JPMorgan: NVIDIA’s Next-Gen Vera Rubin AI Chips Enter Final Pre-Production at TSMC Amid Surging AI Demand
NVIDIA’s AI roadmap remains firmly on track, according to a new investor note from JPMorgan analyst Harlan Sur, following a meeting with Toshiya Hari, NVIDIA’s VP of Investor Relations and Strategic Finance. The update paints a picture of overwhelming AI demand, stretched supply chains, and reassurance on the company’s next-generation GPU platform, Vera Rubin.
Sur reports that lead times for NVIDIA’s Blackwell Ultra GPUs remain at “quarters, not months”, even two years into the AI spending boom. Despite ramping up sharply in NVIDIA’s fiscal Q2 and now comprising around 50% of the Blackwell mix, supply continues to lag behind demand, underlining just how strong enterprise appetite for NVIDIA hardware remains.
The most critical takeaway: NVIDIA confirmed that all six GPUs in the Vera Rubin platform have entered their final pre-production stages at TSMC. This effectively disproves recent rumors of delays, with the company reiterating its schedule for a C2H26 (second half of calendar 2026) launch. All chips have already taped out, marking a major milestone in the transition to NVIDIA’s next architecture.
Meanwhile, NVIDIA’s Blackwell-based products for China are generating immense early demand. A Reuters report revealed that tech giants including ByteDance and Alibaba are willing to pay double the price of the H20 GPU to acquire the upcoming B30A, which reportedly delivers a 6x performance uplift compared to its predecessor. Despite U.S. restrictions, NVIDIA’s products remain the most attractive option in China, largely due to the CUDA software ecosystem and superior NVLink interconnect, which ensures higher performance in clustered deployments.
NVIDIA is also expected to see strong demand for its RTX Pro 6000D systems, based on the B40 GPU. These systems do not require a special U.S. export license since they lack high-bandwidth memory (HBM) and are primarily intended for inference workloads, not foundational AI model training. This loophole may allow NVIDIA to move significant volumes into China.
For context, the B30 GPU is a more scaled-down design compared to the B40, optimized for multi-chip clusters with dynamic compression to offset lower per-chip performance. The B40, on the other hand, is positioned as a direct replacement for the banned H20 GPU, offering stronger standalone performance.
Between overwhelming Blackwell demand, early Chinese enthusiasm for its restricted models, and Vera Rubin’s smooth progress toward launch, NVIDIA looks set to extend its dominance in AI hardware for years to come.
Do you think NVIDIA’s Vera Rubin launch in 2026 will mark another leap like Ampere or Hopper, or will competition from AMD and startups finally narrow the gap?
