Explain how self-hosting changes the cost curve and where the crossover point sits.

Ringbay pays about nine cents every time its AI agent picks up a call. That price never moves. What moves is how many calls come in, and there is one exact number where paying by the call stops being cheaper than owning the machine.

How to answer this, stage by stage

Nobody is grading whether you know the word "GPU." They are grading whether you can write two real cost lines and point at the exact spot where one beats the other.

Let's learn

What does it actually cost to stop renting a voice model and start owning one?

Ringbay is a phone line a small business can forward its main number to. An AI voice agent picks up, answers common questions, and either resolves the call or routes it to a person, or takes a message if nobody's free.

At launch, Bellrock ran Ringbay entirely on a hosted API, a company that rents out its speech and language model by the call. That was clearly the right call. Ringbay had four thousand calls a month, and the whole bill came to three hundred sixty dollars. Building and running Bellrock's own machine for that would have been a strange way to spend money nobody had yet.

Eighteen months later, Ringbay handles about sixty two thousand calls a month, growing about nine percent every month as more small businesses sign on. The bill grew right along with it, and it's now about five thousand five hundred eighty dollars a month. Still affordable. Still approved without much of a second look.

The turn: five thousand five hundred eighty dollars a month is not the problem. The problem is that nobody ever wrote down the number of calls where paying by the call stops being the cheaper choice, so the bill can keep looking fine for a long time after it's stopped being the right decision.

At its worst, Ringbay keeps growing on the same path for another year or two, well past the point where self hosting would have been cheaper, and every month past that point is money spent on rent that could have gone toward owning the machine outright. At four hundred thousand calls a month, that gap is about seventeen thousand dollars, every single month, quietly.

What I'd leave alone: self hosting is a decision about where the model runs, not about the model itself. The eval set Ringbay grades every call against, and the bar a call has to clear to count as handled well, stay exactly the same whether the answer comes from a rented machine or an owned one. Moving the wiring is not a reason to touch the wiring's judgment.

The lesson: a decision that was cheap and correct on the day it was made doesn't stay correct just because nobody revisits it. It goes wrong quietly, at a specific number of calls, and the only way to catch that number is to write it down before you need it.

Now here is the same thing as a story

Read the short version above when you want the shape of the answer. Read this when you want to feel why a bill that kept looking fine still went wrong.

Ozren Bremseth can tell you Ringbay's monthly API bill down to the cent without opening a dashboard. He built the call routing pipeline himself, every hop from a ring landing on Bellrock's number to a transcript closing out.

He joined a month before Ringbay launched, back when it handled four thousand calls from four small businesses testing it out. The hosted API was obviously right. Three hundred sixty dollars a month, no machine to babysit, no reason to think twice. For the first year, the good months were genuinely good. Volume climbed, the bill climbed with it, exactly the way a rented, pay by the call service should behave, and every month finance approved the line item without a meeting.

The habit that faded wasn't dramatic. First, the monthly infra review folded Ringbay's number into a bigger line called vendor costs, general, instead of giving it its own row. Then Rozalind, who signed off on it, stopped opening the breakdown behind that number, since the total always came in close to what she expected. Then Ozren stopped checking the bill against anything except the bill from the month before, because a number that keeps agreeing with itself stops feeling like something worth opening the hood on.

It came back in an ordinary budget meeting. Rozalind was scanning line items out loud and got to Ringbay's. "We're basically renting an apartment forever on this thing," she said, half joking. "At some point people just buy the building." Nobody in the room treated it as urgent. It was one sentence, said in passing, on the way to the next line item.

Ozren couldn't stop thinking about the apartment line. That weekend, about six hours at his kitchen table, he wrote out both equations for real instead of trusting his gut. Paying by the call: nine cents times volume. Owning the machine: fifteen thousand dollars a month flat, plus about a cent a call. He solved for where they crossed. A hundred eighty seven thousand five hundred calls a month. Ringbay was at sixty two thousand. Growing nine percent a month, that crossing was about thirteen months out, and the decision to keep paying by the call had already gone eighteen months without anyone naming that number.

The real cost was never this month's bill. It was that folding Ringbay into vendor costs, general made the one number that mattered invisible inside a category built to shrink, not grow.

The decision Ozren would take back happened in a much smaller room, in Ringbay's first week, four people around a table deciding how to serve the model. Running everything through the hosted API wasn't a compromise. At four thousand calls a month it was obviously correct, and nobody in that room was wrong to pick it. What nobody did was write a number on the whiteboard and say, revisit this once we pass it.

