it computes the gradient using *every* data point before updating the weights. With 4 customers, this is instant. With 2.4 million StreamFlow subscribers, each iteration requires a matrix multiplication with a $(2400000, 47)$ matrix. That is slow.