DeathStack
Every AI tool you build your business around is a bet that it'll still exist next year. Most of the time that bet pays off. Sometimes it doesn't, and a solopreneur wakes up to find the tool they depend on has quietly gone dark, no warning, no refund, no forwarding address.
These are twenty-five that didn't make it. Real companies, real dates, real reasons, sourced and linked. Read them like a horror movie: you already know how it ends, the interesting part is watching how it got there.
The graveyard
ChatGPT Plugins
The landlord can always change the locks. It's still your fault for never asking who owned the building.
It started as the biggest opportunity in AI: OpenAI opened ChatGPT up to third-party plugins, letting it book flights on Expedia, browse the live web, run code, and reach into dozens of outside services. Developers and solopreneurs rushed in, building on ChatGPT the way an earlier generation built on the App Store or Facebook Platform. This was supposed to be the land grab.
Then, one year later, without much warning to the builders who'd bet on it, OpenAI unveiled GPTs and the GPT Store and began quietly winding plugins down. The distribution channel these developers had spent months building on simply switched off. No transition plan, no revenue-sharing consolation prize. Just a support article, filed under Help, titled almost cheerfully: "Winding down the ChatGPT plugins beta."
The PDF Chat Wrappers
A hundred blacksmiths were selling a spell the king was about to cast for free. Only the king knew the announcement date.
For about eight months in 2023, "upload a PDF and chat with it" was one of the hottest weekend-project-turned-startup ideas in AI. A wave of nearly identical tools, ChatPDF, PDF.ai, AskYourPDF, and dozens of clones, launched doing the same trick: wrap an OpenAI model in a file uploader, charge roughly $10 a month, watch the signups roll in.
Then OpenAI shipped native file upload directly inside ChatGPT. Free. No install. No signup. The entire category didn't get a slow, dignified decline, it got made irrelevant in a single product announcement. Most of these apps are technically still online. Almost nobody is paying for them anymore.
Neeva
You cannot out-price a giant who has already decided to lose money forever just to keep you out.
Neeva had everything a startup is supposed to need: founders who ran Google's ad business and built YouTube's monetization engine, $77.5 million raised, and a genuinely good pitch, an ad-free, private, AI-powered search engine that didn't sell your data. Tech press loved it. Early users loved it.
It didn't matter. Google Search is free, Neeva wasn't, and no amount of founder pedigree changes the math on convincing millions of people to pay monthly for something they can already get for nothing. Snowflake eventually acquired Neeva, not for the search engine, for the AI talent. The product itself was shut down within weeks.
Olive AI
The wizard's automaton worked flawlessly, right up until you noticed the seven hundred tiny hands operating it from inside.
At its peak, Olive AI was valued at $4 billion and had raised $902 million promising to automate the mind-numbing administrative work that eats hospital budgets alive: prior authorizations, claims processing, scheduling. It was one of the most hyped healthcare AI companies in the country.
Then the automation didn't automate. Investigations and internal reporting found large parts of what Olive sold as AI relied heavily on manual human labor happening behind the interface. Hospitals that had built workflows around Olive's promises were left scrambling. After mass layoffs, Olive shut down entirely and sold off its remaining assets for a fraction of that $4 billion number.
Builder.ai
Give the golem a name and a face and investors will believe it did the work the villagers did overnight.
Builder.ai pitched itself as an AI so advanced it could build you a working app just by describing what you wanted, no developers required. Microsoft backed it. It hit a $1.5 billion valuation. The company even nicknamed its AI system "Natasha," giving it a face and a personality investors could root for.
Natasha was, allegedly, roughly 700 engineers in India writing code by hand. Investigations found the company had inflated its reported revenue by nearly 300%, claiming around $220 million against a real figure closer to $55 million. When a creditor seized $37 million from its accounts, the whole structure collapsed, triggering bankruptcy proceedings across five countries at once.
Tune AI
The toll bridge business collapses the day the king builds his own bridge and charges nothing to cross it.
Backed by Accel and Flipkart, Tune AI built a genuinely useful product: tools for fine-tuning and running large language models, sold through Tune Chat and Tune Studio. For a while, it filled a real gap for developers who needed LLM infrastructure without building it themselves.
The gap closed. AWS, Google, and Azure all shipped equivalent tooling, bundled into platforms most developers were already paying for, at a lower marginal cost than a standalone vendor could match. Most of Tune's users were on the free tier and never converted to paying customers. The infrastructure costs stayed high. The margins didn't.
Woebot
The healers' guild takes seven years to approve a remedy. The plague, unfortunately, does not wait for paperwork.
Woebot was one of the first genuinely well-regarded AI therapy chatbots, delivering cognitive behavioral therapy exercises to roughly 1.5 million people and earning an FDA Breakthrough Device designation, a real signal of legitimacy in a category full of hype.
