AI graveyard is the name TechCrunch gave on 15 September 2026 to a running list of AI products, startups and bets that have shut down, pivoted or badly missed expectations. The list, written by Lauren Forristal, starts with Relay, an AI workflow automation startup that shut down completely on 14 September. It runs through OpenAI’s Sora app and ChatGPT Atlas browser, Apple’s delayed Siri AI, Microsoft’s Recall, Notion Mail, Humane’s AI Pin, the Rabbit R1, Huxe, Yupp and Figgs AI.
TechCrunch frames the list “less as a scorecard of flops than as a record of what these attempts can (hopefully) teach the rest of the industry”. It also cites S&P Global Market Intelligence research finding that about 42% of AI initiatives are ultimately abandoned by their corporate parents.
Below we set out every entry in one table, then look at each group in turn: OpenAI’s casualties, Big Tech delays, startups that closed, and AI hardware. We then sort the failures by cause, work out how long the products lasted and how much money was involved, and compare the list with independent trackers that count hundreds of dead AI products. We finish with practical lessons for businesses choosing AI tools. For current options, see our AI models and tools hub.
Table of contents
- What TechCrunch’s AI Graveyard List Covers
- The AI Graveyard at a Glance
- OpenAI’s Entries in the AI Graveyard
- Big Tech Delays That Earned an AI Graveyard Mention
- Startups and Products That Did Not Survive
- AI Hardware in the AI Graveyard
- Why Projects End Up in the AI Graveyard
- How Big the AI Graveyard Really Is
- Lessons for Businesses Buying AI Tools
- What the Survivors in the AI Graveyard Have in Common
- Frequently Asked Questions About the AI Graveyard
- References
What TechCrunch's AI Graveyard List Covers
The TechCrunch list is not limited to startups that went bust. It deliberately mixes three kinds of failure: companies that closed, products that big companies shut or folded into other apps, and high-profile features that missed their promises badly enough to cause backlash, lawsuits or years of delay.
Why the list starts with Relay
Relay, an AI-powered workflow automation tool built as an alternative to Zapier, is the entry that prompted the list. TechCrunch notes that as OpenAI, Google and other large platforms built similar automation into their own tools, the five-year-old startup “struggled to hold on to a reason to exist as a stand-alone product”. That pattern, a startup outpaced by bigger platforms, recurs throughout the AI graveyard.
Abandoned inside big companies too
The list also stresses that plenty of AI bets die inside the companies that start them. The S&P Global figure of 42% refers to AI initiatives abandoned by their corporate parents. TechCrunch lists the reasons as insufficient funding, technical challenges, competition, difficulty scaling or weak user demand.
The AI Graveyard at a Glance
The table below lists every product on the TechCrunch list with its status and the reason given, using TechCrunch’s reporting. Dates are 2026 unless stated.
| Product | Company | What it was | Outcome | Reason given |
|---|---|---|---|---|
| Relay | Relay | AI workflow automation | Shut for paying users 14 Sep | Big platforms built the same features |
| ChatGPT super app redesign | OpenAI | Chat, Codex and Work modes in one app | Rolled back after 9 Jul launch | Users found it confusing |
| ChatGPT Atlas | OpenAI | AI web browser | Discontinued 9 Aug | Features folded into ChatGPT |
| Sora app | OpenAI | AI video app | Shut March 2026 | High costs, falling users |
| Siri AI | Apple | Context-aware assistant | Years late; $250M settlement | Engineering problems and bugs |
| Recall | Microsoft | Searchable screenshot history | Delayed nearly a year, still criticised | Privacy and security backlash |
| Notion Mail | Notion | AI email inbox | Shutting 22 Sep | Users let agents run email instead |
| AI Pin | Humane | Wearable AI device | Shut Feb 2025; assets sold to HP | Poor performance, battery concerns |
| R1 | Rabbit | AI companion device | Still updated, pivoting | Unfinished at launch |
| Huxe | Huxe | AI podcast-style audio app | Shut May 2026 | Big platforms offered the same thing |
| Yupp | Yupp | Free model comparison playground | Shut March 2026 | No strong product-market fit |
| Figgs AI | Figgs AI | AI character role-play | Shut 2024 | Free service too expensive to run |
OpenAI's Entries in the AI Graveyard
“Even the biggest players aren’t immune,” TechCrunch writes. OpenAI, one of the most valuable private companies in the world, accounts for more entries in the AI graveyard than any other company on the list.
The super app that was rolled back
On 9 July 2026, OpenAI updated ChatGPT to combine several experiences, including separate “Chat”, “Codex” and “Work” modes, and renamed the familiar version “ChatGPT Classic”. The redesign drew criticism from users who found the new interface confusing and cluttered, and OpenAI rolled it back soon after, restoring the familiar interface.
