What exactly happens when you press send: Journey from Chat to a GPU

========================================================== Learn to write the programs that keep these GPUs busy, in Vizuara's Kernel Engineering Workshop: https://kernelworkshop.vizuara.ai ================================================== What happens in the one second after you press Send? We follow a single question from a phone to a datacenter, into one NVIDIA H100, and through the 36 transformer blocks of an open language model, to see where that second goes. Along the way: what a token is, why every block of a language model comes down to matrix multiplication, how a GPU splits that work across its 132 SMs and their tensor cores, and why the answer took 1.04 seconds when the math alone needed 2.4 milliseconds. Chapters 0:00 Press Send 0:08 154 tokens 0:35 The journey to a GPU 1:06 Inside the model 1:22 Matrix multiplication 1:54 In parallel 2:26 The dream, and reality 2:55 Inside one second 3:24 The road from memory 4:22 Many users at once Numbers and sources Every measurement here is our own: Qwen3-8B in BF16, served by vLLM on one NVIDIA H100 on Modal (28 September 2026), with speeds read from vLLM's own counters. One user: 149 tokens a second. 1,024 users at once: 26,303 tokens a second. H100: 989 TFLOP/s dense BF16 and 3.35 TB/s of HBM3 memory bandwidth (NVIDIA datasheet). Qwen3-8B: 36 transformer blocks, hidden size 4,096 (published configuration).