The Death of Cables in AI Data Centers
Taiwan's Zhen Ding Technology just announced something that should interest anyone building AI infrastructure: they've joined Nvidia's MGX (Modular GPU eXpress) ecosystem to supply cable-free PCB-based interconnect technology. Instead of hundreds of individual copper cables connecting GPU modules in a server rack, the new approach uses high-layer-count PCB backplanes to carry 112 Gbps PAM4 signals between compute modules.
This is a fundamental architectural shift in how AI server racks are wired.
Why This Matters for Engineers
The Cable Problem at AI Scale
A modern AI training cluster has thousands of GPU-to-GPU connections running NVLink at 100+ Gbps per lane. Each connection traditionally uses either:
- DAC (Direct Attach Copper) cables: Cheap but bulky, limited to ~3m
- AOC (Active Optical Cables): Longer reach but expensive, fragile
In a densely packed 80-100kW GPU rack, these cables create multiple problems:
- Signal integrity degradation — Every cable has two connector-to-PCB transitions, each adding impedance discontinuity
- Thermal obstruction — Cable bundles block airflow in thermally critical zones
- Assembly time — Hundreds of cables per rack means hours of manual routing and testing
- Reliability — Connector contacts are the #1 field failure mode in data centers
The PCB Solution
Zhen Ding's cable-free approach replaces discrete cables with PCB backplanes/midplanes that are pre-manufactured with all interconnects integrated. Think of it as a giant multi-layer PCB that sits between server modules, carrying signals directly through impedance-controlled traces.
Technical Specifications (What We Know)
Based on publicly available information about MGX-class backplanes:
| Parameter | Requirement |
|---|---|
| Layer count | 30–60+ layers |
| Signal speed | 112 Gbps PAM4 per lane |
| Impedance tolerance | ±5% on 85Ω differential |
| Board thickness | 6–10mm |
| Board length | 600–1000mm |
| Material | Very low-loss (Dk 3.0, Df 0.002) |
| Copper foil | ULVP (Rz < 1.5µm) |
| Via technology | Sequential lamination + laser drill + backdrill |
These are among the most challenging PCBs manufactured anywhere. The combination of extreme layer count, very-low-loss materials, ultra-smooth copper, and ±5% impedance over 600+ mm lengths pushes every aspect of PCB fabrication to its limits.
Why Nvidia Needs This Now
Nvidia's GB200 NVL72 racks and future Rubin-based systems push interconnect density to levels where discrete cables become impractical:
- 72 GPUs per rack with NVLink mesh connectivity
- Thousands of 112G lanes between compute trays
- 3-5 minute module swap target (impossible with hundreds of cables)
- Liquid cooling leaving zero tolerance for cable obstruction
The MGX modular architecture is designed so compute trays slide into a pre-wired chassis — cable-free by design.
PCB Content Per Server: A Paradigm Shift
This trend dramatically increases PCB value per AI server rack:
- Traditional cabled rack: $2,000–$5,000 in PCBs
- Cable-free MGX rack: $5,000–$15,000 in advanced PCBs
That's a 2-3× increase in PCB dollar content per rack, and AI data center buildout is accelerating. For the PCB industry, this represents a significant structural demand driver.
Implications for PCB Engineers
If you're working on AI infrastructure:
Material selection is critical — Standard FR-4 cannot meet Df < 0.003 at 28 GHz. You need Megtron 7, Tachyon, or equivalent ultra-low-loss laminates.
Copper roughness matters — At 56 GHz (112G PAM4 Nyquist), standard copper foil can double your conductor loss vs. ULVP. Budget 0.8-1.2 dB/inch difference.
Via transitions are the bottleneck — Every layer change needs carefully designed anti-pads, back-drilled stubs, and return-path vias. 3D electromagnetic simulation is mandatory.
Fabricator qualification is essential — Not all fabricators can hold ±5% impedance on 40+ layer stackups with mixed-sequential lamination. Verify capability with test coupons before committing production.
Lead times are long — 6-12 weeks for this class of board. Plan accordingly.
The Bigger Picture: AI Reshaping PCB Supply Chains
This partnership between Zhen Ding (world's 3rd largest PCB maker) and Nvidia further concentrates the AI hardware supply chain in Taiwan. With TPCA projecting Taiwan PCB output exceeding NT$1 trillion (US$33.5B) in 2026, the industry's center of gravity continues to shift toward AI-driven applications.
For hardware teams outside the hyperscaler ecosystem, the takeaway is that advanced PCB manufacturing capacity is being absorbed by AI infrastructure demand. Securing capacity for complex boards (12+ layers, low-loss materials) requires earlier planning and potentially premium pricing.
Source: DIGITIMES, June 2, 2026
Originally published at AtlasPCB
Top comments (0)