Filling the Gap in High-End PCB Manufacturing to Accelerate the Growth of the AI Electronics Industry

In the digital economy era, AI and big data technologies are reshaping the global industrial landscape. According to a McKinsey report, the global AI market reached a scale of USD 1.5 trillion in 2024, with a compound annual growth rate (CAGR) of over 25%. Meanwhile, the International Data Corporation (IDC) predicts that global spending on big data analytics will exceed USD 300 billion within the next two years, driving exponential growth in the demand for data center servers.

At the heart of all this lies the server motherboard—regarded as the "brain" of AI systems—making the PCB substrates that support AI computing the lifeline of the electronics industry. These motherboards must host high-speed chips such as CPUs and GPUs and ensure real-time processing of massive data volumes. However, the design and manufacturing of such PCBs come with extremely high technical barriers: they typically require 18+ layers, 4–8 step HDI (High Density Interconnect) structures, and must use high-frequency, high-speed base materials. Unfortunately, manufacturers capable of producing these advanced PCBs remain scarce, particularly in the field of prototyping and small-batch production—severely constraining the pace of innovation in the AI era.

As a leading company with a decade of experience in HDI PCB manufacturing, PCBWay has been focusing on high reliability and high-precision production since 2014. We have continuously advanced our technology in small-batch HDI PCB manufacturing, gradually bridging multiple technical gaps in complex HDI PCB processes.

This article will begin with an overview of the current AI boom, and go on to explore in depth the characteristics of server PCB substrates, existing supply bottlenecks, and feasible solutions—calling on the industry to join forces in driving the transformation of the AI electronics sector.

Industry Landscape and Demand Data

Artificial intelligence and big data technologies have now entered the stage of large-scale commercial deployment. According to IDC, the global AI server market reached USD 30.7 billion in 2023, marking a year-over-year growth of 35.6%. This figure is expected to surpass USD 90 billion by 2028, with a compound annual growth rate (CAGR) of 24%. Among these, cluster servers used for training large AI models place particularly demanding requirements on hardware—each GPU node requires an ultra-high-performance motherboard. For example, NVIDIA’s DGX SuperPOD system utilizes 8-step HDI PCB motherboards to support data throughput at the petabyte-per-second level.

Big data analytics is also fueling demand. According to global data platform Statista, the total volume of global data reached 120 ZB (1 ZB = 1 billion TB) in 2023 and is projected to grow to 180 ZB by 2025. Cloud computing giants like Amazon AWS and Microsoft Azure are deploying servers by the millions. Google alone added over 1 million new servers to its data centers in 2023, a 28% increase from the previous year. These server motherboards must support high-frequency computing (such as 128-core CPUs operating above 5GHz) and low-latency data transmission, making 18-layer and above PCBs a standard configuration.

According to a report by industry research firm Prismark, the high-end server PCB market reached USD 18 billion in 2023, with PCBs of 18–24 layers accounting for over 60% of that total. Demand is projected to grow to USD 25 billion by 2025, with an annual growth rate exceeding 20%.

Detailed Demand Data:

• Layer Count Requirements: The average number of layers in AI server motherboards has increased from 14 layers in 2019 to 18–22 layers in 2024. Dell’s PowerEdge series motherboards, for example, now feature 20-layer designs to support AI inference workloads.
• High-Frequency, High-Speed Requirements: Signal frequencies have risen to over 112 Gbps (e.g., PCIe 5.0 standard), requiring base materials with a dielectric constant (Dk) ≤3.5 and a dissipation factor (Df) ≤0.002.
• Market Gap: In 2024, the global supply gap for such boards has reached 18%. For small-batch orders (fewer than 500 pieces), the delivery delay rate is as high as 60%, significantly hindering time-to-market for server products.

This situation stems from the “acceleration law of innovation” in the AI industry: the cycle from prototyping to mass production has shortened to just 3–6 months. However, the PCB substrates required to support this rapid pace have become a critical bottleneck. Only high-layer-count, high-frequency, high-speed, multi-step, and high-reliability HDI PCBs can meet these product demands—but their manufacturing process is often regarded as the "Mount Everest" of small-batch PCB fabrication.

In the following sections, we will examine the core characteristics of these advanced PCBs in greater detail.

Features and Manufacturing Insights of Server-Grade PCB Substrates

PCBs designed for AI-based large-scale servers represent the pinnacle of electronic manufacturing. Their key features can be summarized as the “Five Highs”:

• High Layer Count (more than 18 layers)
• High-Frequency Signals (GHz-level transmission)
• High-Speed Data (Gbps-level rates)
• High Step Count (4–8 step HDI structures)
• High Density (line width/spacing < 50μm)

Take a typical 8-step HDI server motherboard as an example: it involves a 20-layer copper stack, with millions of microvias interconnecting layers to support 128 high-density BGA chips, all while maintaining an overall board thickness of under 2 mm.

In terms of materials, high-frequency, high-speed substrates dominate the market—such as Rogers RO4350B, Panasonic R5785N—with dielectric constants (Dk) required to remain stable within 3.4 ± 0.05.

