"Time your scale-up to the axis that matters Inception chose to commercialize diffusion LLMs specifically when the industry shifted from competing on training-time scaling laws to competing on inference-time efficiency. As the speaker explains, 'right now everything has shifted to inference time scaling' because 'the price per token or what's needed per token becomes the key metric.' Diffusion models 'scale better than autoregressive models at inference time — they're cheaper to serve, they're faster, you get more tokens per GPU.' The lesson is to time a new approach's debut not when it first technically works, but when the metric it excels at becomes the one the market actually rewards. Source: Arash Vahdat, Latent Space Podcast, Diffusion LLMs with Inception AI"