Nvidia Shifts to Rack-Scale Rubin as 2027 Cloud AI Accelerator Shipments Seen Up 65%
Supply chain sources say Nvidia enters heavy Vera Rubin shipments in the second half of 2026, while DIGITIMES sees 2027 high-end cloud AI accelerator shipments rising 64.5% to 25.95 million units with ASICs overtaking GPUs.
Nvidia's AI platform roadmap is moving to rack scale. Supply chain sources say the company will enter a heavy shipment phase for Vera Rubin in the second half of 2026, followed by the Vera CPU, Rubin GPU and platforms including NVL4, NVL8 and NVL72. The roadmap culminating in Rubin Ultra shifts the competitive unit from the individual accelerator to the integrated rack. That change raises content per rack in power delivery, liquid cooling, switching and board-level integration, and pushes more value into the assembly and thermal supply chain rather than the die alone.
Total accelerator volumes are set to grow sharply. DIGITIMES expects 2027 shipments of high-end cloud AI accelerators to reach 25.951 million units, up 64.5% year on year. Within that total, ASIC shipments are seen at 15.262 million units, up 110.9%, overtaking GPUs for the first time. Nvidia remains the largest single supplier with an estimated 9.096 million units, up 18.2%, although its share is seen falling to 35.1% as custom silicon programmes at Google, AWS and other cloud providers move into volume production.
Packaging choice is becoming a strategic variable rather than a procurement detail. Google's next-generation TPU project, scheduled to enter mass production by the end of 2027, is set to adopt Intel's EMIB-T advanced packaging with support from ASIC partner MediaTek, and Google is reported to be weighing a CoWoS backup as EMIB-T substrate yields lag. Advanced packaging capacity, not die supply, remains the binding constraint for both the GPU and ASIC routes. For hardware buyers, that points to continued tightness in HBM, substrates and rack-level thermal and power components through 2027.