The AI graveyard is a better roadmap than any launch keynote
A running list of dead and delayed AI products tells you more about where the industry is headed than a stage demo does. Here is why readers should weigh failures more.
AI-generated image
Which tells you more about the future of AI: a polished keynote demo or a list of the things that quietly stopped existing? We would bet on the list every time.
The list in question is TechCrunch's running "AI graveyard", which tracks products and startups that shut down, pivoted or missed expectations. It runs from Apple's repeatedly delayed Siri AI to OpenAI's "super app" misfire. It also cites S&P Global Market Intelligence's estimate that about 42% of AI initiatives are ultimately abandoned by their corporate parents.
Demos show what's possible; tombstones show what people actually use
A demo answers one question: can this work, once, under ideal conditions? A shutdown answers a harder one: did anyone keep using it? The graveyard is full of products that passed the first test and failed the second.
Take OpenAI's July 9 update, which combined "Chat," "Codex" and "Work" modes and renamed the familiar version "ChatGPT Classic." According to TechCrunch, users found it confusing and cluttered, and OpenAI rolled it back soon after. No keynote would have told you that. The market did, and quickly.
Or take Notion Mail, which TechCrunch says was scheduled to shut down on September 22 after the company noticed its users were handling their inboxes with separate AI agents instead. That is a very specific lesson: people do not necessarily want AI baked into every app. Sometimes they want one agent that works across all of them.
The pattern: stand-alone AI products get swallowed by the platforms
Read the list as a map and one pattern jumps out. Relay, a workflow automation tool pitched as a Zapier alternative, shut down in August after OpenAI, Google and other larger platforms built similar automation into their own tools. ChatGPT Atlas, the stand-alone browser, lasted less than a year before OpenAI folded its most useful features into ChatGPT. Operator got the same treatment.
Our reading is that the "AI feature versus AI product" question has mostly been settled, and the feature is winning. If your product is something a general-purpose assistant can do as a side effect, you are on borrowed time. That is bad news for a lot of startups. It is also an uncomfortable signal for the giants, because their own stand-alone experiments keep ending up as menu items.
The Browser Company's Arc is a slightly different case. It is being sunset on December 22, with the company pointing to the rising cost of keeping a browser secure in the age of AI. We would read that as a warning: the maintenance bill for AI-era software may be the quiet killer that never shows up in a demo.
Delays and backlash are data too
Not every entry is a corpse. Apple's new Siri slipped repeatedly, and the delay contributed to a $250 million settlement over how the iPhone 16's AI capabilities were marketed. The new version reached the iOS 27 beta in July and began rolling out to English-language users this month. Microsoft's Recall was delayed for nearly a year after a privacy backlash, and a researcher has since built a tool that extracts what it captures.
These are not failures of ambition. They are failures of the gap between the announcement and the product. And that gap is the number readers should track: how long between "coming soon" and "available," and what happened in between.
Hardware makes the point most brutally. Humane raised $230 million and shut down its AI Pin business in February 2025, with HP acquiring most of its assets for $116 million. Rabbit said it sold 100,000 units of the R1 shortly after launch, yet TechCrunch notes early reviews called the device unfinished. Strong launch sales are not a strong product.
What to do with this
We would add a caveat. A graveyard is not a scorecard of flops, and TechCrunch itself frames it as a record of lessons. Dying early is often what experimentation looks like, and some of these ideas may return in better shape. Also, a list assembled by one outlet is a selection, not a census.
Still, our advice is simple. When a company shows you a demo, ask which earlier product it replaces, what it costs to run and what happened to the last attempt. When a startup claims a niche, check whether a platform can absorb it with one update. And when you read a launch, find the matching entry in the graveyard.
Keynotes are marketing. Shutdown notices are the closest thing this industry has to honest feedback, and they're free to read.
No topics yet: start the first one.
More stories
Altman's "terrible" ending shows why scenario talk isn't AI safety
When OpenAI's CEO calls one outcome of the AI race terrible while reportedly saying the company can't afford to cede ground, incentives are doing the talking, not safety plans.
Nvidia the lender: when its best customers struggle to borrow, the moat leaks
By funding the neoclouds that buy its GPUs, Nvidia has tied its fortunes to their survival. As those stocks slide, circular financing looks less like a moat and more like exposure.
AI's missing payoff outside tech is the adoption gap the hype skipped
Apollo says non-tech companies show no AI profit gain. We think that means value is flowing to suppliers, and boards should demand measurable results now.
The AI boom now runs on credit, and credit is getting nervous
A pulled Solidigm IPO and distressed neocloud debt suggest the AI build-out has become a financing story, and the weak link is borrowed capacity without proven demand.
