Technical Deep Dive
The shift from conventional PCB to AI-grade HDI boards is a fundamental change in manufacturing physics. Standard PCBs for consumer electronics or automotive applications typically use 4-8 layers with line widths/spacing of 75-100 microns. AI server HDI boards, by contrast, require 16-30+ layers, line widths down to 30-40 microns, and advanced materials like low-loss laminates (e.g., Panasonic Megtron 6, Isola I-Tera) to handle signals at 112 Gbps PAM4 and beyond.
Key Technical Parameters for AI HDI Boards:
| Parameter | Standard PCB (NEV/Consumer) | AI Server HDI Board |
|---|---|---|
| Layer Count | 4-8 | 16-32+ |
| Minimum Line Width/Space | 75-100 µm | 30-40 µm |
| Aspect Ratio (via) | 8:1 | 12:1 or higher |
| Material | FR-4 (standard) | Low-loss, high-Tg (e.g., Megtron 6) |
| Signal Integrity | Up to 25 Gbps | 112 Gbps PAM4 |
| Thermal Management | Basic copper planes | Embedded thermal vias, metal cores |
| Yield Rate | >95% | 70-85% (at scale) |
Data Takeaway: The jump from standard to AI-grade HDI is not incremental—it requires a complete retooling of lamination, drilling, plating, and inspection processes. The lower yield rates for high-layer-count boards mean that only manufacturers with mature process control can achieve profitability.
Engineering Challenges:
- Stack-up Design: Managing impedance control across 30+ layers with mixed materials is extremely complex. Any mismatch causes signal reflection and data errors.
- Via Formation: High aspect ratio vias (12:1 or more) require advanced laser drilling and uniform copper plating to avoid voids. Unimicron uses a proprietary 'any-layer' HDI process that allows vias to connect any two layers, reducing signal path length.
- Registration Accuracy: Layer-to-layer alignment must be within ±15 microns. This demands precision lamination presses and automated optical inspection (AOI) systems.
- Thermal Dissipation: A single NVIDIA H100 GPU can draw 700W. The HDI board must efficiently conduct heat away from the chip to prevent throttling. Embedded copper coin or vapor chamber integration is becoming common.
Relevant Open-Source Tools:
While PCB design is dominated by proprietary EDA tools (Cadence Allegro, Altium), there are notable open-source projects for simulation and verification:
- OpenEMS (GitHub: ~1.5k stars): A full-wave electromagnetic field solver used for signal integrity analysis of high-speed PCB traces.
- KiCad (GitHub: ~18k stars): While not AI-specific, its latest 8.0 release added improved differential pair routing and impedance calculator plugins, making it viable for prototyping AI accelerator boards.
- PySPICE (GitHub: ~500 stars): A Python wrapper for SPICE simulations, useful for modeling power delivery networks in AI server PCBs.
Key Players & Case Studies
The global HDI board market is dominated by Taiwanese and Japanese firms, but Chinese companies are aggressively investing to close the gap.
Competitive Landscape for AI-Grade HDI:
| Company | HQ | AI HDI Capability | Key Customers | Recent Investment |
|---|---|---|---|---|
| Unimicron | Taiwan | 32-layer any-layer HDI, 30µm L/S | NVIDIA, AMD, Intel | $2.5B (2023-2025 capacity expansion) |
| Ibiden | Japan | 30-layer HDI, IC substrates | NVIDIA, Broadcom | $1.8B (new plant in Gifu) |
| AT&S | Austria | 28-layer HDI, embedded components | AMD, Xilinx | €1.2B (Leoben facility upgrade) |
| Shennan Circuits | China | 24-layer HDI, 40µm L/S | Huawei, Inspur | ¥5B (2024 bond for AI PCB) |
| Mingyang Circuit | China | Currently 16-layer max, targeting 24-layer | Tier-2 server OEMs | ¥1.2B (this bond) |
Data Takeaway: Mingyang is entering a market where the top three players already have multi-year head starts in process maturity and customer relationships. Its competitive advantage lies in cost—Chinese manufacturers can offer 15-20% lower prices than Taiwanese/Japanese peers—but only if it can achieve acceptable yield rates.
Case Study: Shennan Circuits
Shennan Circuits (SCC) is the most direct comparable. In 2023, SCC raised ¥5 billion via convertible bonds to build a dedicated AI HDI plant in Wuxi. By Q1 2025, SCC reported that its AI server PCB revenue had grown 300% year-over-year to ¥2.8 billion, with gross margins of 38% compared to 22% for its traditional automotive PCB business. This validates the economic thesis behind Mingyang's pivot.
Case Study: Unimicron's Dominance
Unimicron, the world's largest HDI board maker, supplies NVIDIA's DGX and HGX server platforms. Its 'any-layer' HDI technology uses sequential lamination to create microvias that can connect any two layers, enabling the dense routing required for GPU-to-GPU interconnects (NVLink). Unimicron's gross margins exceed 45%, and it has maintained a 95%+ customer retention rate. The barrier to replicating this technology is immense: Unimicron holds over 200 patents related to HDI manufacturing processes.
