AI PCBs have officially entered their second super-cycle, with the core drivers being: value leap + technological upgrade + supply gap, not a price war.

I. Core Drivers

• Sustained high growth in overseas CSP capital expenditure: +66% in 2024, +65% in 2025, projected +69% in 2026, indicating extremely strong support for computing infrastructure.

• Exponential value surge per card/machine:

From general-purpose servers → H100 8-card → GB series → Rubin series, single-card cost pushing $600+, with per-machine value increasing up to 7-8 times.

II. Three Key Technological Increments (Most Critical)

1. Midplane Backplane

Debuting with Ruby Ultra 144GPU, starting from 78 layers → pushing 100+ layers, single piece value up to $200k, representing the largest incremental factor in this cycle.

2. LPU Inference Chip

Paired with GPUs at a 1:2 ratio, single LPU PCB value exceeds ¥2,000, adding approximately ¥4,000 per GPU.

3. COP Substrate Integration

Integration of packaging + PCB using MSAP process, with potential mass production in 1-2 years, featuring extremely high technical barriers.

III. Supply and Competitive Landscape

• Industry HPC Capex explosion: +7% in 2024 → +130% in 2025 → projected +140% in 2026, far exceeding the 5G cycle.

• The core contradiction is a high-end capacity gap, not a price war; market share redistribution does not impact leading players' profitability.

• Layer count/material/process thresholds are significantly raised, leading to continued concentration among top players.

IV. Core Conclusion

AI PCB is not a regular cycle; it's a long-term structural upgrade driven by computing architecture iteration.

Midplane backplane + LPU + high-layer-count PCBs are the strongest investment themes for the next 2 years, with leading companies continuing to benefit.

$VGT(300476.SZ)

Stock A's Changfei is still holding the line for me

B a little BC

My liquid cooling performed exceptionally well today too

Probably hitting the limit-up$Envicool(002837.SZ)