Describe how prioritization changes when the cost of a feature scales with usage rather than being fixed.

Simulglot is Quillbrace Media's engine for live translation: it reads a broadcast's audio and turns out real-time captions, or a fully dubbed voice track, in whatever language the viewer picks. Solomiya Krogh owns Simulglot's roadmap. Farideh Winship runs international distribution for Vantport Sports Network, and she is the one who has to decide, every quarter, which Simulglot feature is worth building next.

How to answer this, stage by stage

Nobody is grading whether you can say "usage-based pricing is risky" and sound careful. They're grading whether you can turn "the cost scales with usage" into an actual number, and a ranking that survives someone asking where that number came from.

Let's learn

This whole chain runs again for every viewer, every minute they keep the dub on. That's the part a fixed-cost feature never has to do.

Simulglot listens to a live broadcast's commentary and turns it into captions or a fully dubbed voice track, in a language the viewer picks from a menu, with a lag under two seconds.

Before AI dubbing, the only way to add a language to a live match was a human interpreter in a booth, roughly $400 an hour, and even then the translation ran twenty to thirty seconds behind the picture. Vantport could only afford that for a handful of marquee matches a season. Everything else stayed English-only, or captions-only if a fan happened to read the language.

With Simulglot, dubbing runs on every match, in any language Quillbrace has built, and the bill is no longer a booth and a headcount. It's a per-minute compute charge that scales with however many people actually use it.

Same word, "feature," two completely different cost shapes. Only one of them cares how many people show up.

Here's the turn. The extra mistakes a bigger crowd might cause were never the risk. The real risk is that the very thing that makes a usage-scaling feature look like a win, more people using it, is also the thing that can quietly eat its own margin, and a dashboard showing "reach up 3x" never says so on its own.

What it costs at its worst: run the same peak-load compute cost up by just three-tenths of a cent, from 1.1 to 1.4 cents a viewer-minute, and Simulglot's single biggest month of the year for that feature nets exactly zero. A cent more past that, and the month that was supposed to be the proof dubbing works becomes the month it loses money.

What I would leave alone: a usage-scaling feature nobody expects to go viral doesn't need any of this scrutiny. Simulglot's German dub track has a small, steady audience that has never once doubled in a week. If reach can't plausibly triple, the margin math barely moves, and building a cost ceiling for it would be solving a problem that doesn't exist yet.

The lesson: a feature whose cost is fixed only needs one number to get right: will enough people pay for it. A feature whose cost scales with usage needs a second number nobody thinks to ask for until it's too late: what does it cost at the reach you get if it actually works.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one when you want to feel what a single upset match nearly cost.

The broadcast ops room at Quillbrace Media only gets loud twice a year: launch week, and the week an underdog wins. Solomiya Krogh has owned Simulglot's roadmap for three years, and she can price a new language pair in her head faster than most people can open a spreadsheet, build cost, compute cost, expected reach, done.

Portuguese dubbing launched in the spring, for Vantport Sports Network. For most of a season, it was exactly the win everyone hoped for. Farideh Winship watched reach on dubbed matches climb steadily, and every month, margin climbed right along with it, six thousand dollars, give or take, month after month. Nobody re-ran the numbers past the first few months. Why would they? The math kept agreeing with itself.

Then Tempestade, a club nobody outside its own city had heard of three weeks earlier, won a quarterfinal it had no business winning, on a stoppage-time goal that ended up on every sports account by morning.

Farideh was thrilled. Reach on the Portuguese dub for Tempestade's next match tripled in a single week, 1.5 million viewer-minutes instead of the usual 500,000. She sent Solomiya a message that just said: this is the number we build the whole roadmap around.

Solomiya pulled the real numbers at month's end, out of habit more than worry. Revenue was up, sharply, $17,400 against a usual $10,000. She almost closed the spreadsheet there. Then she checked the cost line. $14,700, against a usual $4,000. A third of that new reach had landed past Vantport's monthly cap, priced at a rate Simulglot's own compute never got any cheaper to run just because the price to serve it had dropped.

Margin for Tempestade's viral month: $2,700. Less than half of what an ordinary calm month brought in, on more than double the revenue.

Here's the decision that traces back to a contract-signing meeting eight months earlier. Vantport's finance team had asked for a lower overage rate, worried a big month could blow their own budget. Sales agreed to 0.6 cents a viewer-minute past the cap, well below what it cost Simulglot to run the model, on the reasoning that overage would be rare. Nobody in that meeting asked what happens to Simulglot's own cost the exact week overage actually happens, because a genuinely popular match and a genuinely strained server both draw on the same overloaded capacity at the same time.

I would take that back. Not the discount itself, Vantport needed the budget protection to sign at all. What I'd change is pairing it with a hard ceiling on Simulglot's own peak compute cost, so a demand spike could never cost more per viewer-minute than the business could still make money serving.

