Build it or buy it · the honest math
Should you build this yourself? Maybe. Here are the numbers.
A detector costs $10, and your first one is free. Over 10,000 units, it costs about $1.10. Here is where that money goes, what it would cost your team to build the same thing, and the point at which building it yourself is the right call. Some companies are past that point. Most of the ones we talk to are not.
Where your $10 goes
We put the split on the page because the rest of this comparison only works if you can check it.
Over 10,000 units: volume pricing, about $1.10 a unit
Ship more than 10,000 units and we drop the software charge. You pay the ~$1 of hardware, manufacturing and shipping, plus $0.10 a unit for us. That applies to every unit in the order, not just the ones past 10,000.
The hardware dollar does not go away if you build it in house; you still need a microphone, a microcontroller and a board. The software portion is what building it yourself can save you, about $9 per paid unit at the standard price and ten cents per unit on orders over 10,000. So the real question is whether your engineering costs less than what you would save on every unit you ship.
What building it in house costs
A voice stop that works in a demo is a weekend project. One that hears “stop” across a loud room, ignores “don’t stop” and the robot’s own speaker, runs on a microcontroller and fails toward stopping takes a few months. This estimate assumes one engineer who already knows embedded machine learning and doesn’t have to learn it on the job.
| Work | Engineer-weeks | Cost |
|---|---|---|
| Speech data. Many speakers, accents and distances saying the word; synthetic voices; the hard negatives (“don’t stop”, “bus stop”) | 3–5 | $10,500–17,500 |
| Model. Train it, shrink it to fit a microcontroller, tune false triggers against hours of real floor noise | 2–4 | $7,000–14,000 |
| Firmware. Audio capture, feature extraction, on-device inference, a latched stop output, USB | 3–5 | $10,500–17,500 |
| Board. Schematic, layout, two prototype spins, plus $1,500–3,000 in boards and parts | 2–4 | $8,500–17,000 |
| Enclosure and mounting. Plus $500–1,500 in prints and tooling | 1–2 | $4,000–8,500 |
| Validation. A test rig, distance and noise sweeps, latency, plus about $500 of equipment | 2–3 | $7,500–11,000 |
| Total | 13–23 | $48,000–85,500 |
Priced at $3,500 per engineer-week: about $175,000 a year fully loaded (salary, benefits, overhead), over 50 working weeks.
Call it $50,000 to $85,000, three to six months, for one language. That doesn’t include keeping it working: firmware fixes, new languages, a new microphone when a part goes end-of-life, or re-testing after each change. It also doesn’t include the safety case, which is the part your safety officer will actually ask about.
Where building it starts to pay
Buying costs $10 a unit after the first, and about $1.10 a unit over 10,000. Building costs your engineering up front plus the same hardware dollar per unit. At volume, each unit you build yourself saves you about ten cents, so you need to ship roughly ten units for every dollar of engineering before building it pays off. At $50,000 of engineering, that is about 500,000 units. At the standard price, building is also cheaper from 5,557 through 10,000 units. The bulk discount makes buying cheaper again from 10,001 units until the long-run break-even.
Building costs a lot up front, so its cost per unit starts high and falls as you ship more. It never drops below the ~$1 of hardware. Building can be cheaper before the bulk discount starts. On orders over 10,000 units, our price drops to $1.10 per unit; building becomes cheaper again above about 500,000 units. Shaded areas show where buying costs less.
- Buy from us
- $49,990
- Build in house
- $55,000
- Long-run break-even
- 500,000 units
| Units shipped | Buy from us | Build in house ($50k) | Cheaper |
|---|---|---|---|
| 1,000 | $9,990 | $51,000 | Buy, by $41,010 |
| 5,000 | $49,990 | $55,000 | Buy, by $5,010 |
| 10,000 | $99,990 | $60,000 | Build, by $39,990 |
| 20,000 | $22,000 | $70,000 | Buy, by $48,000 |
| 100,000 | $110,000 | $150,000 | Buy, by $40,000 |
| 500,000 | $550,000 | $550,000 | About even |
| 1,000,000 | $1,100,000 | $1,050,000 | Build, by $50,000 |
Buy is $10 a unit after a free first one, or $1.10 a unit for orders over 10,000. Build cost is engineering plus $1 of hardware per unit. That assumes your hardware costs what ours does. At small volumes it usually costs more, which moves the break-even point further out.
When you should build it yourself
Yes, building it in house can be the right call, and we would rather say so here than waste your time on a call. Build it if:
- You ship hundreds of thousands of units of one device. With volume pricing you save about ten cents a unit by building, so at $50,000 of engineering it takes about 500,000 units to come out ahead.
- You already have people who do embedded machine learning and have time on their schedule.
- The detector has to live on your own main board, sharing a processor you already pay for, so even our hardware dollar goes away.
If two of those are true, building it is probably cheaper, and we will tell you the same thing if you ask.
Who we built this for
Teams that want to add a voice stop without developing and maintaining a detector themselves. At 1,000 units, buying costs $9,990 versus $51,000 to build under this estimate. At 10,000 units, buying costs $99,990 versus $60,000 to build, so building may make more sense at that order size. Orders over 10,000 qualify for the $1.10 bulk price on every unit. At 100,000 units, buying costs $110,000 versus $150,000 to build. Use the calculator with your order size and engineering costs.
At that volume, a voice stop is a feature you need, not a product you want to own. Spend the engineering on the robot.
What these numbers assume
- The $1 / $9 split is approximate and averaged across a production run; it moves a little with volume and freight. Volume pricing is that same ~$1 plus $0.10.
- Engineering at $3,500 per engineer-week, for someone who has built this kind of system before. Learning on the job costs more.
- One language, one device, one board design. Each additional language or board adds to the in-house cost, but not to ours.
- No certification work on either side. Today’s units are uncertified evaluation units; see the terms.
Try one before you decide
The first unit is free. Put it on your robot, then do this math with real numbers.