Nockchain's AI proof-of-work, measured in the unit the AI industry runs on
Multiply two numbers, add the result to a running total. That is one multiply-accumulate. Every layer of a neural network is a matrix multiplication, and a matrix multiplication is nothing but MACs — training runs them over data to fit the weights, inference runs them over those weights to answer a prompt. MAC/s is the unit the AI industry runs on, whether or not it uses the word.
The AI puzzle is a tiled INT8 matrix multiplication whose useful work is counted in MAC-equivalents of the job's shape. Like hashrate, the network figure is implied from difficulty and block cadence, not measured on hardware. Realized above the difficulty-implied rate means blocks are landing faster than the target: capacity arriving ahead of the retarget.
Compared against the H100's dense INT8 tensor throughput — 1,979 TOPS, 0.99 PetaMAC/s — because the puzzle cannot use structured sparsity. Real GPUs rarely sustain their datasheet peak, so the equivalence is a floor on the hardware involved. Price ($25,000) and cloud rate ($2.50/h) are stated assumptions for scale, not measurements.