Humane raised $230 million. Rabbit R1 sold 100,000 units. Both failed. Here's why the software mindset kills hardware companies — and how to survive the transition.
In April 2024, Humane shipped its first AI Pins to paying customers. Five months later, more units were being returned than sold. By February 2025, HP bought the company's assets for $116 million — roughly half of what investors had poured in. Ten days after the deal closed, every AI Pin in the wild was bricked remotely.
Rabbit R1 fared slightly better on paper: 100,000 units out the door. But the return rate followed a similar arc. By late 2025, the company couldn't make payroll. Employees went on strike. The device that had been CES 2024's breakout hit was a cautionary tale.
Combined, these two companies burned through over $300 million. Both were founded by experienced tech operators. Both had excellent pre-launch narratives. Both built products that users didn't want.
The pattern isn't coincidence. It's a structural gap between how software founders think and what hardware actually demands.
The Iteration Gap: Why "Move Fast and Break Things" Doesn't Work
In software, a bad release costs you a deploy rollback. In hardware, a bad design decision costs you a mold.
Here's what one iteration cycle actually looks like when you're building physical products in Shenzhen:
| What changes | Time | Cost |
|---|---|---|
| PCB redesign + new prototype boards (5pcs, 4-layer) | 5-7 days | $400-600 |
| New CNC-machined enclosure (single unit) | 3-4 days | $80-150 |
| New 3D-printed test parts | 4-8 hours | $15-50 |
| Injection mold modification (aluminum soft tool) | 2-3 weeks | $2,000-4,000 |
| Full design revision cycle (PCB + enclosure + test) | 10-14 days in Shenzhen; 6+ weeks in the West | $1,000-5,000 per cycle |
Now multiply that by the number of iterations your software team is used to. Most SaaS startups ship multiple times per day. If you approach hardware with the same cadence, you'll burn through your runway before reaching EVT (Engineering Validation Test).
The rule of thumb: plan for exactly 2-3 hardware iterations, maximum. Every cycle beyond that eats months and tens of thousands of dollars. Invest heavily in design validation before sending files to the factory. Simulation, 3D printing, and CNC prototypes cost a fraction of a mold mistake.
The DFM Tax: What $300 Today Saves You Tomorrow
DFM — Design for Manufacturability — is the most skipped step in hardware development, and the most expensive one to skip.
Most software founders have never heard of it. They hand a polished CAD file to a factory and expect it to work. It doesn't. The factory can make exactly what you drew — including the part that's impossible to eject from the mold, the PCB trace that's too thin for volume production, and the component placement that requires a human hand where a machine should be.
The cost curve is brutal:
| Fix stage | Cost multiplier |
|---|---|
| CAD phase (before any tooling) | 1x ($100-1,000) |
| After mold is cut | 10-50x ($5,000-50,000) |
| During pilot production | 50-200x ($50,000-200,000) |
| During mass production + recall | 500x+ ($500,000+) |
A $300 DFM review from your contract manufacturer — which most offer for free on serious projects — can prevent a $30,000 mold redo. Every factory in Shenzhen has a wall of abandoned molds from founders who skipped this step.
DFM checklist your manufacturer should review:
- Draft angles ≥ 1° for injection-molded parts; uniform wall thickness (2-3mm for ABS)
- PCB trace/space ≥ 6mil; vias > 0.3mm; no 0201 components unless absolutely necessary
- Component accessibility for automated pick-and-place assembly
- Test points accessible for production-line testing fixtures
Your BOM Won't Be $50
Software founders consistently underestimate BOM (Bill of Materials) costs. The instinct is to price components at DigiKey single-unit quantities, add 20%, and call it a day.
Reality is more complicated:
- Component costs at scale drop significantly, but you won't hit those thresholds at 1,000 units
- Assembly labor for complex products can be $5-20 per unit in Shenzhen
- Packaging alone can run $2-8 per unit for consumer-grade boxes
- Testing and QA typically adds 5-10% to BOM
- Scrap and rework during pilot runs can consume 5-15% of units
- US import costs in 2026 add roughly 45% on top of factory-gate price (40% tariffs plus customs brokerage)
A useful sanity check: your retail price should be 4-5x your BOM cost to leave room for distribution, marketing, returns, and margin. If your BOM is $50, you need to sell for $200-250 — not $99.
Certification Isn't a Final Step
The Humane AI Pin shipped to customers in April 2024. It was FCC and CE certified. But certification is not a checkbox at the end of development — it's a constraint that shapes your design from day one.
Key certification realities for AI hardware:
| Certification | Cost | Timeline | Notes |
|---|---|---|---|
| FCC + CE pre-scan (EMC lab, half day) | $1,500-2,500 | Book 2 weeks ahead | Do this before cutting molds |
| FCC + CE full certification | $15,000-40,000 | 4-8 weeks | Requires PVT production samples |
| FCC + CE-RED (wireless products) | $20,000+ | 6-10 weeks | Required for Bluetooth/WiFi/cellular |
The important thing: if you fail EMC testing at DVT stage, you're looking at a PCB redesign. If you fail after molds are cut and production tooling is built, you're looking at a disaster. Always pre-scan before committing to tooling.
The Real Decision: Build a Hardware Company, or Don't
Here's the uncomfortable truth most AI founders need to hear: hardware is not a feature extension of your software product. It's an entirely different business.
Software companies that successfully launch hardware (Ray-Ban Meta: 7 million units in 2025, revenue tripled YoY) don't succeed because they're good at software. They succeed because they partner with companies that are good at hardware. Meta didn't build the glasses — EssilorLuxottica did. Meta provided the AI stack, the brand distribution, and the software ecosystem. The hardware layer came from a company that's been manufacturing eyewear for decades.
If you're an AI founder looking at hardware, the question isn't "can we build this?" The question is: "do we have a partner who already builds this category of thing, or are we attempting to learn injection molding, PCB assembly, EMC compliance, and supply chain management from scratch while also shipping AI features?"
The latter path is what Humane and Rabbit attempted. The former is what Meta did. The results speak for themselves.
The Numbers, for Your Spreadsheet
| Milestone | Time | Budget Range |
|---|---|---|
| POC (proof of concept, on-site in Shenzhen) | 1-2 weeks | $500-3,000 |
| EVT (engineering validation, 10-30 units) | 3-4 weeks | $4,000-12,000 |
| DVT (design validation, 50-100 units, soft tooling) | 6-10 weeks | $15,000-35,000 |
| PVT (production validation, 200+ units, hard tooling) | 4-8 weeks | $30,000-80,000+ |
| Total, concept to pilot production (medium complexity) | 4-6 months | $70,000-120,000 |
These are Shenzhen prices with an experienced local partner. Double the timeline and triple the cost if you're doing this remotely from California.
"Hardware is not a feature extension of your software product. It's an entirely different business."
Ready to make the leap?
We've helped 50+ AI teams navigate the software-to-hardware transition — from early-stage concept validation all the way to mass production in Shenzhen. Whether you have a working prototype or just a napkin sketch, our team can assess your hardware feasibility and map the path forward.
Book a Free 15-Minute Project Assessment →
No commitment. No pitch deck. Just practical feedback from people who've done this before.
Next in this series: Three AI Hardware Launches, Three Lessons → — What Humane got wrong, what Ray-Ban Meta got right, and what Meta Orion tells us about the future.
Have an AI software demo that needs a hardware path?
Share the current stage. Aixumo will help you identify the hardware route, productization risks, and practical next move before you spend money on the wrong supplier conversation.