Breaking numbers from the chip world, and these are worth stopping the scroll for.
Nvidia's own claim was that Vera Rubin NVL72 would beat Blackwell by roughly 3x on inference for large models, in the 1 to 3 trillion parameter range.
New independent benchmarks from SemiAnalysis just blew past that claim entirely.
On a massive 1.6 trillion parameter DeepSeek model, Vera Rubin hit up to 7x better token throughput than Blackwell, at the same power draw per megawatt.
Think of it like two factories on the exact same electricity bill, except one is producing seven times more output than the other. Not just faster. Dramatically more efficient per watt.
That efficiency number matters more than raw speed right now. Power, not silicon, has become the real bottleneck for AI companies scaling inference.
Tokens per megawatt is quietly becoming the metric that decides who can afford to run bigger models at lower cost.
If these numbers hold up at scale, this could quietly reshuffle the entire chip roadmap conversation, and make anyone betting long-term on Blackwell alone think twice.
AI doesn't just need to think faster anymore. It needs to think cheaper.
🔗 Original Source & Reference: https://www.techmeme.com/260914/p41#a260914p41
Published automatically via FeedMind AI Content Pipeline.
Top comments (0)