Humane AI Pin: $230M raised, $116M acquired, 7,000 active devices. Rabbit R1: 100K sold, can't make payroll. Ray-Ban Meta: 7 million units, revenue tripled. What separates the winners from the dead?
There's no shortage of AI hardware failure postmortems. But the discussion usually focuses on product features — was the AI good enough? Did the device overheat? Was the latency tolerable? These are symptoms. The real pattern across every failed AI hardware launch is deeper: execution failure at the supply chain and go-to-market level.
Let's look at three launches, what actually happened, and what founders should learn.
Case 1: Humane AI Pin — The $230M Demo That Couldn't Ship
What they raised: $230M from Tiger Global, SoftBank, OpenAI, and others. What they sold for: $116M to HP (February 2025). What happened next: All AI Pins bricked on February 28, 2025.
Humane had everything a startup could want. Two ex-Apple founders. Five years in stealth. A TED stage reveal that went viral. A Coperni runway moment with Naomi Campbell. TechCrunch exclusives at every funding round. Even the Series C cap table looked like an AI hall of fame.
What they didn't have was a product that worked, a supply chain that could deliver, or a go-to-market strategy that survived launch week.
The supply chain failure: Humane manufactured a complex wearable with cellular connectivity, a projector, cameras, and a Snapdragon processor — in relatively small volumes, with no existing manufacturing relationships to lean on. Every component was custom or semi-custom. The BOM was rumored to be north of $400 for a $699 retail device, leaving almost no margin after distribution and marketing. When MKBHD's April 14, 2024 review called it "the worst product I've ever reviewed," Humane had no installed user base to push back, no second product to pivot to, and a unit economics model that couldn't survive a return wave.
The returns cascade: By August 2024, returns were exceeding new sales. Internal data leaked to The Verge showed roughly 7,000 active devices in the wild. The math was brutal: each returned unit represented a $400+ BOM loss, plus reverse logistics, plus the customer acquisition cost already spent. Humane couldn't iterate fast enough to fix the software, and couldn't afford the hardware revision that would have been needed.
The lesson: A hardware company cannot survive its launch reviews if the supply chain and unit economics weren't designed for iteration. Humane bet everything on version 1 being good enough. It wasn't — and there was no version 1.1.
Case 2: Rabbit R1 — CES Hype, Consumer Disappointment
What they sold: 100,000 units at $199 each. What happened next: Mass returns, payroll crisis, employee strikes by late 2025.
Rabbit R1 was the darling of CES 2024. A bright orange pocket device that promised to use AI to navigate apps on your behalf — the "Large Action Model" that would book flights, order food, and send messages without you touching a screen. 50,000 units sold in the first two weeks.
The reality was less magical. Reviews found that the R1 was essentially an Android app running on dedicated hardware. The "action model" could barely navigate any real-world service. Users discovered they could sideload the same software onto their phones, removing any reason to carry a separate device.
The manufacturing trap: Rabbit manufactured 100,000 units based on pre-launch hype, not validated demand. When returns started flowing, the company was sitting on inventory it couldn't move. The unit economics were worse than Humane's: $199 retail price with a custom-designed hardware platform meant razor-thin margins even if everything went perfectly. It didn't.
The software-hardware disconnect: Rabbit's core value proposition was software (the action model). The hardware was just a delivery vehicle. But the software wasn't ready, and once it shipped, the hardware couldn't be updated. Unlike a SaaS product that can push fixes overnight, Rabbit's 100,000 units in the field were stuck running version 1.0 until they connected to WiFi and downloaded an update — which many users never bothered to do because they'd already returned the device.
The lesson: Don't manufacture at scale until you've validated that version 1.0 actually works. Rabbit had the resources to do a smaller pilot run (1,000-5,000 units), gather real user feedback, and iterate. Instead, they went straight to 100K based on CES hype. The manufacturing lead time meant they were committed to a product their users had already rejected.
Case 3: Ray-Ban Meta — The Quiet Success Story
What they sold: Over 7 million units in 2025, revenue more than tripled YoY. What they did differently: Partnered with a company that already knew how to make glasses.
