NVIDIA Rubin CPX Launch Delayed Due to Failed Base Orders Last September, NVIDIA unveiled the design roadmap for Rubin CPX, anticipating a launch by the end of 2026. However, at GTC this March, the Rubin CPX accelerator was completely absent from any slides or agenda, sparking speculation of its cancellation. NVIDIA Vice President Ian Buck subsequently confirmed to ComputerBase that Rubin CPX had been "delayed" and was positioned for the Feynman generation (expected launch in 2028). The economics of CPX—it was originally a product of GDDR/HBM price arbitrage, using cheap, high-capacity memory to handle compute-bound tasks with less demanding bandwidth requirements. When the absolute price of GDDR7 was caught in the DRAM price increase cycle (consumer-grade DRAM increased by 75-80% in Q1 and another 45-50% in Q2), this selling point was indeed damaged. NVIDIA's official reason for cancellation: Changes in workload structure. Ian Buck's original words: CPX only improves TTFT, not actual token speed, and prefill accounts for a small percentage of the overall workload. In the new agent-to-agent scenario with 400-500K KV context and 1000 tokens/sec, the bottleneck shifts from prefill latency to decode throughput. thelec.net/news/articleView.ht...…