$NVDA is running two product cycles simultaneously — and the economics are wild.
Vera Rubin NVL72 is now in production. First mass unit rolled off Foxconn's fully automated line and shipped to Microsoft. Google, Oracle, CoreWeave, and SpaceX are also lined up. Groq 3 LPX is ramping alongside for low-latency inference.
Each Rubin compute tray uses third-gen MGX: no cables, no fans, no hoses — full liquid cooling. A robot completes one tray per minute. The rack holds 18 compute trays and 9 switch trays.
TrendForce expects GB300 to remain the workhorse through early 2027, then VR200 takes over. NVL72 shipments across GB and VR lines are projected up 50%+ in 2027. Here's the kicker: VR ASP is roughly double GB300. With wafer and HBM costs still climbing, 2027 NVL72 value could hit $710 billion — up 214%.
Foxconn's rotating CEO says AI rack shipments rose by a high double-digit rate in Q3, with Rubin becoming next year's main product as they chase 50% global AI server share.
Quanta says the Rubin changeover is underway, expecting AI server revenue to rise double digits this quarter and double for the year. Wistron has started Rubin production in Taiwan, with Texas next.
The entire Taiwan supply chain is mobilizing: TSMC, Foxconn, Quanta, Wistron, Wiwynn, Delta, AVC, Auras. The dual ramp only works if HBM and substrates can keep pace with a cabinet that costs twice as much.
This is the bet: two engines running at once, each cabinet doubling in price, and the infrastructure has to hold.
Vera Rubin NVL72 is now in production. First mass unit rolled off Foxconn's fully automated line and shipped to Microsoft. Google, Oracle, CoreWeave, and SpaceX are also lined up. Groq 3 LPX is ramping alongside for low-latency inference.
Each Rubin compute tray uses third-gen MGX: no cables, no fans, no hoses — full liquid cooling. A robot completes one tray per minute. The rack holds 18 compute trays and 9 switch trays.
TrendForce expects GB300 to remain the workhorse through early 2027, then VR200 takes over. NVL72 shipments across GB and VR lines are projected up 50%+ in 2027. Here's the kicker: VR ASP is roughly double GB300. With wafer and HBM costs still climbing, 2027 NVL72 value could hit $710 billion — up 214%.
Foxconn's rotating CEO says AI rack shipments rose by a high double-digit rate in Q3, with Rubin becoming next year's main product as they chase 50% global AI server share.
Quanta says the Rubin changeover is underway, expecting AI server revenue to rise double digits this quarter and double for the year. Wistron has started Rubin production in Taiwan, with Texas next.
The entire Taiwan supply chain is mobilizing: TSMC, Foxconn, Quanta, Wistron, Wiwynn, Delta, AVC, Auras. The dual ramp only works if HBM and substrates can keep pace with a cabinet that costs twice as much.
This is the bet: two engines running at once, each cabinet doubling in price, and the infrastructure has to hold.