ChatGPT Atlas, Operator and DALL-E folded in
OpenAI has also closed stand-alone apps and moved their features into ChatGPT, a strategy TechCrunch says “worked out better than the redesign did”. ChatGPT Atlas, a stand-alone AI web browser, lasted less than a year before being discontinued on 9 August 2026. Operator, an agent that browsed the web for users, went the same way, and image generation inside ChatGPT has shrunk DALL-E’s role as a separate destination.
Sora: a costly video app
The best-documented OpenAI entry is Sora, its video-sharing app, which OpenAI shut in March 2026, about six months after releasing it to the public. TechCrunch, citing a Wall Street Journal investigation, reported that Sora’s worldwide users peaked at around one million and fell below 500,000, while the app burned through roughly $1 million a day in compute.
Disney had committed $1 billion to a partnership, TechCrunch reported, and learned Sora was being shut down less than an hour before the public did. The deal died with the app. Chief executive Sam Altman chose to free up compute for areas where OpenAI faced stronger competition, including coding tools.
Big Tech Delays That Earned an AI Graveyard Mention
Two entries on the list are not dead products. They are features that missed their promises so publicly that the damage became part of the story.
Apple’s Siri AI and a $250 million settlement
When Apple unveiled Apple Intelligence in 2024, a smarter, context-aware Siri was a headline feature. Apple repeatedly pushed back its release, citing engineering issues and bugs. The delay contributed to a $250 million settlement of a class action over how Apple marketed AI features ahead of the iPhone 16. Eligible US buyers of an iPhone 15 or iPhone 16 between 10 June 2024 and 29 March 2025 can claim up to $95 per device.
Siri AI finally reached the iOS 27 public beta in July 2026 and is rolling out to English-language users this month, with other languages to follow. So this AI graveyard entry has a partial resurrection. It shows how expensive an overpromise can be even when the product eventually ships.
Microsoft Recall and the privacy backlash
Microsoft announced Recall at its 2024 Build conference as an AI-powered “photographic memory” for Windows PCs. It periodically takes screenshots so users can search their own digital history. Critics warned it could create a searchable archive of passwords, private messages and financial details, and Microsoft delayed it for nearly a year to rethink its protections.
The controversy has not ended. A cybersecurity researcher recently built a tool that could extract and display data Recall had captured, reviving questions about whether the underlying problem has been fixed. Recall’s place in the AI graveyard is about trust rather than technology: a feature can work and still fail if people do not feel safe using it.
Startups and Products That Did Not Survive
Five entries are products that closed outright. Most were not bad ideas. They were ideas that larger platforms copied, or that cost more to run than users would pay.
Relay: automation that platforms absorbed
Relay launched in 2021 aiming to become the new Zapier. Free customers lost access on 15 August 2026 and paying customers on 14 September, after a closure first announced in July. Founder and chief executive Jacob Bank rejoined Google as vice president of product for Chrome, where he called the browser “a perfect place to collaborate with agents”.
Notion Mail: replaced by agents
Notion launched Notion Mail in April 2025 as an AI-focused email inbox and is shutting it on 22 September 2026. The reason is unusual. “Today, more than half of Notion Mail users manage emails without ever opening their inbox,” Notion said, so it is “going all in on using agents to run your inbox”. Here the AI graveyard entry is a product made redundant by its own company’s newer AI.
Huxe: podcasts became a feature
Huxe, founded in late 2024 by former NotebookLM developers Raiza Martin, Jason Spielman and Stephen Hughes, turned prompts into podcast-style audio. It raised $4.6 million from investors including Conviction, Figma chief executive Dylan Field and Jeff Dean. It shut in May 2026, a day after Spotify released a similar personal podcast feature. Users had seven days before the app stopped working and their data was deleted.
Yupp: a big seed round was not enough
Yupp offered a free playground where users compared answers from about 800 AI models and earned cryptocurrency for voting. It said it signed up 1.3 million users and raised a $33 million seed round in 2024 led by a16z crypto’s Chris Dixon, plus more than 45 angel investors. It closed in March 2026, less than a year after launch, because it “didn’t reach a strong enough product-market fit”, in part because models improved so quickly.
Figgs AI: free was too expensive
Figgs AI ran from 2023 to 2024, letting users create and chat with customisable AI characters for role-play and storytelling. It reportedly drew more than a million users, but its developers said keeping the service free became too expensive to sustain. It is the clearest case on the list of compute costs outrunning a business model.
AI Hardware in the AI Graveyard
AI hardware appears twice on the list, and the two devices show different outcomes of the same problem: shipping before the software was ready.
