The last three years have handed us a lot of worked examples. A wearable raised a quarter of a billion dollars and sold ten thousand units. A battery company took thirteen billion and never made cells at rate. A children’s robot stopped speaking when its investor walked away.
None of these were run by fools, which is the part worth sitting with. Every one had good engineers, credible investors and a product that demoed well. They died anyway, and mostly of a small number of repeatable causes.
Below are the specific mechanisms by which hardware companies went from funded to finished between 2023 and 2026 (if any of the vocabulary here is unfamiliar, the glossary defines it), and what we would do differently. It is a coroner’s report, written with sympathy for the deceased.
Cause of death: shipping the demo, at a price nobody agreed to
Humane raised roughly $241 million, launched the AI Pin at $699 plus a mandatory $24/month, and sold about 10,000 of them for around $9 million in revenue. By August 2024, according to internal figures obtained by The Verge, returns were running ahead of new sales; only about 7,000 units were still in customers’ hands. In February 2025 HP bought the assets (the Cosmos software platform, the patent portfolio, most of the staff) for $116 million, and the Pins lost their servers on 28 February 2025.
Manufacturing was not the problem. The pins were built. The product did the thing in the keynote and not the thing in your hand: it was slow, it ran hot, it gave wrong answers, and it never had a reliable core loop.
Rabbit went the other way. The R1 launched into the same wave, sold about 130,000 units, and was pilloried when someone got its software running on a Pixel, which showed the hardware was optional. By September 2024, of roughly 100,000 buyers, about 5,000 were using it daily. But Rabbit kept shipping software weekly, and rabbitOS 2 arrived in September 2025 with a genuinely different interface. Both companies got the same hype and the same criticism. One iterated its way down to a smaller product that works, the other sold its patents.
There was a second flaw underneath the first, and it was in the price. That mandatory $24 a month was not an accessory to the $699; it was load-bearing, a phone plan attached to a device most buyers would not have kept for free. Bird ran the same play in a different category, insisting for years that the per-ride model would out-earn scooter depreciation, and filed Chapter 11 in December 2023. The shape repeats: a hardware price set by what the market will tolerate, and a cost structure that only closes if a recurring revenue line materialises at a rate nobody has evidence for. The subscription was an assumption inside a model, and no customer had signed it.
So before you take pre-orders, write down the single task your product must do better than a phone, and the measurable bar it must clear: latency, accuracy, success rate on the first try. Then run that bar with strangers, not colleagues, on the production intent build. If you cannot pass it, do not set a launch date. What you have is still a research project with a marketing budget attached. Delaying a launch is embarrassing for a quarter; returns outpacing sales is what closes the company.
Price it in the same pass. Build the landed-cost model before the industrial design is locked, and put the ugly lines in it: returns, warranty replacements, support minutes per unit per year, the cost of the cloud you just committed to for five years. Test the recurring revenue separately from the hardware, ideally by selling it to people who already own something else. Then check whether the hardware survives at a 0% attach rate for the first eighteen months. If it does not, the number on the box is a bet.
Cause of death: a product that dies when your servers do
Embodied’s Moxie was a $799 social robot for children. The company had raised about $19 million since 2016. In December 2024 its lead investor withdrew from a funding round, the company wound down, and because every meaningful function ran in the cloud, the robots stopped working. Embodied said it would “try” to refund recent orders. Families were left explaining to children why a companion had stopped talking.
The same story keeps repeating in this category. Spotify bricked the $90 Car Thing on 9 December 2024, which produced a class action in the Southern District of New York and, eventually, refunds. Belkin ended support for most Wemo devices on 31 January 2026, a line on sale since 2011, and the survivors were the ones set up with HomeKit or built on Thread, because those keep working without a cloud.
Regulators noticed. In November 2024 the FTC reviewed 184 connected products and found that 163 of them, 89%, published nothing about how long software support would last; staff could not find support information at all for two-thirds of the devices. The FTC’s position is that this can implicate the Magnuson-Moss Warranty Act and the FTC Act. In Europe the Cyber Resilience Act pushes the same way, requiring a declared support period of at least five years unless the product’s expected lifetime is shorter.
Decide at architecture time what your product does with no internet and no company behind it. Ask what a customer is holding if you die on a Tuesday. Local control for the primary function, a documented support window published on the product page, and a written sunset plan (final firmware that removes the cloud dependency, key material exportable to the owner) cost a few weeks of engineering when you design them in, and cannot be retrofitted once an administrator is running the company. They are also cheap, and buyers trust a brand that publishes them.
