Nvidia groups up with DeepSeek for R1 optimizations on Blackwell, boosting income by 25x

Whereas Western and even home firms like OpenAI and Alibaba (respectively) are attempting to take down DeepSeek by pushing their AI fashions into excessive gear, Nvidia is one firm that sees this breakthrough as a constructive. We noticed this when DeepSeek was first launched, with an Nvidia spokesperson calling it “a wonderful AI development” and stating that they do not see it as a damaging as a result of “inference requires vital numbers of NVIDIA GPUs.”

Nvidia is sticking to its stance on DeepSeek, as the corporate has introduced a partnership with DeepSeek to carry DeepSeek R1 optimizations for his or her Blackwell structure. This seems to ship vital leaps in AI inference efficiency and is anticipated to spice up income ranges way over earlier than.

AI inference is now considerably cheaper and sooner

NVIDIA AI Developer introduced the information on X, claiming that these new optimizations ship 25 instances extra income at 20 instances decrease value per token in comparison with the H100 GPU simply 4 weeks in the past. To get a tough thought of the income enhance, let’s say your AI system usually generates $100,000 in income and prices $50,000 to run, leaving you with a internet revenue of $50,000. With a 25x income enhance, that quantity jumps to $2,500,000, and with 20x decrease prices, your operating bills drop from $50,000 to simply $2,500, bringing your potential complete revenue to $2,497,500.

A key a part of this enhance comes from FP4 precision, which permits B200 GPUs to course of extra information with much less energy, making AI inference considerably cheaper and sooner. Whereas the accuracy is not as excessive as with higher-bit codecs like FP8 or FP16, NVIDIA claims the optimized mannequin nonetheless achieves 99.8% of FP8’s accuracy in benchmark checks, which means it performs almost as properly whereas being much more environment friendly.

Blackwell B200 crushes H100 with over 25x sooner AI processing

Nvidia additionally shared a efficiency chart to indicate the affect of those enhancements the place Blackwell-based B200 GPUs had been in a position to course of 21,088 tokens per second. This may not sound like a giant deal however if you examine this from 5,899 tokens per second on the H200 (February 2025) and simply 844 tokens per second on the H100 (January 2025), it reveals simply how huge of a leap that is. For builders desirous to attempt it out, NVIDIA has launched an FP4-optimized DeepSeek checkpoint on Hugging Face, giving early entry to those effectivity features. That stated, DeepSeek’s extra open method has been a key spotlight amongst specialists, and with Nvidia pushing issues even additional, it reveals how impactful open fashions will be.