🐯虎嗅•Stalecollected in 15m
PCB Boom from Nvidia AI Servers
💡Nvidia's new AI chips spike PCB demand—key for data center builders
⚡ 30-Second TL;DR
What Changed
Groq 3 LPU server: 32 trays/cabinet with 256 LPU chips, increasing PCB needs
Why It Matters
AI inference demand drives PCB capacity expansions, benefiting suppliers and ensuring Nvidia's scaling; expect tiered performance leaps in 2026+.
What To Do Next
Assess PCB suppliers for custom high-speed boards in your AI cluster builds.
Who should care:Enterprise & Security Teams
Key Points
- •Groq 3 LPU server: 32 trays/cabinet with 256 LPU chips, increasing PCB needs
- •Rubin Ultra orthogonal backplanes replace copper cables for higher density
- •PCB makers like Peng Ding invest 11B CNY in AI high-end production
- •M9/M10 materials testing for next-gen AI chips by沪电股份
🧠 Deep Insight
Web-grounded analysis with 4 cited sources.
🔑 Enhanced Key Takeaways
- •The Rubin Ultra NVL576 architecture introduces a massive 78-layer orthogonal backplane (synthesized from three 26-layer boards) that replaces over 20,000 copper cables, reducing signal degradation and increasing assembly efficiency by 40%.
- •Nvidia's acquisition of Groq in December 2025 has positioned the Groq 3 LPU as the 'seventh chip' of the Vera Rubin platform, utilizing a deterministic 'LPU-Link' architecture that relies on high-speed PCB spines rather than traditional NICs for rack-scale scaling.
- •Next-generation M9 and M10 materials utilize Quartz Fiber (Q-cloth) instead of standard glass, achieving a dissipation factor (Df) below 0.0007 and an 80% reduction in thermal expansion, which is critical for 224Gbps SerDes signaling.
- •The industry faces a projected 'Material Shortage Year' in 2026, with a 25% supply-demand gap for HVLP4 copper foil and a 45% gap for Low Dk glass fabric as manufacturers pivot to Rubin-class production.
📊 Competitor Analysis▸ Show
| Feature | Nvidia Rubin Ultra (LPU/GPU) | AMD Helios (MI400 Series) | Traditional AI Servers (H100/B200) |
|---|---|---|---|
| Interconnect | Orthogonal PCB Backplane (Cable-less) | High-speed Copper Cabling | InfiniBand / NVLink Copper Cables |
| PCB Layers | 70 - 78 Layers | 28 - 44 Layers | 18 - 28 Layers |
| Material Grade | M9 / M10 (Quartz-based) | M7 / M8 (Glass-based) | M6 / M7 (Standard Low-Loss) |
| Max TDP | 2300W - 3700W | 1500W - 2000W | 700W - 1200W |
| Bandwidth | 3.6 TB/s (NVLink 6) | 2.0 - 2.5 TB/s (Infinity Fabric) | 900 GB/s - 1.8 TB/s |
🛠️ Technical Deep Dive
The shift to the Rubin Ultra architecture represents a fundamental change in PCB engineering and material science:
- Orthogonal Backplane Synthesis: To achieve the required density for 576 GPUs, manufacturers are bonding three 26-layer boards into a single 78-layer structure. This creates a board thickness of ~2cm with an extreme aspect ratio (>100:1) for through-hole copper plating.
- M9/M10 Material Specs: These materials feature an ultra-low dielectric constant (Dk < 3.0) and utilize 40% spherical silica filler to double thermal conductivity compared to M7, essential for the 2300W+ TDP of Rubin GPUs.
- Groq 3 LPU Architecture: Ditching Von Neumann architecture for a 'data flow' model, the Groq 3 LPU features 500MB of on-chip SRAM with 40PB/s bandwidth. The PCB must support 96 C2C links per chip at 112Gbps SerDes speeds.
- Drilling Precision: The use of quartz fiber (Q-cloth) requires specialized diamond-coated drill bits and slower feed rates to prevent 'delamination' and 'fiber protrusion' during the high-layer count manufacturing process.
🔮 Future ImplicationsAI analysis grounded in cited sources
PCB Supply Chain Consolidation
Only a small tier of global manufacturers (e.g., WUS, Peng Ding, TTM) can currently meet the 70+ layer and M9 material requirements, leading to a high-margin monopoly for elite suppliers.
End of the Copper Cable Era in Racks
As SerDes speeds move toward 448Gbps, the physical bulk and signal loss of copper cables will force all high-density AI clusters to adopt orthogonal PCB backplanes.
Thermal-Driven PCB Redesign
With GPU TDP reaching 3700W by late 2026, PCBs will transition from being mere signal carriers to active thermal management components integrated with liquid-cooling cold plates.
⏳ Timeline
2024-03
Nvidia Blackwell Architecture Unveiled
2025-03
Vera Rubin Roadmap and Kyber Architecture Announced at GTC 2025
2025-08
Peng Ding (Zhen Ding) Approves 11B CNY AI PCB Investment
2025-12
Nvidia Completes Acquisition of Groq to Integrate LPU Technology
2026-03
GTC 2026: Rubin Ultra and Groq 3 LPX Systems Officially Launched
2026-03
Nvidia Commences M10 Material Testing for Feynman Platform
📎 Sources (4)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