switchcasestudios.com
Building the whole stack feels like a moat. The numbers say it's mostly overhead in disguise. The companies running away with this cycle built one thing that was actually theirs — and rented everything else.
There's a particular flavor of founder energy that wants to build the whole thing. Own every layer. No vendors, no dependencies, nobody else's API sitting in the critical path. It feels like control. It feels like a moat. Most of the time it's just a very expensive way to reinvent plumbing that already exists and works better than whatever you'll ship.
The numbers get brutal the second you actually look.
Want your own foundation model? GPT-4 cost an estimated $78–100M+ in compute alone to train, and Gemini Ultra ran roughly $192M, per the Stanford AI Index. Frontier runs are on pace to cross a billion dollars per model by 2027. Fine-tuning an existing open model to your domain costs 1–5% of that — often a few thousand dollars with techniques like LoRA. The build-it-yourself version isn't ten percent pricier. It's a hundred times pricier, and you still finish behind the lab you could have just paid.
Don't want a model — just want users to log in? Building custom authentication in-house runs an estimated $250,000–600,000 in engineering time, and a robust system carries six-figure annual maintenance on top of that. FusionAuth's own framing is blunt: fewer than 5% of engineering teams should build auth from scratch. Meanwhile Clerk hands you 50,000 monthly users free and then charges two cents a head; Auth0's free tier covers 25,000. You can run real, SOC 2-grade authentication for the price of a team lunch — and skip the part where you're personally on the hook for a credential breach.
This is the tax nobody writes on the roadmap. Every system you build, you also own — forever. The patching, the on-call rotation, the security audit, the one engineer who actually understands it handing in their notice. Amazon has a name for this category of work: undifferentiated heavy lifting. Necessary, real, and completely invisible to your customer. Building it doesn't make your product more yours. It just makes it more expensive.
The overhead you can't see on day one is the overhead that sinks you in year three. Every layer you build, you babysit — indefinitely.
— SWITCHCASE STUDIOS · PRODUCT PRACTICEHere's the company that should retire the "but real builders build everything" argument: Cursor.
When it launched, Cursor was — by the dismissive definition everyone loves to throw around — a wrapper. It was a fork of VS Code, Microsoft's open-source editor, with AI bolted in, and every AI request routed out to somebody else's model: Claude and GPT. They didn't build the editor. They didn't build the model. They built the thin, specific layer in between — the part where AI finally felt native to writing code.
That "wrapper" became the fastest-growing B2B software company ever measured. $100M ARR in January 2025. $500M by June. $1B by November. $2B by February 2026 — zero to two billion in roughly three years, outrunning Slack, Zoom, and Snowflake at the same milestone. By mid-2026, 64% of the Fortune 500 were on it, and SpaceX agreed to buy the company for $60 billion.
Look closely at what Cursor did and didn't build. They found a gap — the existing editors weren't designed around AI, and the model labs weren't in the business of shipping editors — and they owned exactly that gap. Everything else, they rented. They paid Anthropic and OpenAI per token and let those companies eat the nine-figure model-training bill.
And here's the part that should land on your roadmap: Cursor only started building its own model, Composer, in late 2025 — after it had a billion dollars in revenue and a concrete reason. By then, every query routed to a third-party model was a line item bleeding straight out of gross margin. Building their own inference wasn't ego; it was a margin decision they'd earned the right to make. They rented right up until renting became the thing holding them back — and not one day sooner.
Before you go wrap an API and call it a company — the graveyard is full of these.
Jasper started almost identically to Cursor: a clean interface on top of OpenAI's API, pointed at marketing copy. It worked. It hit a $1.5 billion valuation in about two years. Then ChatGPT got good enough, the thing Jasper wrapped became something anyone could use directly for twenty bucks a month, and Jasper's revenue reportedly cratered by more than half. The wrapper was the whole company, and the wrapper had no floor underneath it.
That's the real lesson, and it's subtler than "wrappers good" or "wrappers bad." Thin wrappers — the ones that are nothing but a prompt and a logo — see brutal churn, often shedding most of their users inside 90 days. The model provider can rebuild them in an afternoon, and your only remaining lever is price, which you will lose, because you're paying that same provider for the tokens.
The difference between Cursor and Jasper isn't that one wrapped and one didn't. They both wrapped. The difference is what they built underneath the wrapper. Cursor stacked up the things that don't ship free with an API key: deep codebase indexing, the editor workflow, enterprise trust, proprietary usage data, eventually its own inference. Jasper stacked up a nice interface. When the ground shifted, one had roots and one had a logo.
So the niche, by itself, isn't enough. The niche is where you start. The real question is whether — once you're sitting in that gap — you can grow something underneath you that the company you're renting from can't trivially copy.
The strategy was never "never build." It's "build the one thing, rent everything else, and stay honest about which is which." Four ways to keep that line clean:
The best niches sit in the seam between giants. The model labs don't want to build your vertical's workflow; the infrastructure vendors don't want to own your customer relationship. Cursor lived in the gap between "editor" and "model." Your gap is the job that's too small for the platform above you and too specialized for the generalist tools beside you.
Auth, payments, email, hosting, the base model — if it works the same for you as for everyone else, it isn't your product, and building it is a tax. The test is one question: would a customer ever choose you because of how you built this layer? If not, rent it. Nobody picks a tool for its bespoke login screen.
Renting has a crossover point. Cursor rented models until per-token costs ate their margin at a billion in revenue — then building paid off. Managed auth is a steal at 50,000 users and gets ugly at fifty million. Build the thing you're renting when, and only when, the spreadsheet says renting is now the more expensive option. Not before, on a hunch.
Owning a gap is the start, not the moat. The defensible part is what accumulates beneath it that your vendors can't copy: proprietary data, workflow lock-in, domain trust, the integrations nobody else bothered with. If everything you have evaporates the day the model provider ships one feature, you don't have a company — you have a countdown.
Building everything feels like ambition. Usually it's overhead wearing ambition's clothes. The companies that won this cycle didn't build more — they built less, on purpose, and spent all their effort on the one gap that was theirs to own.
Find the gap. Rent the rest. Put roots down where it counts. The model labs spend nine figures so you don't have to; the auth vendors carry the breach liability so you don't have to. Your job is the thin, specific layer that's worth building — and the discipline to leave the rest alone.
Build less. Own the part that matters. It's the boring framing — and, like most boring framing, the one the numbers keep backing up.