Humane AI Pin: $230 million raised, $116 million sale
The Humane AI Pin is probably the best-known failure in AI hardware. Humane, founded by former Apple employees Bethany Bongiorno and Imran Chaudhri, raised more than $230 million. It launched the Pin in April 2024 at $699 as a smartphone alternative. Reviews were poor, returns at one point outpaced sales, and Humane warned customers to stop using the charging case over battery fire concerns.
In February 2025, HP bought most of Humane’s assets for $116 million, about half of what the company had raised. Humane had reportedly sought a sale at $750 million to $1 billion less than a year earlier. The Pins stopped working on 28 February 2025, and only buyers from the previous 90 days qualified for refunds. For owners, this AI graveyard entry meant a device that simply stopped.
Rabbit R1: still alive
The Rabbit R1, unveiled at CES in January 2024, sold 100,000 units shortly after launch, according to Rabbit. Early reviews described it as unfinished, with unreliable performance and few useful integrations. Rabbit has kept updating it, now positioning it as a controller for agentic tasks, and has announced “Project Cyberdeck”, a portable device for vibe coding. It is on the list as a warning, not an obituary.
Why Projects End Up in the AI Graveyard
Looking across all twelve entries, the causes fall into a small number of groups. Several products fit more than one.
Absorbed by a bigger platform
The most common cause is a bigger company shipping the same feature to an existing audience. Relay lost to automation inside OpenAI and Google tools. Huxe lost to Spotify and NotebookLM-style features. Atlas and Operator became parts of ChatGPT. As TechCrunch’s Huxe coverage put it, core products of startups “often turn into commoditized features of large companies”.
Compute costs that outran revenue
Sora’s roughly $1 million a day and Figgs AI’s free service both show that generative AI‘s running costs can sink a product with real demand. Video, image and character chat are among the most expensive things to serve per user.
Trust, hardware and overpromising
Recall shows that privacy concerns can stall a working feature. Humane and Rabbit show that AI hardware magnifies software weaknesses, because customers pay up front. Apple’s settlement shows that marketing ahead of delivery has legal as well as reputational costs.
| Cause | Entries | Warning sign for buyers |
|---|---|---|
| Absorbed by a platform | Relay, Huxe, Atlas, Operator, Notion Mail | The product is one feature a large vendor could add |
| Compute costs | Sora, Figgs AI | Free or cheap plans for expensive media generation |
| No product-market fit | Yupp, ChatGPT super app | Usage without a clear paying customer |
| Hardware before software | Humane AI Pin, Rabbit R1 | Up-front device cost tied to cloud services |
| Trust and privacy | Microsoft Recall | Features that capture sensitive data by default |
| Overpromising | Apple Siri AI | Marketing ahead of shipped capability |
How long the products lasted
For the closed products with clear start and end dates, we counted the months between public launch and shutdown. ChatGPT Atlas launched on 21 October 2025 and was discontinued on 9 August 2026, 292 days later. The chart compares these approximate lifespans with the 26-month average that the 404Tomb tracker reports across its archive, scaling each bar against that 26-month figure.
How Big the AI Graveyard Really Is
TechCrunch’s list is curated and short. Independent trackers try to count everything, and their numbers give a sense of scale, with caveats about how they collect data.
42% of AI initiatives abandoned
The S&P Global Market Intelligence figure cited by TechCrunch, that about 42% of AI initiatives are ultimately abandoned by their corporate parents, is the broadest measure. It counts internal projects that never become public products, so most of this AI graveyard is invisible from outside.
Trackers counting hundreds of dead products
The AI Graveyard website, aigraveyard.org, says it records “$25.5B of capital burned across 99 obituaries”. That averages about $258 million per entry ($25.5 billion ÷ 99), a figure pulled up by large failures such as autonomous-vehicle and health AI ventures. Its most recent entry was ChatGPT Atlas on 9 August 2026.
The 404Tomb directory lists 597 dead AI products across 21 categories and 11 years, with an average lifespan of 26 months. 404Tomb notes that most of its archive is community-submitted and not independently verified, apart from a set of verified case studies, so its counts are indicative rather than exact.
Money involved in this year’s list
The money on TechCrunch’s list ranges from a few million dollars to hundreds of millions. The chart below shows the figures TechCrunch reported, plus one estimate: Sora’s roughly $1 million a day over about six months, or around 180 days, suggests about $180 million in compute. Bars are scaled against Humane’s $230 million.
Funding does not buy survival
Yupp’s $33 million and 1.3 million users worked out to about $25 raised per user ($33 million ÷ 1.3 million), and it still closed. Humane’s buyers recovered about half of what investors put in. A large round buys time to find a paying market. It does not guarantee one.
Lessons for Businesses Buying AI Tools
For most businesses the AI graveyard is not an investment story. It is a supplier risk story. Every entry on the list had customers who relied on it and then had to move.
