Technical Deep Dive
The PCB is the unsung hero of modern electronics, a complex laminate of copper and insulating materials that provides the electrical connections between all components. For AI accelerators, the demands are extreme. Nvidia's H100 and B200 GPUs, for instance, require PCBs with 20 to 30 or more layers, featuring High-Density Interconnect (HDI) technology with microvias as small as 50-75 microns in diameter. These boards must manage immense power delivery (often exceeding 700W for a single GPU) and signal integrity at speeds approaching 112 Gbps for PCIe Gen 5 and NVLink interconnects. The manufacturing process involves sequential lamination, laser drilling, precise copper plating, and stringent impedance control—a process that is as much a materials science challenge as it is a manufacturing one.
Chinese manufacturers like Shennan Circuits (深南电路) and WUS Printed Circuit (沪电股份) have invested billions in state-of-the-art facilities to master these processes. Shennan, for example, has been a key supplier for Nvidia’s high-end server-grade PCBs. The company’s technical capability is evidenced by its ability to produce boards with aspect ratios (board thickness to via diameter) exceeding 20:1, a critical metric for signal integrity in high-speed designs. Open-source repositories like the PCB-Design-Resources collection on GitHub (with over 5,000 stars) provide community-driven guides on HDI design rules, but the actual fabrication tolerances are a closely guarded secret of these manufacturers.
Data Table: PCB Complexity Comparison for AI Accelerators
| Parameter | Consumer GPU (e.g., RTX 4090) | Enterprise AI GPU (e.g., Nvidia H100) | Next-Gen (e.g., Nvidia B200) |
|---|---|---|---|
| Layer Count | 12-16 | 20-24 | 28-32 |
| Minimum Trace/Space | 3/3 mil | 2.5/2.5 mil | 2/2 mil |
| Via Type | Through-hole + Blind | HDI (Microvia) | HDI + Any-layer |
| Max Data Rate | PCIe Gen 4 (16 GT/s) | PCIe Gen 5 (32 GT/s) | PCIe Gen 6 (64 GT/s) |
| Power Delivery | ~450W | 700W | >1000W |
| Material | Standard FR-4 | High-Tg, Low-Loss | Ultra Low-Loss (e.g., MEGTRON6) |
Data Takeaway: The jump from consumer to enterprise AI GPUs represents a step-change in PCB complexity, not just an incremental improvement. The move to any-layer HDI and ultra-low-loss materials for next-gen accelerators creates a massive barrier to entry for non-Chinese manufacturers, who have not invested at the same scale.
Key Players & Case Studies
The PCB supply chain for AI is dominated by a few key players. On the design side, Nvidia specifies the board stack-up and routing constraints. The actual fabrication is then outsourced to a select group of manufacturers. The primary Chinese players are:
- Shennan Circuits (深南电路): The clear leader for Nvidia’s high-end server PCBs. The company has a dedicated R&D center for high-speed digital boards and has been a supplier for multiple generations of Nvidia DGX systems. Its revenue from PCB business exceeded $2.5 billion in 2024, with a significant portion attributed to AI-related products.
- WUS Printed Circuit (沪电股份): A major supplier for networking and server infrastructure, including backplanes for AI clusters. They have a strong focus on high-layer-count boards (up to 40 layers) and have invested heavily in new plants in Huangshi, Hubei province.
- Unimicron (欣兴电子): A Taiwan-based manufacturer that is also a key player, but its capacity is increasingly strained by demand from both Nvidia and AMD. Unimicron’s focus is on advanced IC substrates, which are a step above PCBs but share similar manufacturing challenges.
On the non-Chinese side, TTM Technologies (US) and AT&S (Austria) have capabilities, but their capacity for the specific high-volume, high-mix needs of AI accelerators is significantly smaller. AT&S has invested in a new plant in Malaysia, but it is years away from reaching the scale of Chinese competitors.
