Building hardware in Detroit: how we shipped 4,000 industrial sensors

The contract-manufacturing decision almost killed us. Through summer 2024 I tried to find a US-based contract manufacturer who would take a 5,000-unit run at our margins. Six firms quoted us. Five of those quotes were unworkable; the sixth was workable but required a $480k upfront tooling commitment we couldn't afford.

In September 2024 I flew to Shenzhen for two weeks. I had three meetings a day with contract manufacturers a friend had introduced me to. The cost-per-unit quotes I got were 40-60% lower than the US options. The tooling commitments were a quarter the size. The risk was real — supply chain, IP, communication — and not abstract.

§The four-hour board meeting

I came back and called an emergency board meeting. Two board members wanted us to stay US-only and slow growth. One wanted us to go to Mexico. Aleksandr wanted Shenzhen. I was unsure. We talked for four hours. The eventual decision was Shenzhen with a structured risk-management plan: a US firmware team, US-based final test and packaging, on-site QA in Shenzhen for the first two production runs, and a parallel Mexican contract manufacturer ramping in the background for redundancy.

The first Shenzhen production run shipped in February 2025. Yield was 94% on 1,200 units. The second run, in April, was 97% on 2,400. As of April 2026 we've shipped about 4,000 cumulative units to 38 industrial customers. The redundancy strategy was the same one Stilt Robotics later adopted across Tijuana and a backup line. Revenue is ~$680k from hardware sales and ~$320k from cloud subscriptions, total roughly $1M ARR, with a much steeper growth curve now that we're not capacity-constrained.

The actual product breakthrough — the thing that compounded after the manufacturing problem was solved — was a feature we'd dismissed twice during prototyping. A maintenance manager at a paper mill in Wisconsin asked us in October 2024 whether the sensor could detect not just bearing failures but pump cavitation, which has a different vibration signature. Aleksandr initially said it would require new firmware. The customer said: "Then write new firmware. We'll pay $40,000 for it."

We wrote the new firmware in eleven days. We shipped it as an over-the-air update via AWS IoT Device Management to every sensor in the field, not just the Wisconsin customer. Within sixty days, 14 of our 38 customers had upgraded their cloud subscription to a tier that included pump cavitation detection. The new tier was $24/unit/year instead of $9. Average revenue per sensor effectively doubled.

The broader pattern: the hardware was the same hardware. The firmware updates kept changing what the same hardware could do. We started thinking of every sensor as a software-upgradeable platform, not a fixed product. That mental shift changed our entire roadmap.

§From 4,000 to 19,000 units

The fourteen months after we crossed unit number 4,000 were when manufacturing scale stopped being an art project and became a system. We shipped unit 19,000 in March 2026. Per-unit COGS dropped from $84 to $51, mostly from a switch to a single SMT line at our contract manufacturer Jabil in Auburn Hills and from re-spinning the board to use a single STM32U5 instead of the STM32L4 plus nRF52 combo we'd shipped in v1. Yield climbed from 88.6% to 97.4%.

We made one hire that ended six months of operational pain: Marcus, a former production engineer at Tesla's Sterling Heights body shop, who joined as VP of Manufacturing in mid-2024. He brought a quality-discipline framework — daily Pareto chart reviews, weekly walkthroughs of every reject — that I'd seen on factory floors but never run myself. Field failure rates fell to 0.31% per unit-year, which is the metric our Ford, Stellantis, and GM Tier-1 customers actually buy on. The "buy where credibility lives" pattern echoes what the Stilt Robotics warehouse story in this cohort found — that operator-credentialed hires do work no marketing budget can replicate.

1. 1
   Hand-assembly is a phase, not a strategy.

   If your unit economics depend on someday switching to contract manufacturing, switch sooner than feels comfortable. We waited six months too long.

2. 2
   Visit the contract manufacturer in person.

   Two weeks in Shenzhen taught me more about our actual cost structure than six months of remote conversations. Hardware founders cannot run their supply chain over Zoom.

3. 3
   Redundancy across geographies is not paranoia.

   We pay slightly more by running parallel production in Shenzhen and Mexico. That redundancy paid for itself the first time a single customs delay would otherwise have stopped shipping.

4. 4
   Firmware is the real moat in industrial IoT.

   The hardware will get cloned. The accumulated library of detection algorithms — and the over-the-air update channel into deployed sensors — is what makes you durable.

5. 5
   The customer who offers $40k for a feature is telling you the product roadmap.

   Listen to that customer. Ship the feature. Then ship it to everyone, because the next four customers will pay too.

What this looked like in practice.

The thing nobody tells you about *building hardware in Detroit* is that the talent pool is the moat. Every person on our manufacturing team has shipped real product before. The cost of hiring is higher than San Francisco for adjacent software roles and lower than San Francisco for production engineering, by a factor of three.

The thing nobody tells you about *unit cost reduction* is that the line item is rarely the part — it is the assembly time. Our $33 per-unit COGS reduction broke down as $7 in BOM savings and $26 in cycle-time reduction.

The thing nobody tells you about *contract manufacturers* is that they will quote you a price and then a schedule, and the schedule is what kills startups. We built contractual penalties for missed lead times into our Jabil master supply agreement after our first launch slipped four months.

The thing nobody tells you about *Tier-1 automotive customers* is that they will not buy you on innovation — they will buy you on PPAP and IATF 16949 certification. We invested $310k and nine months in compliance before our first revenue dollar from a Tier-1.

The thing nobody tells you about *firmware updates in industrial IoT* is that the field cannot be assumed to have connectivity. Our update path uses LoRaWAN failover with a USB-C technician override, and we tested it in three different cold-storage facilities before we shipped.