Weekly Radar 007: NVIDIA Moves to Secure Substrate Capacity - TrendForce

NVIDIA's pre-booking of substrate capacity signals that the AI chip bottleneck has migrated from silicon dies to the packaging substrate layer. ABF substrate area per package has ballooned—B200-class parts exceed 2,500 mm² with 20+ RDL layers—while effective global capacity remains concentrated in Ibiden (Japan) and Unimicron (Taiwan, China), with 18-24-month expansion lead times. This is no longer a procurement issue; it is a strategic resource allocation problem. For AMD, Intel, and hyperscaler ASIC programs competing for the same substrate pool, NVIDIA's long-term reservations compress available headroom and shift pricing power decisively toward the two dominant suppliers. Expect 15-20% unit-cost escalation through 2026. Technically, copper-film uniformity and warpage tolerances demanded by next-gen packages are pushing organic materials toward their physical ceiling, accelerating R&D into glass-core substrates and advanced RDL-on-substrate processes. Over the next 12-24 months, substrates will become the primary constraint on AI compute delivery—replacing the logic wafer as the binding bottleneck. The 'lock capacity, lock delivery' dynamic will mirror the 2021 memory-chip scramble, except this time there is no spot market to arbitrage. The company that secures substrate first secures the AI revenue cycle.