NVIDIA Vera Rubin / variant
576 Rubin Ultra GPUs in a single rack at ~600 kW. First NVIDIA system shipping with co-packaged optics (CPO) NVSwitch, replacing pluggable transceivers. Densest AI training pod NVIDIA has shipped.
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 |
|---|---|---|---|---|
| 2027 H2 | 50–200–400 | 28.00M | rumor | CPO ecosystem timing |
| 2028 FY | 300–800–1.5k | 26.00M | rumor | CPO ecosystem timing |
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.
Top-bin Rubin (likely 2x reticle), full HBM stack allocation.
88-core Vera CPU.
HBM stacks attached to the compute die.
Top-bin 16-Hi HBM4; supply is the gating factor.
Advanced packaging (CoWoS-L) + ABF substrate.
NVL576 absorbs a large share of TSMC CoWoS-L 2027 capacity.
Cold plate, CDU, hoses, manifolds.
High-performance copper cold plate, 600W+ per die.
600 kW rack; multiple CDUs feed a single rack.
PSUs, busbars, BBU.
600 kW rack needs higher density PSU shelves.
NICs, switches, cables, CPO modules.
NVL576 ships with co-packaged optics NVSwitch — first NVIDIA system to do so. Replaces pluggable transceivers.
Single-mode optical fiber + connectors between CPO modules.
Rack, sleds, trays, brackets, heatsink fins.
Motherboard, backplane, BMC, signal integrity.
