Nvidia Cuts Rubin GPU Output Forecast to 1.5M Units with Rack Shipments Halved
NVDA•KeyBanc projects Nvidia’s CoWoS capacity to increase 76% to 650,000 units in 2026 and 29% to 840,000 in 2027, supporting Blackwell, Rubin and Hopper GPU production. Rubin output forecasts have been reduced from 2 million to 1.5 million units, and rack shipments cut to 6,000 units due to HBM4 delays.
1. CoWoS Capacity Expansion
Nvidia’s chip packaging capacity under CoWoS is projected to rise 76% to 650,000 units in 2026 and a further 29% to 840,000 units in 2027. This expansion underpins production plans for Blackwell, Rubin and Hopper GPUs, ensuring manufacturing capacity aligns with anticipated AI demand.
2. Rubin Output Revision
Production forecasts for Rubin GPUs have been revised downward from an earlier 2 million units to approximately 1.5 million units for 2026. This adjustment reflects updated expectations for availability of next-generation AI accelerators.
3. Rack Shipment Reductions
Server systems outfitted with Rubin GPUs are now expected to ship around 6,000 racks, down from previous projections of 12,000 to 14,000 units. This halving of rack volumes will affect initial deployments in hyperscale and enterprise data centers.
4. Cause and Outlook
Qualification delays for HBM4 memory have been identified as the primary bottleneck slowing Rubin ramp-up. Resolution timelines for memory validation will determine when production and system shipments can return to original forecasts.
