víctor pérez

Krea's journey from API wrapper to training its own diffusion model (Crea 2) on owned GPU clusters in Ohio demonstrates that controlling the full stack — model architecture, post-training data curation, and inference infrastructure — is necessary to differentiate on creative quality and avoid single-point-of-failure dependency on third-party APIs.

With base architectures (Transformers, diffusion) commoditized and open-source models (Flux, Llama) closing the gap to closed-source, competitive advantage shifts to post-training: human-curated preference data, creative evaluation, and style-specific fine-tuning — areas where domain experts outperform generic RLHF.

Application-layer companies are subsidizing inference costs (up to 75%) to acquire users, but this mirrors early telecom/data pricing that eventually converged to flat-rate subscriptions. The winners will own the compute and model stack to control margins, not arbitrage subsidized APIs.