The things that stopped working

On February 28, 2025, the Humane AI Pin stopped responding to voice commands. HP had purchased the company's patents and engineering talent for $116 million the week before, a figure that worked out to roughly half of the $230 million Humane had raised since its founding in 2017. The acquisition did not include the pins themselves. Owners who had paid $699 plus a $24 monthly subscription kept the magnetic chest clip, the battery booster, and the laser projector that beamed text onto their palm. None of it did anything. The servers were off.

This is the pattern. AI hardware does not fade out the way AI software does. A failed app lingers in app stores, gradually losing reviews and updates until someone at the company remembers to pull it. A failed AI device gets a shutdown date, a farewell email, and a brick. The clock starts the day it ships, not the day the company incorporates.

The Humane shutdown was not the largest of the cluster, only the most recent. Magic Leap, founded in 2011, raised $2.6 billion across roughly a decade and shipped approximately 6,000 units of its original headset against a 100,000-unit target. On December 31, 2024, the company turned off the cloud services that powered Magic Leap One and pivoted what remained of the business to enterprise customers. The headsets, which had sold for $2,295, became paperweights with eye-tracking cameras.

A pattern starts to emerge if you line up the dates. Jibo, the social robot from MIT Media Lab spinout Jibo Inc., shut down on March 4, 2019, after raising $73 million. The company had been winding down for months. When the servers went dark, the robots delivered a brief farewell speech to their owners, swiveling their heads as they had been programmed to do. Then they stopped. Owners filmed it. The clips circulated.

Embodied, the maker of Moxie, followed almost six years later. Moxie was an $800 social robot designed to teach social-emotional skills to children, many of them on the autism spectrum. The company had raised $86.3 million since its 2016 founding. On December 31, 2024, the servers went off. There were no refunds. Parents posted on Reddit about how to explain "robot death" to children who had been told the robot was their friend.

Anki, founded in 2010, was the largest consumer robotics shutdown in the cluster by unit count. The company had raised roughly $200 million and shipped about 1.5 million Cozmo and Vector robots before a last-minute funding round fell through. It closed on May 1, 2019. Vector, which depended on cloud servers for voice processing, was eventually rescued by a third party that acquired the assets and stood up a subscription service. The official servers were already gone.

Amazon Astro for Business, launched in November 2023 at $2,349, lasted ten months. Amazon discontinued the patrol-and-monitor variant of its home robot on September 25, 2024, and bricked the devices that had been sold. Business customers who had deployed Astro units for after-hours security received refunds. The hardware itself, complete with periscope camera and SLAM navigation, became unusable. The consumer Astro continued, sort of, in a limited invitation-only program.

Lighthouse AI shut down on December 18, 2018, after raising $17 million since its 2015 founding. The product was a $300 home security camera that used computer vision to recognize family members, pets, and strangers, and to answer natural-language queries like "what did the dog do today." The company's farewell note was unusually candid about why the math had not worked. Wyze had launched a $20 camera the previous year. Lighthouse could not justify a 15x price gap on the strength of recognizing the babysitter.

Google Clips, a $249 clip-on camera that used on-device machine learning to decide when to take pictures of children and pets, launched in October 2017 and was discontinued on October 15, 2019. The technology worked. The concept did not. Reviewers and customers consistently described the always-watching premise as uncomfortable. The phone in their pocket already took better pictures, on demand, when they wanted them.

Amazon Glow, a $300 projector-and-video-call device aimed at children, shipped in September 2021 and was killed on October 4, 2022. It had projected interactive games onto a tabletop while a remote grandparent participated by video call. The reviews were warm. The use case was thin. Families on FaceTime did not appear to need a dedicated projector to feel connected.

Six structural problems show up across these obituaries, and they are not the same problems that kill software products.

The first is cloud dependency. Almost every device in the cluster, including Jibo, Moxie, Astro for Business, Magic Leap One, and the Humane Pin, routed core functionality through company-controlled servers. Voice processing, vision models, content delivery, and in some cases basic startup checks all required a live connection. When the company died, the device died. There was no offline fallback because building one would have added cost and complexity to products that were already losing money on each unit shipped.

The second is the phone. The Humane Pin, Google Clips, Amazon Glow, and arguably Lighthouse all had to answer the question of what they did that a phone did not. The Pin captured photos, took notes, sent texts, and answered questions. So does an iPhone, with a better screen, better camera, better battery life, and no $24 monthly fee. Clips took candid pictures. So does the burst mode every parent already has. Phones absorbed each category before the dedicated hardware could establish a market.

The third is the privacy threshold. Consumers, when asked, will say they want smart cameras and ambient assistants in their homes. When the products ship, they buy fewer of them than the projections assumed. Clips was the cleanest example: a camera designed to watch the family all day, marketed by a company that already knew too much about the family. Astro, Moxie, and the various social robots ran into the same friction at a lower temperature. Always-on microphones and cameras pointed at children clear a bar most households do not want to clear.

The fourth is the cost curve. Humane was $699 plus $24 monthly. Magic Leap was $2,295. Moxie was $800. Lighthouse was $300. Glow was $300. None of these prices held up against the software alternative running on hardware the customer already owned. Phones get cheaper every year on a performance-adjusted basis. Dedicated AI hardware does not, because the volumes never get high enough to ride down the cost curve.

The fifth is capital intensity. Anki burned roughly $200 million to ship 1.5 million units, which sounds like a success until the unit economics are unpacked. Magic Leap burned $2.6 billion to ship about 6,000 headsets, or roughly $433,000 of investor capital per unit sold. Humane raised $230 million and shipped, by most estimates, around 60,000 pins before the shutdown, or roughly $3,800 per unit. These are not the numbers of a category that closes the loop.

The sixth is the development cycle. Hardware takes years. The Humane Pin was in development for seven years before launch. Magic Leap took roughly nine years to ship the original headset. Jibo's Indiegogo campaign ran in 2014 and the product shipped in late 2017. The AI models the world wanted in 2017 are not the models the world wanted in 2024. By the time the device arrived, the cloud-based alternative on the customer's phone had been quietly absorbing the use case for years.

What the cluster suggests, taken together, is that the AI hardware category is not failing because the technology does not work. In most cases the technology worked fine. It is failing because the economics of physical objects, the dependency on living servers, the threshold for cameras in the home, and the gravitational pull of the phone all sit on top of the same product. The archive will keep getting longer. The farewell speeches will get shorter.