FACT As of April 30, 2026, the Kyber rack (576-Rubin Ultra) announced by NVIDIA with a production target of 2027 only operates on the 800V HVDC power architecture. This is a key change specified by NVIDIA's own technical blog (NVIDIA Tech Blog, 2025-05; Engineering.com, 2026-04). Going to 1 megawatt or more per rack is impossible without GaN·SiC·800V DC conversion devices, and the Infineon White Paper (2025) states that +92 GW of additional data center power demand will occur by 2027. Both sources point to the same conclusion — the bottleneck is caught first in the power conversion stage, not the GPU die.
NARRATIVE Memory rides a separate cycle. Multiple channels such as moomoo and NerdWallet are designating 2026~2027 as the entrance to the AI memory supercycle, and full-scale mass production of HBM4 simultaneously forces burn-in/test capacity. The same logic applies to the 600 V-class SiC MOSFET transition — if full inspection (burn-in) capacity is not increased, the yield of next-generation AI server PSUs and EV inverters will collapse. This is the exact macro entry condition for a small company's sales to increase 5x within a year.
① Rule Zero (Contemporaneity): Since there is a reference point of April closing price, stocks that have already run up +200% or more in 6 months lower the position cap with a late-entry warning. ② Ownership of Structural Bottleneck: Cases where critical parts of the BOM are occupied by a narrow margin (A1), or have theme-scarcity (A2). ③ Disqualification Marker: We simultaneously look at whether GAAP operating cash flow is already in sight of turning positive and whether the debt-equity capital structure is clean.