NVIDIA announced Vera Rubin - a platform for 'continuous post-training' of agentic models, meaning systems that don't stop learning from their own actions after the initial training. The company claims it does the same job with a quarter of the chips compared to the previous generation. The numbers come from NVIDIA. The direction is real though, and everyone paying the AI bill pays for it.
- NVIDIA announced Vera Rubin - a platform for continuous post-training of agentic models (they learn from their own actions, not just from data).
- Claims a quarter of the chips for the same work compared to the Blackwell generation - per NVIDIA's own data, no independent test.
- Partners with live workloads: Prime Intellect, Perplexity, Together AI. Demo model: Nemotron 3 Ultra.
Until now an AI model was made roughly like this: you train it once on a mountain of data, freeze it, and ship it as is. NVIDIA is proposing something else - a model that never stops learning while it works. To make that cheap you need different hardware. That's what they showed.
Slow down on the excitement, though. When a chipmaker tells you its new chip is four times better than the old one, that's not news - that's a price sheet. I'll believe it once someone who doesn't make money off that chip measures it.
But under the marketing there's a real direction, and it shouldn't be dismissed. 'Learning on the go' means the model is no longer a finished product. It runs and changes while you use it. And the question shifts - from 'how smart is it today' to 'what does it cost to keep it smart tomorrow'. Whoever pays the bill knows which question matters.
This is where AI pricing is heading - from 'buy a model' to 'pay to keep it alive'. For the small player that's both an opportunity and a trap: it's easier to get in, but you pay every month. There's no magic here and nothing free. And NVIDIA's number is still waiting on independent verification. For now it's theirs.