The lifespans of dead AI products

Microsoft Tay went live on Twitter on March 23, 2016, at roughly 8:14 a.m. Pacific. By the next morning, the account was suspended. The chatbot had been coaxed by users into posting racist, sexist, and Holocaust-denying messages, and Microsoft pulled it after about sixteen hours. Tay is the shortest-lived entry in our archive. It is also, in a sense, the most honest one, because its failure was visible in real time and the cause of death was not in dispute. Most AI products take longer to die, and most deaths are harder to explain.

The thirty products documented in this issue of The Graveyard span a range of lifespans that is wider than the category itself feels. Tay lasted sixteen hours. Meta's Galactica, a science-focused language model demo released on November 15, 2022, was withdrawn three days later after researchers found it confidently fabricating citations and producing plausible-looking nonsense about real scientific topics. At the other end, Vicarious, the brain-inspired robotics company founded in 2010 by Scott Phoenix and Dileep George, was acquired by Alphabet's Intrinsic in April 2022, after twelve years and roughly $250 million in funding. Humane's AI Pin, announced in 2017 and shut down in February 2025, ran for seven years and burned through $230 million.

Looked at as a whole, the archive sorts roughly into three lifespan bands, and the band a product lands in says something useful about how it died.

The first band is the fast deaths, measured in hours or days. These are almost always public demos from large companies, not startups. Tay and Galactica are the canonical examples. Neither had a business model to defend, neither had paying customers, and neither had a board to consult before pulling the plug. They died fast because they could. The cost of leaving them up, measured in reputation, exceeded the cost of taking them down, measured in nothing. Speed of death in this band is a function of optionality, not of severity.

The second band runs from about three to six years. This is where most of the venture-funded entries cluster. Drive.ai, the Stanford self-driving spinout founded by Sameep Tandon and Carol Reiley in April 2015, was acqui-hired by Apple in June 2019 after raising $77 million. Voyage Auto, which tried autonomous shuttles in retirement communities, folded into Cruise in March 2021 after about four years and $52 million. Starsky Robotics, which removed the driver from highway trucks before anyone else, shut down in March 2020 after roughly four years and $21.7 million. Forefront Chat, a free consumer chat wrapper, lasted barely over a year before closing in February 2024.

The pattern in the middle band is that money runs out before the technology arrives. Founders raise on a thesis, the thesis requires a breakthrough or a regulatory shift, the breakthrough or shift does not come on schedule, and the next round does not close. Stefan Seltz-Axmacher's post-mortem on Starsky said this plainly: the supervised machine learning approach was harder than they had budgeted for, and the safety case was not where investors wanted it to be by 2020.

The third band is the slow deaths, six years and longer. These are the most interesting entries, because the companies in this band almost always had real revenue, real customers, and real products that worked, at least partially. Argo AI, founded in November 2016 by Bryan Salesky and Peter Rander, lasted just under six years and absorbed $3.6 billion from Ford and Volkswagen before being wound down in October 2022. Olive AI, the healthcare RPA company founded in 2012 by Sean Lane, lasted eleven and a half years and raised $902 million before shutting in October 2023. Babylon Health ran for ten and a half years under Ali Parsa, raised $1.15 billion, briefly went public via SPAC, and ended in August 2023.

Slow deaths tend to share a structure. The product works well enough to attract customers and capital, but not well enough to defend a valuation set during a period of cheaper money. Olive AI was reportedly valued at $4 billion in 2021. Babylon's SPAC valued it at $4.2 billion in 2021. Neither could grow into those numbers. When the funding market reset in 2022 and 2023, the gap between what the businesses produced and what they had promised became unbridgeable, and the companies were unwound rather than rescued.

Hardware AI dies slowly, and expensively. Jibo, the social home robot from MIT's Cynthia Breazeal, was announced in 2014, shipped in late 2017, and shut down its servers in March 2019. Anki, the consumer robotics company behind Cozmo and Vector, lasted nine years and raised about $200 million before closing in May 2019. Humane's AI Pin lasted seven years from founding to wind-down, with the actual product on the market for about ten months. Lighthouse AI, a smart home security camera with computer vision, ran from 2015 to December 2018 on $17 million.

