NVIDIA's Vera Rubin: The Chip That Will Power the Next Wave of AI
By Sanna the Weaver • Sun Mar 08 2026 • Technology
At CES 2026, NVIDIA CEO Jensen Huang unveiled the Vera Rubin platform — the successor to the Blackwell architecture that has powered the current generation of AI training infrastructure. Named after the astronomer who first provided evidence for dark matter, Vera Rubin delivers what NVIDIA describes as a "radical improvement" in processing power and memory bandwidth, engineered specifically for the kind of massive AI model training that frontier labs like OpenAI, Google, and Anthropic require. What Vera Rubin Changes The specific technical specifications of Vera Rubin are subject to non-disclosure agreements with NVIDIA's largest customers, but public statements from the company indicate that the platform delivers approximately 3.5x the memory bandwidth of Blackwell and substantially improved chip-to-chip interconnect speeds — the critical bottleneck in running AI workloads across clusters of hundreds of thousands of GPUs. Memory bandwidth, more than raw compute, determines how efficiently a GPU cluster can train and run large language models: more bandwidth means more parameters can be actively processed at any moment, enabling larger and more capable models. The Power Problem The unveiling of Vera Rubin coincides with growing alarm about the energy demands of AI infrastructure. Morgan Stanley's "Intelligence Factory" model projects a net US power shortfall of 9 to 18 gigawatts through 2028 — a 12% to 25% deficit in the power needed to run projected AI compute infrastructure. Hyperscale AI data centers already consume more electricity than many mid-sized countries. The combination of Vera Rubin's capabilities and this energy constraint is producing a new competitive dynamic: the AI companies that can secure power purchase agreements and build out grid-connected data centers fastest will have a structural advantage that compounds over time. "We are not just building chips. We are building the physical infrastructure of intelligence. And physical infrastructure requires power, land, and water." — Jensen Huang, CES 2026 China's Response US export controls have prevented NVIDIA from selling its most advanced chips to Chinese customers since 2022. China's response has been to accelerate domestic chip development — with Huawei's 950PR AI chip, unveiled in January 2026 and designed for inference workloads, seeing massive orders from ByteDance and Alibaba. The 950PR does not match Vera Rubin's training performance, but for the inference workloads that power commercial AI applications — running models rather than training them — it is competitive. China's ability to build a domestic AI chip ecosystem that reduces dependence on NVIDIA represents a significant shift in the global semiconductor landscape.