Check whether the product is a feature
Ask whether a large platform you already use could ship the same capability. Relay, Huxe and Notion Mail all lost to features inside bigger products. If the answer is yes, prefer the platform feature or make sure the smaller vendor gives you something the platform cannot.
Insist on data export and notice periods
Huxe gave users seven days before deleting their data. Humane’s devices stopped working with a few weeks’ warning. Before adopting an AI tool, confirm you can export your data, prompts, workflows and outputs in a usable format, and look for a contractual notice period for shutdowns.
Keep an exit plan for every AI dependency
List which AI tools sit inside critical processes and what you would switch to. Switching suppliers is easier with a written plan, as with any IT service; our guide to switching IT support providers without disruption sets out the same discipline. Test the fallback before you need it.
Watch pricing that looks too generous
Figgs AI and Sora show that heavy media generation on free or cheap plans can be unsustainable. A price well below what the service must cost to run is a signal to ask how the vendor makes money, not only a bargain.
Review privacy before you deploy
Recall is a reminder that AI features which capture screens, emails or conversations create new stores of sensitive data. Run a privacy and security review before rollout, not after staff have already switched a feature on.
| Risk signal | Question to ask the vendor | Protection |
|---|---|---|
| Single-feature product | What stops a major platform copying this? | Short contracts, platform fallback |
| Heavy compute on low prices | How do unit economics work at our volume? | Budget for price rises |
| No export tools | Can we export everything in a standard format? | Regular exports, data clauses |
| Hardware tied to a cloud service | What happens to devices if the service ends? | Refund and end-of-life terms |
| Sensitive data capture | Where is captured data stored and who can read it? | Privacy review, off by default |
What the Survivors in the AI Graveyard Have in Common
Not every entry is final. Several companies on the list turned a failure into a next step, and the pattern is useful.
Pivot to where users already are
Notion is closing Notion Mail but keeping the email agents users actually adopted. OpenAI kept Atlas and Operator’s best features inside ChatGPT. Rabbit is repositioning the R1 around agentic tasks. In each case the company followed observed behaviour instead of defending the original product.
Ship late rather than never
Apple’s Siri AI and Microsoft’s Recall both arrived years later than promised, at real cost. Both are still alive. Delay damaged trust and triggered a settlement in Apple’s case, but it did not end the product. The AI graveyard mostly holds products that ran out of money, users or patience before they could be fixed.
People move on to bigger platforms
Relay’s founder joined Google Chrome. Some Yupp staff joined a “well-known” AI company, according to its chief executive. Humane’s engineers became HP IQ. Talent from failed AI products often reappears inside the platforms that outcompeted them, which is one reason the same ideas return in new forms. For more on how talent moves between labs, see our report on OpenAI’s run of executive departures.
Frequently Asked Questions About the AI Graveyard
What is the AI graveyard?
The AI graveyard is TechCrunch’s running list of AI products, startups and bets that shut down, pivoted or badly missed expectations, published on 15 September 2026. Independent websites also use the name for larger databases of dead AI products.
Which AI products are on the list?
Relay, OpenAI’s ChatGPT super app redesign, ChatGPT Atlas, Operator and Sora, Apple’s Siri AI, Microsoft Recall, Notion Mail, Humane’s AI Pin, the Rabbit R1, Huxe, Yupp and Figgs AI.
Why did OpenAI shut down Sora?
TechCrunch, citing a Wall Street Journal investigation, reported that Sora’s users peaked at about one million and fell below 500,000 while the app burned roughly $1 million a day in compute, so OpenAI freed the chips for other priorities.
How many AI initiatives fail?
S&P Global Market Intelligence research cited by TechCrunch found that about 42% of AI initiatives are ultimately abandoned by their corporate parents. Trackers such as 404Tomb list hundreds of discontinued AI products.
How can a business avoid being hurt by an AI shutdown?
Choose vendors whose product is more than one easily copied feature, insist on data export and shutdown notice terms, keep a tested fallback for every critical AI tool, and review privacy before deploying features that capture sensitive data.
References
The AI graveyard: a running list of projects and startups that didn’t make it (TechCrunch)
AI automation startup Relay shuts down, staff joins Google’s Chrome team
Why OpenAI really shut down Sora
Notion Mail shuts down amid agent takeover
Audio-generation app Huxe shuts down
Yupp shuts down after raising $33M from a16z crypto’s Chris Dixon
Humane’s AI Pin is dead, as HP buys startup’s assets for $116M
Apple to pay $250M to settle lawsuit over Siri’s delayed AI features
Generative AI shows rapid growth but yields mixed results (S&P Global Market Intelligence)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.