Cause of death: building the factory before the yield exists
Northvolt is the largest European example and worth going through slowly. It raised roughly $13 billion in equity and debt, hired at 100 to 150 people a month, and set out to make Skellefteå produce 16 GWh a year. In Q3 2023 it delivered 79.8 MWh, under half a percent of the annual target. Losses went from €113 million in 2022 to about €950 million in the first nine months of 2023. In June 2024 BMW cancelled a €2 billion contract over the delays. September 2024 brought 1,500 redundancies, November 2024 a Chapter 11 filing, and on 12 March 2025 bankruptcy in Sweden.
Arrival ran the same play in vehicles. Peak valuation around £9 billion after its Nasdaq listing, more than $1.5 billion spent on R&D, a “microfactory” strategy that was going to make the whole thing capital-light. When EY was appointed in February 2024, production had not started. Operating losses at the main UK entity were £562.3 million in 2022 alone.
Neither company was short of demand signals. What both lacked was a process that made good units repeatably, and both built capacity on the assumption that the process would be solved by the time the building was finished. Manufacturing does not work that way. You find out what your yield is by running the line, which is why it cannot sit in a Gantt chart as a date.
Gate every capacity commitment on a measured number from the previous stage. Do not sign the second tool until the first tool has produced a run at your target first-pass yield, and wait for a pilot run that has held the yield across three builds with different operators before you take a factory slot. Every step of a ramp should be paid for with data from your own line rather than a customer’s forecast.
Cause of death: putting compliance at the end of the Gantt chart
Certification is the one dependency you cannot compress with money or effort, because most of it is queue time in somebody else’s lab. Since 1 August 2025, the cybersecurity articles of the EU Radio Equipment Directive have been mandatory for radio equipment, with EN 18031-1/-2/-3 as the harmonised route, and where the standards’ restrictions bite, self-declaration is off the table and a notified body has to be involved. The Cyber Resilience Act layers on top: reporting obligations from 11 September 2026, full application 11 December 2027, and a machine-readable software bill of materials with every release.
The extreme version of this cause of death is European aviation. Lilium and Volocopter between them consumed billions and years pursuing type certification. Lilium filed for insolvency in October 2024, briefly recovered, then filed again on 21 February 2025 when a €200 million rescue collapsed, holding an order pipeline it put at more than 780 aircraft and having never carried a paying passenger. Volocopter entered insolvency in December 2024. Nobody in either company misunderstood that certification was required. What they got wrong was how much runway it eats when it sits at the end.
Book the lab before you need it, and treat pre-compliance as a design activity rather than a test activity. Run radiated emissions on the first engineering build in a cheap chamber, not the last. Ask your notified body what their lead time actually is this quarter and put that number, not an optimistic one, in the plan. Budget one failed test campaign and one board respin as the expected case; if you do not need them, you have found six weeks. Keeping the SBOM current from the first firmware commit costs nothing and is miserable to reconstruct later.
Cause of death: the gap between the purchase order and the payment
Cana raised $30 million in January 2022 for a molecular beverage printer, built a working prototype, signed brand partnerships and hired Patrick Stewart as an ambassador. In May 2023 it shut down and laid off everyone, because it could not raise the money to build a production line. It had a product. It did not have the capital to convert the product into units.
This is the least glamorous killer and the most common. The arithmetic does not negotiate: your contract manufacturer wants a deposit against tooling and components, then the balance before the goods leave the building. Long-lead parts want committing months earlier still. Your retail or distribution customer pays 60 to 90 days after delivery. Between those two facts sits a hole the size of your first production run, and you either finance it or you fall into it. Embodied’s shutdown had the same structural cause from a different direction: a single lead investor withdrew, and there was no reserve behind them.
Build a weekly cash model rather than a monthly P&L, and extend it through the second production run. The first is usually funded by the raise; the second is where teams discover they have sold everything and cannot afford to build more. Negotiate payment milestones with your CM as a real commercial term, not an afterthought; a 30/40/30 split against milestones is worth arguing for and often available. Get an inventory financing or receivables facility approved before you need it, because approval takes weeks you will not have. And never rely on one investor to bridge a build. Two smaller commitments that exist beat one large one that has only been discussed.
