If you've been following the AI hardware race, you've probably seen plenty of coverage about GPU shortages, HBM capacity, and advanced packaging. But there's a layer beneath all of that which gets far less attention — and it may be the most critical bottleneck of all.
The substrate.
And in most cases, the substrate is a printed circuit board.
Why the Substrate Matters More Than Ever
For decades, the PCB was treated as a passive carrier. You designed your circuit, sent the Gerber files to a fabricator, and got back a board. The PCB didn't determine performance — the chips did.
That assumption is breaking down, fast.
Consider the data:
- PCB value per AI server rack has grown to nearly 10× that of a standard server rack.
- AI server PCBs now require 20–40+ layers, compared to 8–12 layers for standard servers.
- CoWoS packaging and similar 2.5D/3D technologies require substrates with line width and spacing below 15μm — approaching IC substrate precision.
- Glass core substrates, the next big thing in advanced packaging, are being developed by Intel and Samsung to replace organic substrates, with a CTE of ~3.2 ppm/°C compared to 12–17 ppm/°C for organics.
The substrate is no longer just connecting chips. It's becoming the performance enabler — and in some cases, the limiting factor.
Three Forces Driving the Substrate Revolution
1. AI Chip Packaging Is Pushing Substrates to Their Limits
AI accelerators are getting bigger. Much bigger. NVIDIA's next-generation chips are pushing package sizes beyond 100mm × 100mm. At these dimensions, organic substrates exhibit warpage that degrades yield.
This is why the industry is moving toward glass core substrates. Glass offers:
- CTE closer to silicon (~3.2 ppm/°C vs. 12–17 ppm/°C)
- Better dimensional stability at large sizes
- Lower dielectric loss for high-speed signals
- Higher interconnect density (line/space down to 2–3μm) The challenge? Glass processing is difficult, and the supply chain for glass substrates is still in its early stages.
2. Signal Integrity Requirements Are Reaching New Extremes
112G PAM4 and 224G SerDes are pushing PCB materials to their limits. Standard FR-4 can't support these speeds. The industry has moved to M7, M8, and M9-grade laminates — with dissipation factors dropping from 0.02 to below 0.002.
But materials alone aren't enough. The substrate design — layer stackup, via structures, impedance control — must be optimized for these speeds. This is why back-drilling, via-in-pad, and mSAP processes are becoming standard in high-end substrates.
3. Supply Chain Constraints Are Reshaping the Industry
ABF substrate capacity has been tight for years, and AI demand is making it worse. Major substrate manufacturers are prioritizing AI customers, leaving other segments to compete for remaining capacity.
This is driving investment in new capacity, particularly in Southeast Asia and China. But new capacity takes time — and the gap between demand and supply isn't closing quickly.
What This Means for Hardware Engineers
1. The PCB Is No Longer a Commodity
If you're designing high-performance systems, the PCB can no longer be treated as an afterthought. It's a performance-critical component that must be designed with the same rigor as the chips it connects.
2. Substrate Design Is Becoming a Specialized Skill
Designing a 30-layer AI server board with M9-grade materials, 112G signal integrity requirements, and glass-core substrate technology is not the same as designing a standard 4-layer board. The skills required — high-speed signal integrity, power integrity, thermal management, material science — are increasingly specialized.
3. Supply Chain Knowledge Matters
When substrate materials have long lead times and limited supply, understanding the supply chain becomes part of the design process. Engineers who know which materials are available, which are constrained, and which have alternatives will design better products.
4. Testability and Reliability Are Non-Negotiable
AI servers run 24/7 at full load. Reliability isn't optional. Thermal cycling performance, via reliability, and material stability are all under scrutiny. Designers need to think about long-term reliability from day one.
The Bottom Line
The AI hardware race isn't just about chips. It's about everything that connects them — and the substrate is at the center of that.
For hardware engineers, this is both a challenge and an opportunity. The challenge: keeping up with rapidly evolving substrate technologies. The opportunity: positioning yourself as someone who understands the full stack — from chip to package to board.
The engineers who thrive in the next decade will be those who see the substrate not as a passive carrier, but as the connective tissue that makes advanced systems possible.
If You're Designing High-Performance PCB Projects
AnyPCBA has over a decade of experience in small-to-medium batch PCB manufacturing, supporting 2–64 layers with HDI, rigid-flex, and high-frequency hybrid capabilities. Our engineering team provides DFM/DFA design reviews to help you identify potential issues in stackup, impedance, material selection, and manufacturability.
Top comments (0)