Run Tempestade's match again, with that ceiling live. Overflow traffic past the cap now automatically drops to a lighter, slightly less rich TTS voice, capping compute at 0.95 cents a viewer-minute instead of drifting toward 1.1, or worse. Same tripled reach, same $17,400 in revenue. Margin: $4,050, not $2,700. And the real change isn't even that one number. It's that reach would now have to hit 5.3 times forecast, not 3, before the same month could go negative at all.

One design leaves the business exposed to whatever a viral moment happens to cost that particular week. The other puts a floor under it no matter how big the moment gets.

What I'd tell myself, sitting in that contract meeting: a discount that only shows up "in the rare case" isn't rare risk. It's risk you've agreed to pay for exactly when you can least predict the bill.

BOUND, or pricing a feature that gets more expensive exactly when it wins

Not a way to sound cautious about usage-based pricing. BOUND is what turns "the cost scales with usage" from a warning label into a real number, and a ranking that survives someone asking where that number came from.

The five moves, in order. The fourth one, the check, is the step a rushed answer skips, and it's the one this whole question is actually testing.

Caption styling never leaves the bottom-left corner, no matter what happens on a broadcast. Dubbing does, and the direction it moves depends on one number.

Four points on one line. The gap between the third and the fourth is three-tenths of a cent, and it's the whole difference between a good quarter and a bad one.

Three things worth naming directly, since this is where the real judgment sits. The AI-specific failure mode here is a silent margin collapse under demand: nothing breaks, no error appears, the model keeps translating fine, but the compute cost of serving a real spike in live inference quietly outgrows the revenue attached to it, and a reach chart alone never shows that happening. The guardrail is a per-viewer-minute cost ceiling: once inference demand nears real-time capacity, overflow traffic automatically drops to a lighter TTS voice tier instead of paying full peak compute cost for it. That's a real trade-off, accepted on purpose: viewers past the committed cap get a slightly less rich synthetic voice during a spike, not a worse translation, not a dropped stream, in exchange for a cost curve that can't run away. The alternative Quillbrace rejected was a hard usage cap, stopping the dub entirely once the monthly ceiling was hit. It lost because cutting a live translation off mid-broadcast, for real fans watching a real match, is a worse outcome for Vantport than the margin risk it would have solved.

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

Same five letters, a hospital instead of a broadcast booth. The soundtrack becomes a chest X-ray, and the viral match becomes flu season.

Different building, same shape of decision. A viewer-minute became a scan, and a viral match became a flu season.

Colworth Diagnostics builds SecondGlance, which gives a radiologist an AI second opinion on every chest X-ray before they sign off. Dr. Wilhelmine Applewhite, chief radiologist at Radnor Memorial Hospital, is choosing between two features too: a one-click structured PDF report formatter, built once for $12,000, and expanding AI second-reads to every scan, not just flagged ones, at a real compute cost per scan.

Run BOUND on it. Break it down: the PDF formatter's value is minutes saved per report, times reports filed, against a fixed build cost. The second-read expansion's value is scans analyzed, times the value of a caught finding, minus a real per-scan compute cost. Own the numbers: second-reads cost about 35 cents a scan to run normally, and the hospital pays Colworth 90 cents a scan for the service, a healthy margin, at a forecast of 20,000 scans a month. Use a range: flu season can triple scan volume to 60,000 in a month, and that's exactly the range that matters. Nail the sanity check: at forecast, margin is $11,000 a month. Tripled, if overflow scans route to a costlier on-demand GPU tier at 65 cents each instead of the usual 35, margin only rises to about $22,500, not the naively expected $33,000, because a third of that volume now costs nearly twice as much to analyze. Direction: the swing assumption is the per-scan compute cost under surge load, the same shape as Simulglot's peak-load number, just measured in scans instead of viewer-minutes.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: run the usage-scaling feature's margin at the reach it would get if it actually went viral, not the reach you forecast, and rank on that.
Cost: no budget to build the automatic downgrade this quarter. Ship a hard alert on the peak-cost number instead, and treat the swing assumption as a known, watched risk, not a solved one.
The model got better, for real: say Simulglot's compute cost per viewer-minute drops to 0.4 cents. The pattern doesn't disappear, it just moves. Re-run the same equation with the new number. A cheaper model lowers the floor. It doesn't make usage-scaling cost stop existing.

Where people run it wrong.
They rank by total addressable reach instead of margin at that reach.
They see revenue go up during a spike and call the feature validated, without ever checking whether cost went up faster.
They fix the risk by throttling the user instead of designing the cost curve itself to bend, the same mistake a hard usage cap makes.

How to use it live. Ask yourself, before naming any number: "does this feature's cost move when nobody's watching, or only when everyone is?" If the answer is the second one, reach alone is the wrong thing to rank it on, and that question alone buys real thinking time.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0