CITIC Securities said AI compute is entering a full‑liquid‑cooling era, with liquid cooling shifting from an optional configuration to required infrastructure. Mainstream platforms — NVIDIA Rubin, Google TPU v7/v8, AMD Helios and ByteDance’s AI Rack 3.0 — are moving from high‑share liquid or air‑liquid hybrid designs to 100% liquid‑cooled, fanless architectures. Coupled with tight local capacity overseas, accelerating exports by Chinese suppliers and the roll‑out of China’s national standards, 2

CITIC Securities said AI compute is entering a full‑liquid‑cooling era, with liquid cooling shifting from an optional configuration to required infrastructure. Mainstream platforms — NVIDIA Rubin, Google TPU v7/v8, AMD Helios and ByteDance’s AI Rack 3.0 — are moving from high‑share liquid or air‑liquid hybrid designs to 100% liquid‑cooled, fanless architectures. Coupled with tight local capacity overseas, accelerating exports by Chinese suppliers and the roll‑out of China’s national standards, 2026 could be a turning year for concentrated order and earnings realization across the liquid‑cooling supply chain.