Two billion dollars in consumer AI hardware burned in two years. The graveyard holds two distinct species of failure — one that kills instantly and one that kills slowly — and the counter-example that explains why the category is being reborn.
On February 28, 2025, at noon Pacific Time, every Humane AI Pin in existence stopped working. The company had raised $230 million from investors including Tiger Global, Microsoft, and Qualcomm Ventures. It shipped roughly ten thousand units. When HP acquired what remained for $116 million — half the capital raised — the servers were the first thing to go. The devices did not degrade. They bricked.
Six months earlier, Humane had sought a buyer at $750 million to a billion dollars. The final price was a confession that the business was worth less than the sum of its patents and personnel. HP folded the engineering team into an enterprise lab called HP IQ and pointed them at making existing hardware smarter. The consumer AI hardware dream that produced the Pin died in a press release.
Two deaths
Humane's death was instant. The server dependency that powered every function meant that when the company could no longer sustain the infrastructure, neither could the product. Every device became inert plastic on the same day. Owners had no warning and no recourse. Only customers who had purchased within the prior ninety days received refunds.
Rabbit's death was slower. The R1 launched off $30 million in funding from Khosla Ventures and shipped roughly 100,000 units — mass returns followed, and by September 2024 only about 5,000 were in daily use. By mid-2025, employees were going unpaid. By October, three of the company's twenty-six remaining staff went on strike. Rabbit released a software update and teased a next-generation device, but no new funding materialized. The hardware still technically functions. The company around it is dissolving.
The two failures are structurally different. Humane's Pin required constant server connectivity for its core functions — the device was a terminal, not a computer. When the servers went dark, the terminal went with them. Rabbit's R1 relied on a cloud-based Large Action Model for its central purpose of operating apps on the user's behalf, but could perform some operations locally. The server dependency was real but partial. The result was not instant death but a slow drain: each missed paycheck, each departed engineer, each month without funding dimmed the product further without a single definitive moment of failure.
The distinction matters for anyone evaluating AI hardware investments. A bricking event is binary — it either happens or it does not, and when it happens, the capital destruction is total. A zombie decline is probabilistic — each quarter the company survives, the product survives with it, but the probability of eventual abandonment compounds. Both are server-dependency risks. They differ in speed.
What worked
EssilorLuxottica sold seven million pairs of Meta Ray-Ban AI glasses in 2025, more than tripling sales from the prior two years combined. Production is scaling toward twenty million units or more by the end of 2026 — the international rollout was paused due to what the company called unprecedented domestic demand. Prescription lens models launched in March 2026, expanding the addressable market to the hundreds of millions who already wear corrective eyewear.
The structural differences explain the outcome. The Ray-Bans are sunglasses first. If Meta's servers go dark, the owner still has functional eyewear. The AI features — camera, assistant, translation — are additive. They do not replace anything the user already does. Distribution runs through EssilorLuxottica's global retail network, not a direct-to-consumer website nobody visits twice.
Humane tried to replace the smartphone. Rabbit tried to replace individual apps. Meta added AI to sunglasses. The graveyard holds the replacements. The market rewards the additions.
The resurrection
On May 21, Brett Adcock raised $700 million in a Series A at a $6 billion valuation for Hark, his AI hardware company. Adcock, who previously founded Figure AI and Archer Aviation, had seeded the company with $100 million of his own capital. The round was led by Parkway Venture Capital.
The investor list reads like a semiconductor industry consortium: Nvidia, AMD Ventures, Intel Capital, Qualcomm Ventures, ARK Invest, Brookfield, and Salesforce Ventures. Humane's backers were software investors and celebrity technologists. Hark's backers are the companies that manufacture the chips that go inside the devices. They have direct visibility into which designs are viable. When every major chip company invests in a consumer AI hardware startup, they are investing downstream of their own component roadmaps.
Era Computer raised $11 million in April to build an operating system for AI glasses, rings, and pendants — a platform layer, not a device. Apple, Samsung, and Google are each developing AI glasses. The category that failed as standalone replacement devices is being rebuilt as additive extensions of existing form factors, financed by the companies that supply the silicon.
The capital lesson
The conservative estimate of capital destruction in first-generation consumer AI hardware exceeds two billion dollars. Humane and Rabbit are the visible cases. Builder.ai burned through $445 million before filing for bankruptcy in May 2025 — investigations revealed its AI was largely offshore human workers. Roughly 3,800 AI startups shut down in 2025, about a quarter of the fourteen thousand launched the prior year.
Most were not hardware companies. But the hardware failures carry a lesson that software failures do not: when the product physically exists in a user's hand and then stops working, the trust damage extends beyond the company to the category.
Hark's investors are betting that the lesson has been absorbed. No server-dependent standalone devices. No phone replacements. Additive features on existing form factors, backed by the chip companies whose economics depend on the hardware succeeding. The graveyard did not kill the category. It taught the next generation what not to build.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)