When does buy still beat build? Pricing software's LLM survival threshold

Brandur Leach, leaving Stainless to run his open-source job queue River as a business, tackles the obvious objection to selling software in the LLM era: why pay for anything a model can clone internally? His answer is that cheaper isn’t free. LLM-built systems still demand dozens of human-supervised feedback loops to reach acceptable quality, plus perpetual maintenance whose real cost is the engineer overseeing the model. He runs the numbers on a viral LinkedIn boast about replacing a $400/month Jira bill with a Claude-built tracker: at a $200k engineer’s ~$96/hour, even a generous two hours of monthly upkeep plus a two-week initial build pushes the break-even point out to roughly 37 months. The math simply doesn’t pencil out.

From this he derives a ‘zone of viability.’ Software stays worth buying when two conditions hold: it carries enough genuine novelty and ongoing maintenance burden that an LLM rebuild is non-trivial, and its price isn’t so steep that rebuilding becomes the obvious move. A 50-seat Salesforce contract at ~$25k/month flips the calculus toward build, since that buys 1.5 full-time engineers — and the market seems to agree, with Salesforce down 30% year-to-date. The ‘minimum viable unit of saleable software’ sits at the floor of this zone, below which cloning costs less than even going through a purchasing process.

Leach bets River clears that bar. Its core features are free and open source, but advanced capabilities (workflows, sequential and concurrency-limited jobs) and invoice billing sit behind a Pro tier whose API design and performance would take real effort to replicate. Priced sublinearly by team size — $125/month for up to 20 developers — it aims to stay comfortably inside the zone of viability. Whether the gamble pays off, he admits, is something only the coming months will reveal.