter" remains unchanged, the number of high-end chips that can ultimately be delivered per square meter will decrease.

Therefore, what deserves more attention than nominal production capacity is:

Effective capacity = Nominal capacity × Proportion of high-end production that can be converted × Yield rate × Certification rate × Production efficiency.

AI is simultaneously lowering the latter few coefficients.

2. Yield is becoming the most hidden supply constraint:

Intel's EMIB-T provides a good example.

Industry research shows that Ibiden, Shinko, and others are currently prototyping EMIB-T substrates, with customers setting a yield target of approximately 57%, but the actual prototyping yield is still relatively low at this stage. Ibiden completed some prototyping in June of this year and delivered a small number of good products to customers for packaging and testing; Intel plans to enter mass production in the first quarter of 2027.

The challenge for the AMD MI450 stems from its large size of 100mm x 120mm and its 24-layer structure. The larger the area, the fewer substrates can be cut from a single panel, and any local defects, uneven thickness, or warping will be magnified, ultimately affecting the yield.

The gap between domestic and international markets is primarily reflected in yield rates. Industry research indicates that the yield rate gap between domestic PC substrate manufacturers and leading overseas companies can be controlled within 5 percentage points for low-end PC substrates, but this gap can exceed 10 percentage points for high-end AI products.

In other words, a company announcing the construction of 100,000 square meters of production capacity does not mean that the market will immediately have 100,000 square meters of usable Rubin or next-generation CPU carrier boards.

3. Expansion is already stalled due to equipment issues:

More advanced signals come from the device itself.

A recent Goldman Sachs survey of Eternal Precision, a major global supplier of ABF (Absorbent Polymer) vacuum laminators, revealed that lead times for high-end equipment orders have extended from 2028-2029 to 2031, with new orders primarily coming from high-end 90-ton three-stage vacuum laminators. Some customers have even proactively offered to pay a premium to shorten delivery times, a practice uncommon during the previous ABF upswing cycle of 2020-2022.

Goldman Sachs' research further estimates that the demand for mid-to-high-end equipment will be approximately 1,500 units from 2027 to 2031, while the cumulative shipments from 2022 to 2025 will be approximately 600 units. This figure of 1,500 units is a projected value based on equipment manufacturers' capacity and industry expansion plans, and cannot be equated with locked-in orders. However, the equipment scheduling itself is enough to show that substrate manufacturers are vying for future capacity access in advance.

Domestic industry research also confirms the equipment constraints: core equipment such as laser drilling and high-resolution exposure still heavily relies on imports from Japan, and the delivery time for new equipment can exceed two years. This forms a clear timeline:

Ordering equipment → Installation and commissioning → Membrane material and chemical compatibility → Customer certification → Yield ramp-up → Formation of effective production capacity.

The fact that equipment orders are booked until 2031 does not mean that ABF substrates are in short supply until 2031, but it does mean that the supply side is unlikely to expand rapidly within a year like ordinary PCBs.

4. Having equipment is not enough; certification further delays the supply:

High-end electronic materials share a common characteristic: the construction of the equipment is only the starting point.

Morgan Stanley's research on high-end CCLs points out that materials of M7 and above typically require 12-18 months of certification to enter the customer base of NVIDIA, AWS, Google, Meta, and other companies. Moreover, the certification is specific to the material, production line, and factory location. The final product must also pass reliability tests on the actual PCB structure after drilling, lamination, and electroplating.

Customer certification for ABF substrates is more complex. As product size and number of layers increase, material properties, equipment precision, and manufacturer process capabilities must all meet standards simultaneously. This explains why the industry can simultaneously exhibit two seemingly contradictory phenomena: all manufacturers are expanding production, yet high-end substrates remain in short supply.

Goldman Sachs estimates, based on current demand and capacity expansion assumptions, that the ABF substrate shortage rate may widen from approximately 14% in the second half of 2026 to 34% in 2027 and 51% in 2028. These figures are highly dependent on AI chip shipments, capacity expansion progress, and yield assumptions, and are more suitable to be viewed as stress scenarios rather than actual industry statistics; however, the direction is clear—new nominal capacity may not be able to keep up with effective demand.

5. This gives the price increase a more solid foundation:

The future price elasticity of ABF substrates will mainly come from two parts.

One reason is the upgrade in product structure: larger size and higher layer count products consume more materials, take longer to occupies equipment, and have lower yield rates, so their prices are much higher than those of PC substrates.

Another part comes from the premium for scarce capacity: when leading customers lock in equipment and capacity in advance, peripheral customers who want to obtain additional supply need to accept higher prices or longer delivery times.

However, this does not mean that all customers and all substrates will raise prices simultaneously. According to Ibiden's industry research, it is still difficult for substrate manufacturers to unilaterally raise prices for large customers who have access to supply chain cost information. Significant price increases due to supply and demand tensions are more likely to occur first for smaller customers, new projects, and high-difficulty products.