That designation became the problem. The regulatory approval process for AI mental health tools moved on FDA time while the underlying AI moved on large language model time, and the gap between the two eventually became too expensive to bridge. Woebot shut down not because it failed users, but because the rules built for a slower era of software never caught up to it.
Coqui AI
The tools loved by every apprentice in the guild are not always the tools any lord will pay for.
Coqui built the Coqui TTS library, an open-source text-to-speech tool that became the go-to option for developers who wanted voice synthesis without licensing a black-box commercial API. It was genuinely loved in the parts of the AI community that build things rather than just talk about them.
Being loved by developers and being funded by investors turned out to be two different problems. Coqui ran out of money in an increasingly crowded voice AI market, competing against well-capitalized commercial players. Before going dark, the team released its final models to the open-source community rather than letting the work simply vanish.
Yara AI
Most quests end when the hero runs out of gold. This one ended because the hero decided the dragon wasn't worth slaying.
Yara was a UK-based AI mental health companion offering CBT-style exercises and conversational support, the kind of product that usually keeps shipping features and raising rounds for as long as anyone will fund it.
Instead, its founder made a call almost nobody in this graveyard made voluntarily: after looking honestly at what the product was and wasn't capable of, they decided AI chatbots simply aren't safe enough for people in genuine mental health crises, and shut the company down themselves rather than keep operating.
Ghost Autonomy
Standing next to the king at the feast does not mean the king remembers your name by winter.
Founded as Ghost Locomotion by John Hayes and Volkmar Uhlig, the company set out to let ordinary passenger cars drive themselves on highways, raising $63.7 million from investors including Founders Fund and Khosla Ventures. When that didn't pan out on schedule, Ghost pivoted to crash-prevention software and raised another $100 million in 2021 to chase it.
Along the way, Ghost joined the OpenAI Startup Fund's portfolio, gaining early access to OpenAI's models and Microsoft Azure credits, plus a direct $5 million investment. Five months after taking that money, with roughly $220 million raised across its lifetime, Ghost shut down its offices in Mountain View, Dallas, and Sydney and laid off about 100 people.
Forward Health
The self-operating door worked in every demonstration. In the actual castle, it kept locking the guests inside.
Founded by ex-Google exec Adrian Aoun, Forward reached unicorn status in 2021 and rolled out CarePods, self-serve kiosks that used AI to screen and diagnose health conditions, into malls, gyms, and offices across the Bay Area, New York, Chicago, and Philadelphia. The company raised close to $700 million chasing a vision of primary care without a waiting room.
The pods themselves were the problem. Self-service blood draws routinely failed, patients occasionally got stuck inside, and after years of burning through funding on hardware and engineering, Forward had generated less than $100 million in revenue since founding. In November 2024, it shut its locations, killed its mobile app, and laid off around 200 employees, almost overnight.
Humane AI Pin
Everyone lined up to buy the amulet before anyone had worn it in daylight.
Built by former Apple designers, the AI Pin promised to replace the smartphone with a screen-free wearable: a laser projector on your palm, a voice assistant handling the rest. It launched with over $230 million raised and some of the loudest hype in consumer hardware that year.
Then it actually shipped. Reviewers found it slow, prone to overheating, and hard to justify at $700 plus a $24 monthly subscription. Within months, daily returns were outpacing daily sales. HP acquired what was left for $116 million, not for the device itself but for the underlying CosmOS platform, the patents, and the founding team. Every AI Pin still in someone's pocket stopped connecting to Humane's servers at a scheduled deadline, noon Pacific on February 28, 2025, and the cloud data behind it was deleted for good shortly after.
Tome
The whole village came to see the free magic show. Nobody bought a ticket to the second one.
Tome became the fastest productivity tool ever to reach a million users, hitting that mark 134 days after launch, faster than Dropbox, Slack, or Zoom. It kept growing past 20 million users and raised $81 million from investors including Lightspeed, Coatue, Greylock, and former Google CEO Eric Schmidt, at one point valued around $300 million.
Almost none of those millions of users were paying anything close to what the company needed. Most sat on the free plan or a $10-a-month tier, nowhere near enough to fund the business behind it. Founder Keith Peiris shut Tome down in March 2025, laid off nearly all of a 70-person team down to six, and redirected the survivors toward a new product entirely, an AI sales CRM called Lightfield, a category where buyers actually expect to pay real money.
Sora
The spell cost a fortune to cast and the audience tossed a single copper coin into the hat.
OpenAI's standalone Sora app let anyone generate short AI videos, including ones dropping your own face into them through an invite-based "cameo" feature. It launched to enormous buzz, peaked at roughly a million active users within its first weeks, and landed Disney as a partner willing to commit $1 billion in equity to be part of it.
