Nvidia Halts H200 for China, Shifts to New 'Rubin' Chips
Nvidia is reportedly halting production of its China-bound H200 AI GPUs due to shifting export controls. The company is reallocating its TSMC foundry capacity to focus on its next-generation Vera Rubin AI chips, highlighting the direct impact of geopolitics on the global semiconductor supply chain and product roadmaps.
The upcoming "Rubin" architecture, named after astronomer Vera Rubin, represents a full platform shift, not just a GPU upgrade. The system integrates a new Rubin GPU, a "Vera" CPU, NVLink 6 switches, ConnectX-9 SuperNICs, and BlueField-4 DPUs, all designed to function as a single AI supercomputer. This follows Nvidia's accelerated one-year release cadence, with Rubin planned for 2026, Rubin Ultra for 2027, and a successor architecture, "Feynman," slated for 2028. The Vera CPU component is an ARM-based chip featuring 88 custom "Olympus" cores and is designed to work cohesively with the Rubin GPUs via a high-bandwidth NVLink-C2C interconnect. This "superchip" combination of a CPU and multiple GPUs is aimed at dramatically reducing inference token costs and training times for massive Mixture-of-Experts (MoE) models. Performance specifications for the Rubin platform point to a significant generational leap. The sixth-generation NVLink interconnect will offer 3.6 TB/s of GPU-to-GPU bandwidth, double that of the Blackwell architecture. The platform will also be the first to incorporate HBM4 memory, with early reports suggesting a leap to 22.2 TB/s of memory bandwidth, a 70% increase over some initial projections. For context, the H200 GPU that was bound for China is based on the older "Hopper" architecture and was primarily an upgrade to the H100. Its main advantage was being the first GPU to use HBM3e memory, boosting memory capacity to 141GB and bandwidth to 4.8 TB/s—a 1.4x increase over the H100. The U.S. export controls, first implemented around October 2022, are designed to restrict China's ability to obtain and manufacture high-performance computing chips for advanced AI and military applications. These regulations have created a moving target for chipmakers, forcing continuous adjustments to product roadmaps for hardware sold to China.
