The Minimum Viable Unit of Saleable Software

2026-06-22 · page · source ↗ · source summary

Brandur (leaving Stainless to go full-time on River) works through the buy-vs-build economics in the age of LLMs. His core insight: LLMs made building cheaper but not free, and the expensive input is human oversight — the engineer who verifies, refines, and maintains what the LLM produces. That creates a "zone of viability" where software is novel enough that rebuilding is non-trivial and priced low enough that buying still beats building. The minimum viable unit of saleable software sits at the low end of that zone.

Key Quotes

"LLMs have made software considerably cheaper to build, but they haven't brought it to zero."

The thesis in one sentence. The LLM writes the code, but a human still has to read it, test it, debug it, and maintain it. That human costs ~$96/hour. The LLM eliminated typing, not engineering.

"Anything you ship can be instantly displaced by an internal package built by an LLM."

The existential threat for SaaS founders. But the operative question isn't can it be displaced — it's whether the displacement is worth it. Brandur's 37-month Jira break-even says no.

"Maintenance will be an ongoing cost."

The line every "just vibe-code a replacement" argument misses. LLM-built software doesn't maintain itself. Each Jira update, each security patch, each edge-case bug — you're paying engineer hours. The initial build is the cheap part.

"The math here doesn't pencil out."

On the LinkedIn Jira replacement story. $400/mo is a rounding error. Even with LLMs, the build cost amortized over 37 months makes it a worse deal than just paying Atlassian. Brandur, to his credit, admits he hates Jira too — but hatred isn't economics.

"The most expensive element being the part-time labor of the human in the equation who oversees and verifies results."

Why Brandur's framework survives even as models improve. Better LLMs reduce the refinement-loop cost but don't eliminate the human-in-the-loop. Someone has to decide whether the output is correct, and that someone bills by the hour.

The Zone of Viability

Brandur's framework is a two-condition filter for software that can survive as a paid product:

  1. Sufficient novelty — non-trivial to rebuild, even with LLMs. River's workflow engine and sequential job orchestration have enough design thought behind them that replicating them is real work, not prompt work.

  2. Moderate pricing — not so expensive that rebuild-by-LLM becomes economically rational. Salesforce at $500/seat/mo is in danger; Jira at $400/mo total is not.

This is a surprisingly durable model. It doesn't depend on AI not getting better — it depends on novel software being hard to specify, which is a property of the problem domain, not the tool.

Key Themes

#build-vs-buy #saas-economics #llm-economics #pricing-strategy #indie-dev #maintenance-cost #business-model

Critical Analysis

Brandur is right about the economics but undersells the second-order effects — and Unit Economics of AI Software adds another: even if build costs drop to zero, operating costs may not. Every LLM call a product makes carries a per-user variable cost that traditional SaaS never had. A product that's trivially cheap to build can still be uneconomical to run if its inference costs eat the margin. The build/buy calculus Brandur works through assumes the operating economics are similar for both sides; AI makes that assumption false. The real threat to his framework isn't that LLMs get better at coding — it's that they get better at specification. If an LLM can generate a Jira clone from a three-sentence prompt and maintain it with zero human review, the entire buy-vs-build calculus collapses. We're not there yet, but the direction of travel is unambiguous.

The piece is also, transparently, an exercise in self-persuasion. Brandur is quitting his job to bet on River; the blog post is him convincing himself the math works. There's nothing wrong with that — public reasoning is how good decisions get stress-tested — but the optimistic framing (River's "sufficient novelty" as a durable moat) is doing a lot of work. Workflow engines are novel now. Whether they stay novel is a bet on the speed of AI progress, not a law of economics.

The strongest part is the Jira break-even math, because it's falsifiable and specific. The weakest part is the Salesforce comparison, because it assumes a 50-seat deployment and ignores the switching costs, integration complexity, and organizational inertia that keep Salesforce entrenched regardless of per-seat price. Nobody replaces Salesforce because the math pencils out; they replace Salesforce because they've decided to endure a multi-year migration.

Brandur's pricing insight — sublinear, team-based, not per-seat — is genuinely underrated. Per-seat pricing is the SaaS industry's original sin: it creates an incentive for customers to minimize usage while the vendor maximizes features. Team-based pricing (River Pro at $125/mo for up to 20 devs) aligns incentives: the vendor wins when the team grows, and the customer doesn't pay a tax on headcount. More SaaS products should price this way, and the AI-displacement threat might finally force them to.

The Golden Age of Open Source Applications complicates the framework from a third angle Brandur never considers: adoption instead of building. Graham's company doesn't rebuild Jira — it takes an existing open-source clone 80% of the way there and has an agent wire up the last 20%. That route has no build cost and no novelty requirement, so it slips entirely past the zone-of-viability filter. The cost it doesn't escape is maintenance — which lands on the open-source maintainer, not the company.

Verdict: A crisp, useful framework wrapped in a founder's self-talk. The zone-of-viability model is portable to any SaaS product decision. The 37-month Jira break-even is the specific number worth remembering. The open question — how fast the "sufficient novelty" bar rises as LLMs improve — is the one Brandur is betting his livelihood on.

Related

Source: brandur.org, May 31, 2026. Fetched 2026-06-22.