NVIDIA Vera Rubin Maximizes Intelligence per Dollar for
Lowest cost per token from extreme codesign maximizes intelligence per dollar for post-training in the agentic era.
What the video says
Just as an elite athlete refines their game between matches, agentic AI models must constantly adapt to shifting environments and new tools through continuous post-training. Sources indicate this endless learning cycle has become the central workload of the agentic era, driving NVIDIA to introduce its new Vera Rubin platform.
This new hardware is designed to maximize intelligence per dollar, training the largest models with just one-fourth of the GPUs required by the previous Blackwell generation. While pre-training only teaches a model fluency by predicting the next token, NVIDIA notes that post-training is where actual intelligence like planning and coding is built.
To demonstrate this, the company disclosed that its 550-billion-parameter Neutron 3 Ultra model scored over 71% on the SwaeBench verified coding benchmark. Partners are already optimization testing, with Prime Intellect reporting that VERA CPUs deliver 30% greater throughput on average than alternative x86 architectures.
Taken together, these advancements allow platforms like Perplexity and Together AI to run massive reinforcement learning loops that keep their models adapted. Ultimately, the industry is shifting toward these continuous post-training runs, where Vera Rubin aims to deliver the lowest possible cost per token.