Cause of death: assuming your bill of materials and your border stay put
Two things moved under the industry’s feet in 2025 and 2026, and both punish designs that were locked without alternatives.
The first is memory. Manufacturers reallocated capacity toward high-bandwidth memory for AI infrastructure, and the consumer-grade supply collapsed. DRAM rose about 172% across 2025; some NAND wafer categories moved more than 60% month-on-month in November 2025. PC OEMs reported 15 to 20% cost increases, with Dell’s COO saying he had never seen costs escalate at this pace. Smaller buyers get the worst of it: prepayment demands, short quote validity, allocation. SK Hynix has suggested the imbalance may persist past 2030. If your product has a DDR part or an eMMC and no qualified alternate, your BOM cost is now set by someone else’s negotiation.
The second is trade. Duty-free treatment of small parcels is gone on both sides of the Atlantic: the US suspended the $800 de minimis exemption in August 2025 and made that suspension indefinite for every mode except international post in June 2026, and the EU replaced its €150 threshold with a flat per-item charge from July 2026. That removed the direct-to-consumer workaround a generation of hardware brands had built into their model. The 25% Section 232 semiconductor duty that took effect on 15 January 2026, by contrast, landed narrowly, on accelerator-class parts defined by performance thresholds rather than the microcontrollers and power devices in an ordinary BOM. For most consumer hardware the exposure sits in the classification rather than the headline rate, and an HTS code that is wrong now costs real money.
Qualify a second source for every part that is single-sourced and non-trivial to replace, and do it at schematic review while changing a footprint is free. Ask suppliers for quote validity in writing and assume it expires. Buy ahead on the parts with the longest lead times even when it feels premature. This is one of the few places where tying up cash early is defensible. Model landed cost with tariff scenarios rather than a single number, and get your HTS classifications reviewed by someone who does it for a living. At RMBG we run this as a standing part of design review rather than a procurement task, because by the time it is a procurement task the schematic is already wrong.
What kills you
Almost none of these companies died of a technical problem their engineers could not solve. They died of commitments made before the evidence arrived: capacity before yield, launch before the core loop worked, a price before the cost was known, a cloud architecture before anyone asked what happens when the company stops paying for it.
Hardware is unforgiving in a specific way. Every commitment hardens on its own clock, and they harden faster than your understanding improves. Caution alone does not fix that, and plenty of cautious hardware companies fail too. What helps is sequencing: knowing which decision each piece of evidence is allowed to unlock, and refusing to spend the next one until you have it.
The companies above were not unlucky. Most of them were early, and being early is only a mistake if you spend like you are on time.
Sources
- Humane’s AI Pin is dead as HP buys the startup’s assets for $116M — TechCrunch
- Humane AI Pin startup shuts down, sells to HP for $116M — TechStartups
- Humane AI Pin woes worsen as recent returns exceed sales — 9to5Mac
- Rabbit r1 — Wikipedia
- Moxie’s $799 robot companion for children is going to die — Gizmodo
- Embodied Moxie shutdown — OECD AI Incidents Monitor
- Spotify begins offering Car Thing refunds as it faces lawsuit over bricking the device — TechCrunch
- Belkin ends support for most Wemo devices and its Wemo app — TechCrunch
- FTC finds smart product makers silent on software support periods — The Register
- Bird Global — Wikipedia
- Northvolt bankruptcy, latest — Sifted
- Northvolt files for bankruptcy in Sweden — Northvolt
- Arrival: the full story of how a £9bn EV firm collapsed into administration — City A.M.
- British EV maker valued at $13 billion enters administration without making sales — Fortune
- RED cybersecurity requirements mandatory on 1 August 2025 — SGS
- The Cyber Resilience Act explained
- Lilium files for insolvency again as funding deal falls through — Vertical Magazine
- eVTOL maker Volocopter files for insolvency — AeroTime
- Cana, the startup building a make-any-drink beverage printer, shuts down — The Spoon
- 2025–present global memory supply shortage — Wikipedia
- Guidance on Section 232 semiconductor import duties for 2026 — C.H. Robinson
- Narrowly Targeted 25% Section 232 Tariff on Certain Advanced Semiconductors — White & Case
- The end of de minimis for China — Easyship
- Indefinite Suspension of the De Minimis Exemption — Federal Register, 24 June 2026
- EU ends €150 duty exemption, charging €3 per item — PPC Land