Every video generated was burning through OpenAI's own compute at a reported rate of about $1 million a day, against revenue from the app's entire lifetime that added up to roughly $2 million. Usage kept sliding, down to under 500,000 active users by the time OpenAI pulled the plug in March 2026, redirecting the compute toward coding tools with more predictable revenue. Disney reportedly learned its billion-dollar partnership was ending less than an hour before the public announcement did.
ChatGPT Atlas
The kingdom built a second castle just to prove the first castle's gate should have been wider, then moved back into the first castle.
Atlas was OpenAI's own standalone web browser, ChatGPT built directly into the sidebar with an agent mode that could research, click through, and complete multi-step tasks on live webpages. It arrived with the kind of launch fanfare reserved for a real bet on the future of browsing.
Less than nine months later, OpenAI shut the whole thing down, giving users until August 9, 2026, to export their data before it went dark. The company's own explanation doubled as an admission: agentic browsing works better living inside a browser or app people already use, not as a separate destination competing for a permanent spot in the dock, the same conclusion Google, Microsoft, and Anthropic had each already reached. Whatever OpenAI learned building Atlas is now being folded back into ChatGPT's desktop app and a Chrome extension instead.
Yupp
By the time the tailor perfected the measuring tape, the whole village had stopped wearing clothes that needed measuring.
Yupp let anyone test and compare answers from roughly 800 AI models side by side for free, then collected anonymized feedback on which model actually won for a given prompt, data it planned to sell back to the labs building those models. It signed up 1.3 million users, landed a leaderboard people actually checked, and raised a $33 million seed round led by a16z crypto's Chris Dixon, backed by angels including Google DeepMind's chief scientist and Perplexity's CEO.
Less than a year after launching, the founders shut it down. Their own explanation was blunt: the AI models it was built to evaluate improved so fast that the comparison problem it solved kept shrinking, while the real money in model feedback was already going to firms hiring specialty PhD raters to work inside labs' reinforcement-learning pipelines directly, not to a consumer-facing leaderboard.
theGist
A sturdy cart built by the finest wheelwrights in the land still needs somewhere the village actually wants to go.
theGist scanned a company's Slack, email, CRM, and analytics tools to surface personalized summaries of what actually mattered that day, a real attempt to solve information overload rather than just add another dashboard to check. It had offices in Tel Aviv and New York, around 15 employees, and a founding team with real pedigree: alumni of Wix, Gong, Blizzard, and OpenWeb, including Wix co-founder and president Nir Zohar as a backer.
It raised $7 million in pre-seed funding in November 2022 and shipped its product in 2023. What it never found was sustainable traction, a pivot toward sales-focused use cases didn't gain momentum, and prolonged reserve-duty call-ups for key engineers slowed development further. Rather than force a drawn-out pivot on a shrinking runway, the founders closed the company and returned what was left of investors' money.
Reforged Labs
The armorer's craft was flawless. The tournament that needed it simply stopped being held.
Reforged Labs built generative AI tools for mobile game studios, working with genuinely major clients including Supercell, King, Ubisoft, Playtika, and Scopely. With a lean team of around 17 people, the company carved out a real niche helping game studios generate art and content faster.
After three years, the founders closed the company, citing the closure of the specific market gap they had built the business around: as generative tools matured and consolidated, the specialized wedge Reforged Labs occupied between studios and raw AI models simply stopped being necessary. Not a scandal, not a fraud, just a market that quietly stopped needing the middleman.
Relay.app
One morning the workshop just had a sign on the door. No plague, no fire, no debtors at the gate. Just a sign.
Relay.app was a well-regarded AI workflow automation platform built by Jacob Bank, who had previously sold his startup Timeful to Google and gone on to lead product for Gmail and Google Calendar. Relay combined AI actions like summarization and data extraction with human-in-the-loop approvals, connecting over 100 apps including Gmail, Slack, and Notion. It raised $8.1 million across a Khosla Ventures-led seed and an Andreessen Horowitz-led follow-on round.
On July 16, 2026, Relay announced it was shutting down, free accounts losing access August 15 and paid accounts September 14. No cause was given, no acquirer named, no acquihire announced, just a shutdown notice on a product users had genuinely liked.
Astra
Two knights swore the same oath at the same fire. By the second village, they had stopped speaking, and the quest died with the friendship.
Astra pitched itself as a "Chief of Staff for every account executive," an AI copilot built to help sales teams manage their pipeline and outreach. In March 2025, the company announced pre-seed funding from a genuinely notable backer: Aravind Srinivas, CEO of Perplexity AI. It was the kind of validation that usually buys a startup a year or two of runway and credibility.
Instead, roughly four months later, co-founder and CEO Supreet Hegde announced Astra was shutting down, citing an internal rift with his co-founder over the pace of growth. The product itself had also struggled to move past limited pilots, with enterprise buyers uncertain how to evaluate AI agents and long sales cycles working against a young company. But the stated final cause was simpler and more human: the founders chose to go their separate ways, and the company didn't survive the split.