Therefore, this round of ABF is closer to a structural price increase for high-end products: the more complex the demand, the fewer suppliers can stably produce in large quantities, and the more likely prices and profits are to concentrate on high-end effective production capacity.

Supply and demand analyses all point to the same conclusion: the future value distribution of the ABF industry chain will not simply depend on who expands production the most, but on who can ultimately transform equipment, materials, and processes into qualified products. For the Chinese supply chain, this also opens up a more complete path to localization—extending from carriers and ABF films to manufacturing equipment and key materials.

1. Carrier board: Directly supports the triple needs of GPU, ASIC, and CPU.

Domestic ABF substrates are still in the catching-up stage, but this is the most direct link in the transmission of demand.

Overseas high-end production capacity is mainly concentrated in companies such as Ibiden, Unimicron, Nanya, Kinsus, Samsung Electro-Mechanics, and AT&S. Domestic manufacturers entered the market relatively late, with companies like Shennan Circuits and Xingsen Technology starting large-scale deployments around 2021, and some products currently in small-batch production. Industry research shows that domestic manufacturers (such as Shennan Circuits and Xingsen Technology) are currently focusing on developing their own computing power supply chain as their main breakthrough point, as verifying top-tier overseas AI products is quite difficult.

The opportunities for domestically produced carrier boards therefore include two layers: the domestic demand brought about by domestic CPUs, GPUs, and ASICs themselves, and the spillover of mid-range orders to second-tier suppliers after the leading overseas high-end production capacity is simultaneously contested by GPUs, ASICs, and CPUs.

2. ABF membrane: The material with the lowest market share but the greatest flexibility in domestic production.

ABF film remains one of the single materials with the greatest potential for domestic substitution. Ajinomoto holds over 95% market share in the additive coating film market for GPU and CPU substrates, while domestic manufacturers have gradually moved from formula development to customer verification and small-batch supply.

Industry research indicates that some of Huazheng's products, comparable to GL102, have entered the small-batch stage, while the higher-end GX series is still undergoing validation. Whether domestically produced films can enter the low-df, large-size, and high-layer substrate markets is key to assessing the industry's value. In addition, Lotus Holdings (which holds shares in Newface), Hongchang Electronics (with an integrated resin and film material layout), and Shengyi Technology (which already has orders for advanced encapsulation films) also have related investments.

3. Equipment: New Expectations for Domestic Substitution

Another constraint on ABF substrate capacity expansion comes from the equipment side. High-end FCBGA manufacturing requires repeated processes such as lamination, laser drilling, exposure, electroplating, and testing. Large-size, high-layer-count products have significantly higher requirements for equipment precision, stability, and cycle time than ordinary PCBs. Currently, many high-end equipment suppliers still rely heavily on overseas suppliers, and the long lead times for processes like laser drilling and high-resolution exposure make it difficult for substrate manufacturers to quickly convert new capacity into effective supply.

Domestic equipment manufacturers have already achieved breakthroughs in several key processes. Dazhu CNC has received formal orders for drilling and high-precision grooving equipment for BT and ABF materials, and is advancing solutions for smaller hole diameters and ABF thinning and resist removal for large-size FCBGAs. Xinge Microelectronics' high-precision direct-write lithography equipment has entered the ABF substrate manufacturing sector. Dongwei Technology's transfer-type VCP and horizontal electroplating equipment also cover ABF substrate applications. In comparison, the localization of vacuum lamination and some high-end testing processes is still in an earlier stage.

The industrial significance of domestically produced equipment goes beyond simply replacing imported equipment. The effective production capacity of ABF substrates depends on equipment precision, yield, process compatibility, and customer certification; any limitation in any of these will slow down capacity release. With overseas equipment delivery times continuing to lengthen, domestic equipment manufacturers that can be the first to complete verification by high-end customers have the opportunity to benefit simultaneously from both substrate capacity expansion and supply chain localization.

4. Key materials such as fillers and electronic fabrics: Sharing the benefits of specification upgrades

ABF membranes are composed of resin, curing agent, and inorganic fillers. With decreasing density (Df) and miniaturization of circuits, high-end products require fine fillers with an average particle size ≤0.5μm and improved dispersion uniformity; spherical silica powder is gradually evolving from a common functional filler to a performance-critical material, and companies such as Lianrui New Materials and Lingwei Technology have already made relevant investments.

Furthermore, the requirements for low CTE electronic cloth, BT core, high-end copper foil, and wet electronic chemicals are also increasing for large-size substrates. It should be noted that the localization flexibility in these areas may not be as great as that in film materials, but the advantage is that it does not require betting on a particular substrate manufacturer or film manufacturer to ultimately succeed.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.