The hardware pattern is not really about lifespan, it is about the shape of the burn. A robotics or wearable company spends years and tens of millions before it has anything to ship, ships a v1 that is necessarily compromised, and then needs a v2 it cannot afford. Jibo had servers to keep running for customers who paid $899. Humane had a returns problem and a battery that overheated. The clock on these companies starts the day they ship, not the day they incorporate, and it runs faster than the cash.

Autonomous vehicles deserve their own paragraph because they dominate the archive by capital. Argo AI, TuSimple's US operations ($1.648 billion, wound down January 2024), Embark Trucks ($614 million, March 2023), Ghost Autonomy ($220 million, April 2024), Optimus Ride, Voyage, Drive.ai, and Starsky together raised more than $6 billion and produced no surviving independent company. The lifespans varied from four years (Starsky) to nearly eight (TuSimple), but the cause of death was consistent: the gap between a demo on a closed route and a deployable system on public roads turned out to be longer than any single funding cycle could cover.

Healthcare AI dies on a different clock, governed by regulation and reimbursement rather than engineering. Forward Health, the AI primary care kiosk company founded by Adrian Aoun in 2016, closed in November 2024 after raising $657 million across roughly nine years. Pear Therapeutics, which had FDA-cleared prescription digital therapeutics, lasted ten years and raised $396 million before filing for bankruptcy in April 2023. Mindstrong Health, working on digital biomarkers for mental illness, ran nine years on $160 million before shutting in March 2023. IBM Watson Health, which never disclosed a standalone funding number, was sold to Francisco Partners in January 2022 after about seven years.

These companies did not die because the AI did not work. They died because the path from a working model to a billable clinical service is long, and insurers, hospitals, and the FDA all run on timelines that venture funds are not built to wait through. Pear had a product covered by some plans. Forward had thousands of paying members. Neither could reach the scale their funding required before the patience ran out.

Then there are the products that died for reasons unrelated to the AI itself. Neeva, the subscription search engine from former Google ad chief Sridhar Ramaswamy, shut its consumer product in June 2023 after about four years and $77.5 million, and was acquired by Snowflake. Neeva's search worked. It just could not get enough people to pay for it while Google was free. Tome, the AI presentation tool, raised $81 million and reached a reported $300 million valuation before pivoting to enterprise and then shutting its consumer product in April 2025. Rewind, Dan Siroker's Mac app that recorded everything you did, ran from 2020 to December 2025 on $22 million before the company shifted focus to a new product. x.ai's Amy and Andrew, the AI email schedulers founded by Dennis Mortensen in 2014, lasted seven and a half years before shutting in October 2021, eventually overtaken by free calendaring built into the tools their users already had.

Two lessons emerge from the spread. The first is that the speed of an AI product's death is mostly a function of who is paying for it, not how good or bad the underlying model is. Big-company demos die in hours because nobody loses anything when they do. Venture-backed startups die in three to six years because that is how long a Series A and a Series B last without a working business. Companies with real revenue die slowly, over seven to twelve years, because real revenue buys time but not, in the end, escape.

The second lesson is that the long lifespans are not necessarily the successes. Vicarious lasted twelve years and produced research papers and a soft acquisition. Olive lasted eleven and produced an $0.9 billion crater. Adept AI, founded in 2022 by some of the authors of the transformer paper, lasted barely two and a half years before its founders left for Amazon in June 2024 and the remaining business was effectively wound down. A short lifespan can be a clean failure. A long one can be an expensive one. The archive does not reward duration, it just records it.

What the archive does reward is specificity. Every entry has a date, a number, a name. The reason these companies are worth documenting is not that they failed, which is the ordinary fate of ambitious technology, but that the public record of how and when they failed is the closest thing the field has to a control group. Tay's sixteen hours and Vicarious's twelve years are not opposites. They are the same data point, taken at different exposures.