Doppl
The scout who found the shortest road to the capital was thanked, reassigned, and never allowed to walk it herself again.
Doppl was Google Labs' standalone experimental app for AI virtual try-on: upload a photo, see clothing rendered onto your own body before you buy. It shipped to real acclaim and genuine usefulness for online shoppers tired of guessing whether something would fit.
Less than a year later, on April 30, 2026, Google shut Doppl down, not because it failed, but because it worked. The underlying try-on technology was folded directly into Google Search, Google Shopping, and Google Images, letting anyone try on clothes from a single selfie without ever opening a separate app. The standalone experiment had done its job: prove the technology, then get absorbed into the platform it was built to eventually serve.
Cruise
The self-driving carriage promised it saw everything on the road. It just didn't see the person stuck beneath it.
Cruise was General Motors' autonomous robotaxi division, spun out of a startup GM acquired in 2016 for over a billion dollars. By 2023 it was running real, paid driverless rides in San Francisco and other cities with nobody behind the wheel, a genuine engineering milestone very few companies in the world had reached.
On October 2, 2023, a Cruise vehicle struck a pedestrian who had just been hit by a separate human-driven car, then dragged her roughly 20 feet at low speed while attempting to pull over, seriously injuring her. California regulators suspended Cruise's driverless permit within weeks, GM pulled every Cruise vehicle off public roads nationwide, and a year of layoffs, leadership departures, and eroded public trust followed. In December 2024, GM announced it would stop funding Cruise's robotaxi ambitions entirely, folding the remaining team into its own in-house driver-assistance work and ending the standalone robotaxi business for good.
IBM Watson Health
The kingdom spent four billion gold summoning an oracle. It sold the oracle for one billion once the prophecies stopped making sense.
IBM launched Watson Health in 2015, riding the fame of Watson's televised Jeopardy! win, with a stated mission to transform cancer treatment, drug discovery, and medical research using AI. IBM spent more than $4 billion acquiring health data companies to feed it, and the division was pitched publicly as a genuine moonshot for the entire healthcare industry.
The clinical results never matched the pitch. Hospitals using Watson for Oncology reported the system frequently recommended treatments doctors considered unsafe or already obvious, and the promised drug-discovery breakthroughs mostly never materialized at scale. On January 21, 2022, IBM announced it was selling most of Watson Health's data, imaging, and analytics assets to investment firm Francisco Partners for just over $1 billion, a fraction of what IBM had spent building it. The business was relaunched under a new name, Merative, with no more mention of curing anything.
Amazon Halo
Nobody accused this amulet of malfunctioning. The kingdom just decided it no longer wanted to be in the business of amulets.
Amazon Halo launched in 2020 as the company's first wearable health device: a wristband paired with an app that used AI to estimate body composition from a phone photo, analyze tone of voice to flag stress in your speech, and track sleep and activity, a genuinely ambitious bet on Amazon becoming a serious player in personal health.
Less than three years later, on April 26, 2023, Amazon announced it was winding down the entire Halo line with little explanation beyond a strategic shift, effective July 31, 2023. Every device stopped working the next day. Customers who'd bought a device or subscription in the past year were refunded automatically, and Amazon deleted all stored Halo health data. It was one of a wave of quiet product cancellations during a broader company-wide cost-cutting push that year.
Meta Galactica
This scroll read like scholarship and lied like a con artist. The guild pulled it from the shelf before the ink had even dried.
Meta unveiled Galactica on November 15, 2022, a large language model trained on 48 million scientific papers, textbooks, and reference material, pitched as a tool that could write scientific articles, summarize research, and generate Wikipedia-style entries on demand. Meta's chief scientist Yann LeCun championed it publicly as a serious step forward for AI-assisted science.
Within hours of the public demo going live, researchers and journalists found it fabricating convincing but entirely false scientific claims, complete with citations to papers that didn't exist, and in some cases generating biased or offensive content on serious topics. The backlash was immediate and loud enough that Meta pulled the public demo on November 18, 2022, just three days after launch, an unusually fast retreat for a major AI lab. It also happened two weeks before ChatGPT's launch would reset the entire industry's tolerance for AI hallucination, timing that made Galactica's downfall look almost quaint in hindsight.
What this means for your stack
None of these tools died because their founders were foolish. Some had elite pedigrees, huge funding, real users, and government-backed legitimacy, and they still didn't make it. If there's one thread through all twenty-five, it's that betting your workflow on a single AI vendor, especially a platform feature or a thin wrapper around someone else's model, is a real risk, not a hypothetical one. Every review on this site now carries a "last verified" date for exactly this reason. The tools change. Sometimes they disappear.
Know of one that belongs here? Submit it, along with what happened and a source, and we'll add it to the stack.