Run that budget meeting again with one change: a monthly job that compares Ringbay's actual call volume against the crossover point, not a person's memory of a joke about apartments. At month thirteen, the month volume actually crosses a hundred eighty seven thousand five hundred, the alert fires on its own, and Ozren already has the self hosted build tested and ready, because he started it the week he ran the numbers, not the week the bill finally looked wrong.

One design waited on a person to notice a bill that kept agreeing with itself. The other watches the one number the whole decision actually turns on.

What I'd tell myself, back in that first small room: cheap and correct forever are not the same thing. Approving three hundred sixty dollars a month was easy. It just meant nobody ever wrote down the number where that stops being true.

BOUND: the two lines Ozren finally wrote down

Not a story question wearing a framework's clothes. This is an estimation problem, and BOUND is what turns a rented apartment joke into a number you can actually act on.

The honest answer was never one number. It runs from about 150,000 to 225,000 calls a month, with Ringbay sitting at 62,000 today, about a third of the way to the low end of that range.

Three things worth stating directly, since this is where the real judgment sits. The alternative Ozren considered, and rejected, was self hosting from day one instead of waiting for Ringbay to reach the crossover volume, one fewer decision to revisit later. It lost because paying the fixed fifteen thousand dollars a month while volume sat between four thousand and sixty two thousand calls would have cost far more than the API, for well over a year, before the fixed cost ever earned itself back. The AI-specific failure worth naming by name is losing the vendor's automatic model upgrades: a hosted API keeps improving underneath its price without Bellrock lifting a finger, but a self hosted model only gets better when someone at Bellrock decides to swap it in, and nothing stops that swap from quietly getting worse instead of better. The guardrail is an eval gate: any self hosted model version has to clear a scored pass rate on Ringbay's own held out set of real call transcripts before it goes live, not a vibe check from whoever built it. And the trade-off worth saying out loud is quality and latency against cost: self hosting past the crossover point lowers the cost per call, but it trades away the hosted vendor's elastic headroom, the ability to absorb a real spike, a product launch, a local event flooding a business's phone line, without anyone at Bellrock provisioning for it. Bellrock accepted a capped, lower monthly cost and its own reliability engineering, in exchange for giving up that headroom above whatever capacity the self hosted machine is actually sized for.

And if you want to be sure it really works, try it somewhere else

Same two lines, a claims photo instead of a phone call, and this time the lever that swings the crossover isn't the API's price at all.

Claimsight is a photo triage tool Halkirk built for insurance claims. An adjuster's assistant uploads a photo of car or property damage, and Claimsight scores how much of the claim a human still needs to look at closely. Kalani Enescu runs model operations for it.

The build-up: Claimsight handles about 18,000 claim photos a month. The hosted vision API costs more per call than Ringbay's phone conversations do, about 22 cents a photo, since a vision model reads far more detail per request than a short voice turn. Self hosting a vision model needs a bigger, pricier GPU: a fixed cost of about $22,000 a month, plus a marginal cost of about 3 cents a photo.

Solve the same equation and Claimsight's crossover lands at about 115,800 claim photos a month, roughly 116,000. Claimsight is well under that today. But unlike Ringbay, the number most likely to move that crossover isn't a vendor's per-call price cut. It's the rental price of the GPU itself, which swings on its own schedule, independent of how many claims come in that month.

Same method, different lever: for Ringbay, the number to watch was call volume against the API's per call price. For Claimsight, it's not volume alone. It's what a vision GPU costs to rent this quarter, since that price can move even while claim volume stays flat.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: two lines, price times volume against fixed plus a small marginal cost, solve for where they cross, check it against something real.
Cost: there's no budget this quarter for both the self host migration and the push to sign new customers. Growth wins. The migration waits, since Ringbay sits at about a third of the crossover volume.
The model got better, for real: say the hosted API's model gets a real quality upgrade for the same price. That raises the bar the self hosted model has to clear, it doesn't change the crossover math itself, since self hosting means giving up automatic access to that upgrade going forward.

Where people run it wrong.
They self host the moment growth feels exciting, instead of at the volume where it's actually cheaper.
They compare list prices only, and forget the self host marginal cost that's still there after the fixed cost, the power, the bandwidth, the carrier line itself.
They set a calendar reminder to revisit hosting, instead of a volume triggered check, so it only gets looked at when someone remembers, not the month it's actually true.

How to use it live. Say the real question out loud before quoting a number: "before I give you a crossover point, I want to say the two equations it comes from." That buys a beat to think instead of reciting a number nobody in the room could rebuild from scratch.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0