There is a new kind of landmark in Silicon Valley these days — and it isn’t a glass skyscraper. It’s a changelog. When a company like Character.ai files for bankruptcy, or a privacy-first AI assistant named Ollie sends farewell notices to users, it doesn’t make the front page of a business newspaper. It makes a list. TechCrunch’s “AI graveyard” post from November 14 is the latest entry.
As someone who manages ICT systems and studies AI adoption in government and enterprise, I’ve been reading that list. It’s a useful one.
What’s in the graveyard
The graveyard is crowded. It’s a running record of AI projects, startups, and hardware bets that shut down, pivoted, or missed expectations. A few entries jumped out:
- Character.ai (2021) — founded by former Google researchers, raised $150M at a $1B valuation, then filed for bankruptcy in October 2026, six years after launch. Google had licensed its technology for $2.7B. Disney, of all companies, is the opening bidder on some of its IP.
- Ollie (2026) — a privacy-first AI assistant that claimed SOC 2, the gold standard for data controls, shut down November 6 after less than five months. CEO Bill Lennon blamed “Muse, Instinct, and others” for crowding him out.
- Sora (2026) — OpenAI’s video platform died in March, written off due to “high operating costs and user-retention challenges.”
- Apple’s Siri — the flagship AI bet of the last two years kept slipping. Apple eventually paid a $250M settlement over delayed AI features in the iPhone 16 line, and Siri AI only shipped in the iOS 27 public beta in July.
- Microsoft’s Recall — the “photographic memory” for Windows PCs was a failure of trust. Screenshots of your own screen, postponed for a year over security concerns, with the underlying vulnerability still open.
- The hardware bets — Humane’s AI Pin (sold for $116M to HP), Rabbit’s R1, and Yupp’s model-comparison playground all went dark.
The reasons are familiar: insufficient funding, technical challenges, competition, difficulty scaling, weak user demand. According to S&P Global Market Intelligence’s 451 Research, about 42% of companies abandon most of their AI initiatives before they reach production. That’s up from 17% a year earlier.
Why AI products fail
I’ve spent years building and shipping software as an ICT division manager. AI products fail for the same reasons every product fails — and a few that are unique to the era.
First, infrastructure cost is real. Sora died because operating costs and retention didn’t work. AI is computationally expensive. The infrastructure bill doesn’t pause when the publicity cycle ends. As the CoreWeave cash-crunch story showed, the economics of renting AI compute from public clouds are tightening fast. “I expected this to be cheaper,” one operator told me. The graveyard doesn’t record a word of that.
Second, distribution is the moat. Ollie had SOC 2. That’s a technical achievement. It wasn’t enough. Muse, Instinct, and others had distribution. “Trust is a differentiator, but it’s not a growth engine,” Lennon said.
Third, the market shifted under them. Notion Mail shut down September 22 because users had already migrated to separate AI agents to handle their inboxes. Even when airlines tried a different approach, United Airlines banned AI agents from booking flights after early experiments showed how hard it is to get the trust right. The product got to the party after the crowd had moved.
Fourth, the benchmark isn’t the product. A $2.7B Google acquisition, a SOC 2 audit, a $150M raise — these look like strong signals until the market tells you otherwise.
The bigger picture
There’s a deeper point that’s easy to miss. The graveyard is not a eulogy. It’s a curriculum.
A year ago, commentators worried about AI’s hype bubble. Today we’re watching the deflation, and it’s not the end of the technology; it’s the end of the first, sloppy chapter. ChatGPT is more popular than ever. Claude is everywhere. The underlying models keep getting better.
What’s dying is the shortcut: the idea that you don’t need a real product, a real workflow, a real reason for people to come back. That’s a good thing. It means the industry is maturing. AI agents are already moving into the wild — the rise of agentic shopping means LLMs now handle purchases, not just answers. That wave, in turn, is why companies are pulling back. The lessons from the graveyard apply directly to the next generation of agent products, and the companies navigating it are learning the same lesson we are.
What this means for builders
I’m not in the graveyard business. I build systems that stay up. A few takeaways stick with me:
- Don’t build a feature; solve a workflow. I’ve been testing the idea of “getting an AI to do something” versus “getting a workflow to work.” The graveyard is full of people who built a thing; the winners built a process. If you want to see how this plays out in practice, my step-by-step guide on how to run open-source AI models locally with Ollama shows the discipline of keeping the model on your own machine.
- Infrastructure dollars need a plan. As an ICT manager, I watch budgets. The 42% abandonment rate isn’t just a statistic; it’s a capital discipline problem. You can’t throw GPU hours at a distribution problem.
- Design for the next owner. We need to think about what happens when our AI is replaced. The graveyard shows us that products don’t last. If you want your work to outlive your product, you build what’s portable.
Conclusion
For developers and businesses looking at the graveyard: learn the lessons, not the names. The companies that fail are usually the ones asking “does this work?” when the market is asking “is this worth paying for?”
For me, as someone who manages technology in a Philippine government institution, the lesson is about discipline. Don’t let a shiny model become a substitute for a working process. Test the workflow, not the demo. Budget the compute, not the press release. Ship something people will use next month, not next decade.
Because the graveyard is full of projects that worked in demos. I should know — I’ve watched several of my own demos become expensive memories.
The most uncomfortable part of the graveyard is how safe the failures look in hindsight. Build a chat app. Raise money. Ship. It’s an undeniable formula that failed hundreds of times. The market doesn’t reward confidence. It rewards persistence against real demand.
That’s the part I want my colleagues to remember. The graveyard is full of people who were right about the future — just too early, too loud, and too slow at finding the people who actually needed it.
I used to think the next big thing would be the one that shipped first. The graveyard corrected me. The ones that survive are the ones that solve a problem people pay to solve. The rest are just expensive demos.
That’s the lesson I’m taking to my own team. The AI graveyard isn’t a warning. It’s a syllabus. And we’re all enrolled.
The AI graveyard isn’t a eulogy. It’s a curriculum. And the course is free.
For builders, the assignment is clear: solve a workflow, not a demo. For executives, the assignment is clearer: budget the compute, not the press release. And for everyone else, the assignment is the most important: don’t confuse a model with a product.
The technology isn’t going anywhere. The graveyard is just the tuition.