PCB self-sufficiency below 20 percent flags AI hardware dependency

A U.S. defense official confirmed that domestic printed circuit board supply chain self-sufficiency sits below 20 percent, according to Inside Defense. The disclosure came in the context of Pentagon procurement reviews, but the constraint binds well beyond defense. PCBs are substrate to every AI accelerator, every server blade, every NIC that ships into hyperscaler pods.

The number matters because the inference layer scales on hardware deployment velocity, and hardware deployment velocity depends on PCB fab throughput and lead time. When NVIDIA ships a Hopper or Blackwell module, the board is multi-layer, high-density, and sourced almost entirely offshore. The same holds for ASIC packages from Google, Amazon, and Microsoft. Self-sufficiency at 20 percent means 80 percent of the supply chain sits outside tariff walls and outside priority-allocation regimes if those walls go up.

Current hyperscaler capex assumes uninterrupted hardware flow through 2025 and 2026. Microsoft has guided to over $80 billion this year, Google to similar scale, Meta to $60-plus billion. Those budgets are denominated in installed GPUs and TPUs, and installed GPUs require boards. A PCB bottleneck does not announce itself as a model-training delay; it shows up as longer server commissioning cycles, deferred cluster expansion, and eventually as higher effective cost per token when demand grows faster than online capacity.

The 20 percent figure also clarifies the structure of any future export-control or reshoring policy. If the U.S. moves to restrict PCB export from Taiwan or China, or to mandate domestic sourcing for federal AI projects, the domestic fab base cannot absorb the load without multi-year build-out. That build-out requires capital, cleanroom space, and workforce—none of which currently exists at scale. The inference cost curve, which has compressed steadily since GPT-3, depends on that not happening.

Open-source model economics depend on the same substrate. A Llama 3 or Qwen deployment at scale still requires server PCBs, and the cost structure for self-hosted inference assumes those boards ship on time at stable prices. If lead times stretch or prices rise, the open-closed cost wedge narrows, and the substitution dynamic that has pressured closed-model pricing since mid-2023 softens. The defense official's number is a supply-side constraint with a demand-side price signal.