Industry Impact & Market Dynamics
The PCB industry is undergoing its most significant structural shift since the smartphone boom of the 2010s. AI server demand is projected to grow at a CAGR of 45% through 2028, far outpacing the 4% CAGR for traditional PCB markets.
Market Size Projections (USD Billions):
| Segment | 2024 | 2026E | 2028E | CAGR |
|---|---|---|---|---|
| AI Server PCB (HDI + IC Substrates) | $8.2 | $17.5 | $32.1 | 45% |
| Automotive PCB (incl. NEV) | $12.1 | $13.8 | $15.2 | 6% |
| Consumer Electronics PCB | $18.5 | $19.2 | $20.1 | 2% |
| Total PCB Market | $78.0 | $92.5 | $108.0 | 8% |
*Source: Prismark Partners, AINews estimates*
Data Takeaway: AI server PCB is the fastest-growing segment by a wide margin. By 2028, it will account for nearly 30% of the total PCB market, up from 10% in 2024. This explains why every major PCB manufacturer is pivoting.
Business Model Implications:
- From Commodity to Specialty: Traditional PCB makers operate on thin margins (10-15%) and compete on price. AI HDI boards command 35-50% gross margins, but require heavy upfront R&D and capital expenditure.
- Customer Concentration Risk: AI server PCB is a winner-take-most market. NVIDIA alone accounts for 70-80% of AI GPU shipments, meaning suppliers must pass rigorous qualification processes. Losing a single customer can devastate revenue.
- Supply Chain Localization: Chinese AI companies (Huawei, Baidu, ByteDance) are increasingly sourcing PCBs domestically to reduce geopolitical risk. This creates a tailwind for Mingyang, but also invites competition from other Chinese firms like SCC and WUS Printed Circuit.
Risks, Limitations & Open Questions
1. Execution Risk: Building a 24-layer HDI production line is not simply buying new machines. It requires months of process tuning, trial runs, and customer qualification. Mingyang's current maximum is 16 layers; jumping to 24 layers in 18 months is aggressive. Any delays in yield improvement could erode investor confidence.
2. Technology Obsolescence: The industry is already moving toward even more advanced substrates. IC substrates (used for chip packaging) offer even higher density and margins. If Mingyang invests heavily in HDI just as the market shifts to IC substrates, it could be left behind. AT&S and Unimicron are already building IC substrate capacity for NVIDIA's next-generation GPUs.
3. Customer Diversification: Mingyang's current customer base is heavily weighted toward domestic server OEMs (Inspur, Lenovo, H3C). To achieve scale, it needs to qualify with NVIDIA or AMD directly—a process that can take 12-24 months and requires passing rigorous reliability tests (e.g., IPC-6012 Class 3).
4. Geopolitical Headwinds: US export controls on advanced chips and manufacturing equipment could limit Mingyang's ability to purchase the high-end laser drilling and lamination equipment needed for 30µm line widths. Japanese and German equipment makers (Mitsubishi, Schmoll) may face licensing delays.
5. Overcapacity Risk: The current AI PCB shortage is driving massive investment. If every major Chinese PCB maker builds new HDI capacity simultaneously, the market could become oversupplied by 2027, compressing margins. A similar cycle occurred in the smartphone PCB market in 2018-2020.
AINews Verdict & Predictions
Mingyang's pivot is strategically sound but execution-dependent. The company is betting that it can replicate Shennan Circuits' success, but SCC had a two-year head start and a stronger balance sheet.
Our Predictions:
1. Short-term (2025-2026): Mingyang will successfully build its 24-layer HDI pilot line by Q2 2026, but initial yields will be below 70%, leading to a 12-18 month period of negative gross margins on AI-related revenue. The stock will be volatile.
2. Medium-term (2027-2028): If Mingyang can achieve 85% yield on 24-layer boards, it will qualify with at least one major AI server OEM (likely Inspur or Huawei). Revenue from AI PCBs will reach ¥2-3 billion by 2028, representing 30-40% of total revenue.
3. Long-term (2029+): The real prize is IC substrates. We predict Mingyang will announce a follow-on investment in IC substrate R&D by 2028, using cash flow from HDI to fund the next leap. Companies that fail to make this transition will be acquired or marginalized.
What to Watch:
- Yield Reports: Mingyang's quarterly disclosure of HDI board yield rates is the single most important metric.
- Customer Announcements: Any mention of qualification with NVIDIA, AMD, or Huawei will be a major catalyst.
- Equipment Orders: Tracking purchases of laser drilling machines from Mitsubishi or Schmoll will signal capacity buildout pace.
Final Editorial Judgment: Mingyang's pivot is a high-risk, high-reward bet that reflects the broader Chinese manufacturing strategy of moving up the value chain. If it succeeds, it will become a case study in industrial transformation. If it fails, it will be a cautionary tale about the dangers of chasing hot markets without sufficient technological foundation. We lean cautiously optimistic—the management team has a track record of disciplined capital allocation, and the AI PCB tailwind is real. But investors should buckle up for volatility.