The Build-Versus-Buy Line Moved

In April we published 16 Tools, One Platform, arguing that building our own operational tooling beat buying it. That post made the case qualitatively. It was, honestly, the easy version — nobody had audited the claim.

This is the audit. We priced every commercial product our fleet replaces, measured what the fleet actually costs to run on AWS, and pulled the build cost out of our own leverage tracker. All three numbers are below.

The headline is that the fleet costs 15.7× less than the software it replaces. That is the least interesting thing in this post.

What we would be paying

Twenty-six repos in the fleet map cleanly onto a commercial product somebody sells by the seat. Docket stands in for Jira. Chirp for Slack. Chronicle for Datadog and PagerDuty. Trellis for QuickBooks. Beacon for HubSpot. Narrative for Webflow.

At four people, list price, annual billing, that stack is $4,380/month — $52,560/year.

Two things about that number are worth pulling out.

Half of it isn't sold by the seat. The bill splits into two pricing worlds. Seat-priced products bill per person and grow linearly. Platform-priced products — observability, marketing automation, accounting, the site builder — bill on hosts, contacts, records or traffic, and cost nearly the same whether four people use them or forty. At four seats those platform fees are 46.6% of the bill.

Enterprise minimums are the real tax on being small. Okta won't sell a $6 seat without a $1,500 annual contract. PagerDuty won't sell fewer than five seats. Glean has roughly a 100-seat floor and a $60K annual contract minimum — you buy 100 seats or you get no Glean. Anaqua, for patent portfolio management, is quote-only and realistically $15K/year at the bottom.

Force the two unbuyable products in at their floors and the four-person bill becomes $10,630/month — $127,560/year. Two and a half times the price, for the same four people.

We also ran it at 500 employees, allocating seats by role rather than pretending everyone needs a Snyk license. That lands at $159,045/month — $1.91M/year, with Datadog and enterprise search alone accounting for 30% of it.

What the fleet actually costs

This part is a bill, not a model.

The numbers come from AWS Cost Explorer on amortized cost, measured across ten clean days in August 2026 and normalized to 30. Amortized matters here: unblended cost reports Savings-Plan-covered compute at $0 and would have hidden roughly a third of the fleet's real footprint. We attributed by resource rather than by tag, because cost allocation tags aren't activated on the account — a gap we found while doing this and have since filed.

$280/month. $3,354/year. Against the buyable four-seat stack, that is 15.7× cheaper and saves $49,206/year. Against the version with enterprise floors forced in, 38×.

Fifteen production tools run on six vCPUs. That is not a clever optimization; it is what a fleet of small, single-purpose services actually needs when nothing is a vendor's multi-tenant platform carrying everyone else's features.

What it cost to build

Here is the half that decides the argument.

Every non-trivial session against these repos is logged to Fulcrum, our own leverage tracker, with an estimate of how long the same work would have taken a senior engineer already familiar with the codebase. Those estimates are recorded when the work happens, not reconstructed afterwards. Across 33 tools and 397 logged sessions:

The spread is instructive. The highest-leverage builds were the ones with the most mechanical surface — Envoy at 540×, Atlas at 157×, Packed at 133×. The lowest were the ones that needed real judgement and iteration: supporting-services at 6×, the fleet meta-repo at 7×, Lattice at 12×. Leverage tracks how much of a task is transcription versus how much is decision.

The part that actually matters

Run the build-versus-buy arithmetic the traditional way and building this fleet is indefensible.

14,105 hours at a $100/hour loaded rate is $1.41 million. At $150/hour it is $2.12 million. Spending $1.41M to avoid $49K/year of SaaS is a 29-year payback. No competent engineering leader would approve that, and they would be right not to.

What changes the answer is the 62.7×. The build consumed 225 hours of wall-clock time and a flat monthly AI subscription — not a seven-figure engineering budget. Against $49,206/year of avoided SaaS, that pays back in weeks.

So the finding here is not "SaaS is overpriced." The vendors are charging roughly what that software costs to build and support under the old assumptions. The finding is that the assumptions moved.

Build-versus-buy has always been a comparison between a large known cost and a small recurring one. Buy won almost every time, for good reasons: the build number was enormous, and it was enormous because software took human-years. When the build number drops by a factor of sixty, the comparison doesn't shift at the margin — it inverts for an entire category of software that was never close before.

Every company sitting on a $50K/year SaaS bill has been doing correct arithmetic. They've been doing it against the old number.

What we are not claiming

A few honest caveats, because this analysis is easy to over-read.

The SaaS figures are list price. Real enterprise contracts land 10–30% below list. Apply that and the $1.91M becomes $1.4–1.7M. It doesn't change the shape of anything.

The leverage estimates are ours. They're per-session judgements about how long work would have taken, made by the person who did the work. They are a considered figure, not a measurement, and they should be read that way.

Six tools have no commercial equivalent at all, so they're excluded from every total rather than guessed at — which makes each SaaS figure a floor, not a ceiling.

Building means owning. Every one of these tools is now ours to patch, migrate, secure and keep running. That cost is real and it does not appear in the $280. The reason it's tolerable is the same reason the build was: the marginal cost of maintenance fell by the same factor the build did.

This worked for operational tooling. These are internal tools with known requirements and a user base we can walk down the hall to. We are not claiming you should build your own database.

Where this leaves us

Fifteen production tools, $280 a month, and a build that would have cost seven figures eighteen months ago.

The number we'd point at isn't the $49K we save every year. It's the 225 hours — because that's the one that used to be 14,105, and it's the one that changes what a small company is allowed to consider building.

The full analysis, including every priced line item at both four seats and five hundred employees, is published as a standalone cost breakdown.