NVIDIA Vera Rubin / variant
72 Rubin GPUs + 36 Vera CPUs in one liquid-cooled rack. 3.6 EFLOPS FP4, 20.7 TB HBM4, 260 TB/s NVLink-6. Mainstream Rubin rack.
區間 + 信心水準 + 來源；不同分析師、不同時點的預估會並列。
| 期間 | 出貨量 (low–mid–high) | ASP (USD) | 信心 | 來源 |
|---|---|---|---|---|
| 2026 H2 | 1.5k–2.5k–4.0k | 3.50M | low | TrendForce 2026-04 |
| 2027 FY | 8.0k–12k–18k | 3.30M | medium | TrendForce + Reuters 2026 |
每行 = 單一機台的零組件 slot。share % 是該供應商在此 component 的估計份額。
GPU / CPU dies — the headline silicon.
Dual-die reticle design on TSMC N3P (3nm-class).
NVIDIA Arm-based Vera CPU on TSMC 3nm; LPDDR5X 1.2 TB/s on-package memory, PCIe Gen6, simultaneous multi-error correction.
HBM stacks attached to the compute die.
16-Hi HBM4 stacks attached to each Rubin GPU.
Low-power LPDDR5X on Vera CPU, 1.2 TB/s aggregate; SOCAMM serviceable module form.
Advanced packaging (CoWoS-L) + ABF substrate.
TSMC Chip-on-Wafer-on-Substrate with Local interconnect.
Ajinomoto Build-up Film substrate carrying the CoWoS-L stack.
Cold plate, CDU, hoses, manifolds.
Copper micro-channel cold plate per GPU module; VR200 compute tray is 100% liquid-cooled (vs GB200/300 ~85/15 hybrid), fanless.
High-efficiency rack manifold with universal quick-disconnects; redesigned for 100%-liquid, hose-free VR200 trays.
In-row CDU, single-phase warm-water DLC at 45°C; cooling-loop (TCS) flow ~+100% vs GB300.
EPDM hose with dripless QDC fittings.
PSUs, busbars, BBU.
Titanium-rated PSU, 54 V output, hot-swap.
Liquid-cooled 54 V DC busbar feeding compute trays (final Oberon-rack generation).
Rack-level backup battery module；800VDC 世代與 power shelf 同櫃整合（液冷）[2]
Driver + 上/下橋 MOSFET 整合單封裝；GPU 板端 Vcore 多相電源，板面積約離散方案 1/4、毛利 40–50%+。
800VDC 前端高壓開關，搭配高壓 superjunction；HVDC 電源側櫃 / PSU 前級整流升壓。
PSU / 板端整流、保護用二極體陣列；安世(Nexperia)斷料疑慮下台廠二供轉單標的。
NICs, switches, cables, CPO modules.
NVLink Switch ASIC + retimers + copper cabling, 1.8 TB/s per port.
800 Gb/s SerDes, ConnectX-9 generation.
Passive copper cable for intra-rack NVLink (NVL72 uses copper, not optics).
NVIDIA Arc Aerial RAN Computer — 6G-ready telecommunications compute platform that runs on Rubin GPU silicon. Nokia integrates this into commercial AI-RAN base-station products.
Rack, sleds, trays, brackets, heatsink fins.
21" OCP-derived rack frame, 1.5 m deep.
Aluminum sled holding 2× GPU + 1× CPU.
Auxiliary aluminum fin stack on cold plate return.
機櫃滑軌組（tool-less rail / slide），承載運算與交換托盤抽拉維護。
Motherboard, backplane, BMC, signal integrity.
Aspeed AST2700-class BMC, OCP DC-SCM module.
High-layer-count PCB carrying GPU + HBM stack.
