Starbucks just fired its software vendors. The build-vs-buy math has flipped.

Starbucks spends roughly $400 million a year on software. This month they announced they are moving off IBM and Microsoft to build their own systems in-house. IBM dropped 3% on the news. Salesforce dropped 4%.

The market did not react because Starbucks is a big customer. It reacted because of what the move signals: when AI makes building custom software 5–10x cheaper, the case for renting generic software starts to collapse. Chamath Palihapitiya put numbers on it the same week — $1 trillion a year in software licenses, $4 trillion in services and consulting around those licenses, all of it merging into one new category where companies own systems built for exactly how they operate.

Here is the part that matters if you run a mid-market company: Starbucks could always afford to build. You couldn't. That is what changed.

The old math, and why it kept you renting

Five years ago, a custom system that ran your whole operation — projects, billing, documents, reporting — meant a dev shop, a team of four to six engineers, and 12–18 months. At US agency rates of $150–$250 an hour, you were staring at $500K to $1.5M before the first user logged in. So nobody built. You bought.

But buying had its own bill. A mid-market ERP implementation runs $150K–$750K in year one once you count licenses, integrators, data migration, and training — and systems-integrator fees alone can hit $300K–$600K. Then you pay licensing forever, on software designed for the median company in your industry, which your team bends their process around instead of the other way.

The third option most companies actually chose: neither. QuickBooks, a CRM, and ten point tools stitched together with exports, spreadsheets, and someone re-keying the same data four times. The licenses look cheap. The leaked hours are not.

What changed

AI-assisted development collapsed the build side of the equation. Work that took a team of six developers a year and a half now takes a small team a quarter. The same scope of system — dozens of screens, AI reading inbound documents and structuring the data, automations handling the reminders and paperwork nobody should touch manually — now lands at a fraction of the old price.

The enterprise world has already repriced. Chamath's own company, 8090, sells fully managed AI-built software to regulated enterprises — starting at $1 million a year. Salesforce Ventures just led their $135M Series A. That tells you two things: the smartest money believes owning beats renting, and the enterprise players have set their pricing floor well above what a mid-market company will ever pay.

Which leaves the actual opportunity sitting with companies doing $10M–$500M in revenue. Too big for duct tape. Priced out of 8090. No longer priced out of building.

The new numbers, side by side

Path	Year-one cost	Timeline	Who owns it
Mid-market ERP implementation	$150K–$750K, then licensing forever	12–18 months	The vendor
Traditional dev shop build	$500K–$1.5M	12–18 months	You
AI-leveraged custom build	A fraction of either	One quarter	You

Ongoing cost does not disappear when you build — plan on maintenance of roughly 15–30% of the build cost per year, whether through a retainer or your own team. But compare that to per-seat licensing on ten tools, compounding annual price increases, and the workaround hours, and the three-year picture usually is not close.

Where buying still wins

This is not "build everything." Some software is a commodity, and building commodities is how custom projects go to die:

• Your accounting ledger. QuickBooks and NetSuite exist for a reason. Compliance, audit trails, and payroll are solved problems. Integrate with the ledger; never rebuild it.
• Email, docs, calendars. Nobody has ever won by rebuilding Gmail.
• Anything a clean existing system already does well. The rule is simple: keep whatever costs less to keep than to replace, and build on top of it.

Build where your workflows are the business: how projects move, how estimates become invoices, how documents get read and routed, how your team knows what to do next. That layer is where generic software fits worst and where ownership pays most.

A three-step test before you spend anything

1. Count the tools. List every system your operation touches in a week, what each costs per year, and every place someone moves data between them by hand.
2. Price the leak. Estimate hours per week spent on re-keying, status-chasing, and report assembly. Multiply by loaded cost. That number is usually bigger than the license line.
3. Compare three-year costs. Current stack plus leaked hours, versus a custom build plus maintenance. If the build wins by a wide margin, it is worth a real conversation. If it does not, keep your stack — and fix the leaks with automation instead.

One more thing the Starbucks story does not tell you: the build is the easy part now. The hard part is deciding what to build. The projects that fail skip straight to code. The ones that work map every workflow with the people actually doing the work, cut the processes that only exist because the old tools forced them, and wireframe the entire system before anything gets built. If a company does those three steps and then decides not to build, the map alone usually pays for itself.

We are not theorizing. We used this exact approach to build PipeLance, our own 600,000-line pipeline operating system that consolidates 40+ go-to-market point tools into one platform — customizable per client, saving companies millions versus the fragmented stack it replaces.