Ever wondered what the "H", "B" and "GB" in NVIDIA GPUs really mean? I did. For a long time I kept finding specs and benchmarks, but not a simple mental model. So here is the breakdown I wish I had earlier.
The letters are a tribute to the scientists who made modern computing possible, and each generation marks a shift in what AI could do:
- V series (Volta): video analytics AI. Tensor Cores debut and AI acceleration becomes real.
- A series (Ampere): general-purpose AI workhorses, the start of training and inference at scale, with Multi-Instance GPU and massive parallelism.
- H series (Hopper): designed for transformers and large language models. This is where generative AI truly scaled.
- B series (Blackwell): trillion-parameter-era GPUs, built for AI factories rather than individual models.
- GB series (Grace-Blackwell): GPU and CPU fused into one system. Rack-scale AI becomes the unit of compute.
The rule of thumb that finally clicked for me: V = video analytics, A = broad enterprise AI, H = LLM acceleration, B = the AI-native future, GB = AI systems, not chips.
Once you see this, NVIDIA's roadmap stops looking like product marketing and starts looking like an AI infrastructure strategy.
What still strikes me every time I look at it: transistors went from billions to trillions, power from 250 W to multi-kilowatt racks, and AI performance from 12× jumps to exascale systems.
That evolution explains why AI sovereignty, AI factories and private AI infrastructure are now board-level conversations, not experiments.