The Ray-Ban Meta story is genuinely remarkable, and remarkably under-discussed in AI circles. The first generation (Ray-Ban Stories, 2021) was a modest seller. The second generation launched in late 2023 with better cameras, better audio, and Meta AI integration. By 2024, sales were climbing. By 2025, they exploded — 7 million units.
The secret isn't better AI (though the AI is good). It's supply chain maturity.
What Meta got right:
- Partner with an existing manufacturer: EssilorLuxottica has been making glasses for decades. They have the factories, the distribution, the retail relationships, and the quality control. Meta didn't have to learn injection molding or lens grinding. They focused on what they do best — AI, software, and ecosystem integration.
- Start with a product people already buy: Ray-Ban Wayfarers sell millions of units every year without AI. The Meta version looks almost identical. Customers aren't buying "AI glasses" — they're buying Ray-Bans that happen to have AI features. This is fundamentally different from Humane and Rabbit, which asked customers to adopt an entirely new form factor and behavior.
- Iterate on a proven platform: Version 2 built on version 1's manufacturing lines, supplier relationships, and retail channels. The molds existed. The assembly processes were documented. The supply chain team knew the lead times. Each new generation gets cheaper to produce, not more expensive.
- Software updates without hardware changes: When Meta improves the AI assistant, every existing pair of glasses gets better. No hardware revision needed. This is the only viable model for AI hardware — the hardware is a platform for continuous software improvement, not a one-shot product.
The lesson: You don't win AI hardware by building better hardware. You win by having better manufacturing partners, better distribution, and a product roadmap where the hardware enables software improvement rather than constraining it.
The Execution Framework
There's a clear pattern across these three cases. Here's what separates the winners:
| Factor | Humane AI Pin | Rabbit R1 | Ray-Ban Meta |
|---|---|---|---|
| Manufacturing partner | Built from scratch | Built from scratch | EssilorLuxottica (100+ years) |
| Supply chain | Custom components, small volumes | Custom hardware, 100K batch | Existing eyewear lines, millions of units |
| Product category | New form factor | New form factor | Existing category (eyewear) |
| Unit economics | $400+ BOM, $699 retail | Thin margins, $199 retail | Healthy margins at scale |
| Iteration path | No v2 possible without new funding | Software locked to hardware | Continuous software updates |
| Distribution | Direct online only | Direct online only | 15,000+ retail locations globally |
| Outcome | $116M acquisition, product dead | Can't make payroll | 7M units, revenue tripled |
The lesson isn't "have more money" — Humane had $230M and still failed. It's "understand which parts of the stack you should build and which parts someone else has already solved."
What This Means for AI Founders
If you're an AI company considering hardware, there are exactly two paths that work:
Path A: The Meta model. Partner with an existing manufacturer in your target category. Your job is AI and software. Their job is everything else. You're adding intelligence to a product that already exists in the world.
Path B: The Dyson model. Commit to hardware as your core competency from day one. Build an in-house hardware team with manufacturing experience. Accept the longer timelines, higher costs, and fewer iterations. This only works if hardware is your business, not a feature of your software business.
The Humane/Rabbit path — raise a lot of money, build hardware from scratch without manufacturing DNA, ship v1 to consumers, hope it's good enough — is not a path. It's a $300M experiment that's been run twice now, with the same result both times.
"You don't win AI hardware by building better hardware. You win by having a partner who already builds hardware."
Don't learn these lessons the hard way.
We've managed 50+ AI hardware projects through the exact phases described above — from EVT and DFM to PVT and shipping. Before you commit to tooling or sign a factory contract, get a second set of eyes on your plan.
Request a Free Project Evaluation →
Tell us where you are in the process. We'll tell you what you're missing.
Next in this series: The Founder's Guide to Shenzhen → — Real costs, real timelines, real advice from the factory floor.
Before you commit to tooling or factory selection, pressure-test the execution path.
Share where your project is now. Aixumo will help identify what is missing before samples, supplier discussions, or production preparation.