Key Product Characteristics:

• High Layer Count Structure (18–30 layers): Multilayer stacking ensures power integrity. For example, a 24-layer PCB typically integrates 12 signal layers, 8 power layers, and 4 ground layers to minimize noise interference. A critical technical challenge lies in interlayer alignment accuracy (tolerance < 5 μm); a deviation exceeding 10 μm can increase signal reflection by 15%, potentially causing errors in AI computations.
• High-Frequency, High-Speed Performance: Frequencies have increased to 5–10 GHz to support DDR5 memory (with bandwidth up to 6.4 Gbps). The challenge lies in dielectric control: materials must maintain a Df < 0.002 at 10 GHz, or else signal loss becomes significant (e.g., losses > 3 dB/inch can reduce transmission distance by half). High-speed transmission also depends on impedance consistency (100Ω ± 5%), requiring precise calculation of differential pair trace widths (±2 μm tolerance) to avoid increased bit error rates in data centers.
• High-Step HDI (4–8 steps): Higher step counts mean ultra-fine interconnect structures. A 4-step HDI requires four rounds of laser drilling with via diameters ≤ 0.1 mm; an 8-step HDI involves eight sequential lamination and laser drilling processes. At high density, there can be over 5,000 microvias per square inch.
• High-Density Interconnect (HDI): Trace width/spacing is reduced to below 50 μm/50 μm (for reference, human hair is about 70 μm thick), with pin densities reaching 200 pins/cm. This requires “zero-blind-via” technology—integrating through, blind, and buried vias across all layers—to ensure low-latency signal transmission (<10 ps/inch).

Manufacturing Challenges and Key Technical Requirements:

• Interlayer Alignment and Lamination Control: PCBs with more than 18 layers require three or more lamination cycles, making them highly susceptible to layer misalignment, which can result in short circuits.
• Microvia Technology: Laser drilling must penetrate via diameters of 2–4 mils, which often leads to issues such as "taper defects" and "pad separation," both of which negatively affect conductivity. The key lies in pulse energy control—when energy is kept below 20 μJ per pulse, the drilling success rate exceeds 98%.
• Signal Integrity: At high frequencies, the skin effect becomes more pronounced, requiring gold surface finishes (electroless nickel immersion gold combined with electroplated nickel-gold) to ensure impedance stability. The main challenge lies in achieving uniform plating, with a variation tolerance of ±0.05 μm.
• Thermal Management of Materials: High-frequency PCB materials must have a coefficient of thermal expansion (CTE) that matches copper (16 ppm/°C); otherwise, the risk of delamination during high-temperature reflow soldering increases by 40%.

Behind all these process challenges lies an enormous investment cost: setting up a single high-level HDI PCB production line requires over $7 million USD. Without years of accumulated technical expertise, achieving reliable mass production is nearly impossible.

PCBWay has addressed these challenges using a proprietary “Four-Step Manufacturing Process”—precision drilling, high-layer lamination, high-precision etching, and signal testing—which has helped raise the yield rate to over 95%.

Current Supply Status of Server-Grade PCB Substrates

The rapid pace of AI innovation has exposed significant weaknesses in the supply chain. According to Prismark's 2023 report, fewer than 10 PCB manufacturers worldwide are capable of consistently mass-producing high-end HDI server boards with more than 18 layers—and 90% of that capacity is monopolized by overseas giants such as Samsung and TTM. Most remaining manufacturers focus on large-volume orders (10,000 pieces or more), leaving the prototype and small-batch sector (1–500 pieces) in a supply vacuum: the supply-demand gap has reached 30%, and 80% of AI startups report delays of 7 to 9 weeks in motherboard delivery.

Shortage data is as follows:

• Imbalanced Capacity Distribution: In 2024, the average lead time for samples and small-batch production of high-end HDI PCBs has extended to 8 weeks, while the standard lead time should be 2–4 weeks.
• Cost and Economic Barriers: Small-batch trial production requires production line reconfiguration, increasing costs by up to 60%. Most small and medium-sized manufacturers are discouraged by the high capital investment required.
• Technical Thresholds: 40% of PCB factories are unable to manufacture server boards due to insufficient capabilities in lamination control and microvia technology.
• Regional Concentration Issues: 70% of production capacity is concentrated in Taiwan and South Korea (e.g., Samsung, Unimicron), while the absence of major capacity in Europe and North America leads to tariff risks. Amid geopolitical tensions from 2023 to 2024, both import lead times and costs for server PCB prototypes and small batches have risen by over 30%.

The root of this shortage lies in the so-called “Three High Paradox”:

• High Capital Investment (over USD 7 million per production line)
• High Risk (prototype failure rate up to 20%)
• High Technical Barriers (due to the “Five High” characteristics)

PCBWay’s Achievements

Focusing on the niche field of server motherboards, PCBWay has established four core technical pillars: fine trace technology, interlayer alignment, microvia fabrication, and signal integrity control.

In 2023, PCBWay successfully broke through the engineering prototype process for 8-step HDI PCBs. In 2025, we achieve mass production of 8-step HDI boards, with 12-step HDI technology already under preliminary research for 2028.

Below are some sample server-grade PCBs developed by PCBWay:

Artificial intelligence and big data technologies are reshaping human civilization, with the foundation being the signal transmission of every server motherboard. Starting from the current AI boom, we have analyzed the peak characteristics of the products, the manufacturing pain points, and revealed the harsh reality of supply shortages. We also showcased PCBWay’s breakthroughs in the small-batch production sector.

For every 1% growth in the AI server market, millions of high-performance PCBs are required. PCBWay promises to continue filling this gap and accelerating innovation by using technology-driven solutions. At the same time, we invite global partners to join us on this journey, turning every engineering prototype and small-batch server motherboard into a spark that drives the intelligent world forward.