Custom Software Used to Cost $500k. Now It Costs $8k a Month.

Here's the cost of shipping a focused custom product — say, a client portal with auth, document upload, e-signature, and a dashboard — at four points in history:

• 2014: ~$450k, 9 months, 4-person team
• 2019: ~$280k, 6 months, 3-person team
• 2023: ~$160k, 4 months, 2-person team
• 2026: ~$30k, 10 weeks, 1 senior engineer + AI

That's a 15x cost reduction in twelve years, with most of the reduction coming in the last twenty-four months. This is not a rounding error. This is an order-of-magnitude shift.

Why the collapse is real

Three curves compound:

1. Infrastructure. Vercel, Neon, Supabase, Fly, Cloudflare — the platforms absorbed the undifferentiated work of ops.
2. Frameworks. Next.js, Drizzle, Tailwind — the undifferentiated work of glue code dropped to near-zero.
3. AI-assisted engineering. A single senior engineer with Claude or GPT now outputs what a team of three did two years ago. Not by replacing the engineer — by eliminating the hours of grinding glue, tests, migration, and boilerplate.

Most firms evaluating build-vs-buy are using a 2019 cost model. That model says build = $280k. So they buy. That decision is based on a number that is no longer correct.

What the new math makes possible

At $8k/mo, a managed build of one product a quarter means $96k/year for four owned products. Each one replaces at least one SaaS subscription. Most replace two or three. The arithmetic works on almost every firm we look at.

What used to be a $450k capital decision is now a $2k/week operating decision. It's not a different number, it's a different kind of number — capex became opex, and suddenly "try it and cancel a SaaS" is reasonable.

The window

The firms that internalize this in 2026 will have an owned stack by 2028. Their competitors will still be renewing Clio and Karbon. That's a compounding delta — cost, speed of change, and customer experience — that gets very large very quickly.