He had Claude Code break NVLink down into a 3D model, and the comments first ask whether the $20 plan is enough
Nicolas Bustamante wanted to understand NVIDIA NVLink, so in the morning he used Claude Code to build a 3D explainer, breaking it down from the rack all the way to the GPU. In the video you can see compute trays, switch trays, power shelves, and 46TB/s of intra-rack bandwidth. The comments are more concerned with whether the $20/month plan can reproduce it.
One morning, Nicolas Bustamante wanted to understand NVIDIA NVLink. Instead of digging through white papers, he had Claude Code build a 3D explainer. The sequence went all the way from the rack down to a single GPU. He wrote on X that AI is your most patient tutor; you can learn whatever you want.
The video shows a GPU server rack. Compute trays, NVLink switch trays, and power shelves are laid out separately. A compute tray holds multiple GPU cards, which connect via NVLink to switches on other trays. The frame labels intra-rack NVLink bandwidth at up to 46TB/s. Next to it, it also lists per-GPU NVLink bandwidth, HBM capacity, memory bandwidth, and FP64 compute for different GPU models. For people unfamiliar with hardware, these terms are hard to digest when piled together; once made into a 3D model, at least you can first see clearly what connects to what.
The comments first ask whether the output can be shared. Someone says that building a model for what you want to learn helps you remember it better than reading materials, and the order from rack to GPU also makes sense. Another netizen says their learning speed is soaring, and they can have AI explain things even while walking down the street. Someone else directly asks whether this was done with Claude and whether the $20 plan can produce this kind of 3D result. Nicolas replies: Claude Code, $20/month, plus custom skills. There are many tutorials on X: paste the article into Claude and have it help you generate a skill.
The point is not whether the 3D animation is flashy. What is more worth looking at is the learning path. First choose a specific question, then have AI break it apart according to spatial structure, and finally get a model organized by rack hierarchy. The process of making the explainer forces you to organize information: what goes on which layer, how the lines run, and which parameters correspond to whom. This action is closer to understanding than bookmarking ten NVLink articles.
But a 3D model is not understanding itself. It lays out the structure; the remaining follow-up questions still have to come from you. Why does a rack need NVLink switch trays? Under what load does 46TB/s matter? Which tasks will the FP64 gap between different GPUs affect? AI can answer these questions—provided someone is willing to keep asking.
A netizen summed it up accurately: in the past, curiosity was limited by access; now it is limited only by effort. The PDFs in your bookmarks haven't been opened yet, while someone else has already turned a morning into a rack broken down by hierarchy.
Original tweet: https://x.com/nicbstme/status/2117305455223644508
发布时间: 2026-11-03 13:41