Nvidia Invents 'Intelligence Per Dollar' To Sell Vera Rubin Chips
The metric divides model capability and served tokens by total training and inference cost, tested already by Prime Intellect and Together AI.
- Vera Rubin trains the largest models using one-fourth the GPUs of the Blackwell generation, Nvidia says.
- Prime Intellect Lab measured 30% greater per-CPU throughput on Nvidia's new Vera CPU for RL workloads.
- Nvidia's Nemotron 3 Ultra, a 550-billion-parameter model, scored 71.7% on the SWE-bench Verified benchmark.
Why it matters: The company selling the chips invented the metric proving they are worth buying: that is marketing, not measurement.
NVIDIA Blog ↗ · Jul 17, 20267/17/26 · ✓ Checked✓ Check