Data Table: Key AI PCB Manufacturers – Capacity and Capability
| Manufacturer | Country | Est. AI PCB Revenue (2024) | Max Layer Count | HDI Capability | Key AI Customer |
|---|---|---|---|---|---|
| Shennan Circuits | China | $1.5B+ | 30+ | Yes (Any-layer) | Nvidia |
| WUS Printed Circuit | China | $800M+ | 40+ | Yes (HDI) | Nvidia, Cisco |
| Unimicron | Taiwan | $1.2B+ (incl. IC substrates) | 20+ | Yes (HDI) | Nvidia, AMD |
| TTM Technologies | USA | $300M (est.) | 20+ | Limited | AMD, Intel |
| AT&S | Austria | $200M (est.) | 16+ | Yes (HDI) | Nvidia (limited) |
Data Takeaway: Chinese manufacturers hold a commanding lead in both revenue and technical capability for AI-specific PCBs. The gap is not just in cost but in the ability to deliver the most complex boards at scale, a capability that has been honed over a decade of serving the global electronics industry.
Industry Impact & Market Dynamics
The concentration of AI PCB manufacturing in China creates a multi-layered risk. The most immediate is the geopolitical supply chain risk. If the US were to impose export controls on AI PCBs similar to those on advanced chips, the entire AI hardware pipeline would grind to a halt. Nvidia’s lead times for its DGX systems would stretch from months to potentially over a year, as alternative suppliers lack the capacity to fill the gap. This is not a hypothetical; the US Department of Commerce has already signaled interest in broadening the scope of controls.
Second, there is the hardware security risk. A malicious actor could theoretically modify a PCB design to introduce a hardware trojan. For example, a subtle change in a trace length could create a timing side-channel that leaks cryptographic keys. More sophisticated attacks could involve embedding a tiny, passive circuit that acts as a backdoor. Detection is extremely difficult because it requires physical inspection of every layer, which is destructive and time-consuming. The industry standard for trust is the Trusted Foundry program, but this only covers chip fabrication, not PCB assembly.
The market dynamics are also shifting. The global PCB market was valued at approximately $80 billion in 2024, with the AI segment growing at over 25% CAGR. Chinese manufacturers now account for over 50% of global PCB output, and their share of the high-end AI segment is likely higher. This dominance is self-reinforcing: more volume leads to more investment in R&D and capacity, which widens the gap with competitors.
Risks, Limitations & Open Questions
- Detection of Hardware Trojans: Current non-destructive testing methods (e.g., X-ray, automated optical inspection) cannot reliably detect malicious modifications at the inner layers of a 30-layer board. The only sure method is destructive cross-sectioning, which is impractical for high-volume production. This creates a fundamental trust problem.
- Geopolitical Escalation: The risk of a sudden supply cut-off is real. If China were to restrict exports of AI PCBs in retaliation for chip controls, the impact would be immediate and severe. The US has no domestic PCB industry capable of filling the void.
- Technical Limitations of Alternatives: Non-Chinese manufacturers are investing, but they face a steep learning curve. AT&S’s new plant in Malaysia is targeting 16-layer HDI boards, while Shennan is already shipping 30-layer any-layer boards. The technology gap is 2-3 years, and the capacity gap is even larger.
- Open Question: Can the US rebuild PCB manufacturing? The CHIPS Act has focused on semiconductor fabrication, but there is no equivalent for PCB manufacturing. The cost of building a state-of-the-art PCB fab is in the billions, and the skilled workforce is largely in Asia. The question is whether the US government will treat this as a national security priority.
AINews Verdict & Predictions
The PCB is the forgotten layer of the AI hardware stack, but it is becoming a critical vulnerability. The industry’s focus on algorithmic safety and data privacy has ignored the physical substrate on which all computation runs. Our analysis leads to several clear predictions:
1. Within 18 months, the US government will initiate a formal review of the AI PCB supply chain, likely leading to new guidelines or restrictions on sourcing for defense and intelligence applications. This will mirror the actions taken on semiconductor equipment.
2. Nvidia will be forced to dual-source its most advanced PCBs, but will find that non-Chinese alternatives are 2-3 years behind in capability and 30-50% more expensive. This will create a cost and performance penalty for non-Chinese AI clusters.
3. The concept of 'hardware trust' will expand to include PCB-level verification. We will see the emergence of startups offering non-destructive testing services for PCBs, using techniques like terahertz imaging or acoustic microscopy to detect anomalies. This will become a standard part of the AI data center procurement process.
4. The most likely near-term outcome is not a complete decoupling but a 'managed dependency' where the US accepts the risk for commercial AI but mandates domestic sourcing for military and critical infrastructure applications.
The AI security conversation must move beyond the GPU die. The PCB is the silent partner in every AI calculation, and its provenance is now a matter of national security. The industry must wake up to this reality before a crisis forces its hand.