The AI gold rush is leaving bodies in its wake, and the casualty list tells you more about where this is really heading than any pitch deck ever could.

The Summary

  • TechCrunch is now tracking a formal "AI graveyard" of failed projects and startups, including Apple's delayed Siri AI overhaul and OpenAI's botched super app launch
  • The pattern of failures reveals what actually doesn't work in AI: thin wrappers around foundation models, consumer tools without distribution, and features searching for problems
  • Survival in AI requires either defensible infrastructure, embedded distribution channels, or solving a problem expensive enough that accuracy matters more than perfection

The Signal

The graveyard isn't just growing. It's getting interesting. Apple delayed its Siri AI transformation multiple times because the accuracy wasn't there, and when you control a billion pockets, launching broken software is more expensive than waiting. Meanwhile OpenAI's attempt at a super app landed with the thud of a company that forgot distribution is harder than generation.

The failures cluster around three patterns. First: wrapper companies. Any startup that was just GPT-4 with a prettier interface is either dead or shambling toward irrelevance. Second: consumer AI tools that require behavior change. People don't want to learn new workflows, they want their existing workflows to work better. Third: solutions to problems that don't cost enough. If the pain point is mild annoyance, AI isn't your business model, it's a feature some incumbent will ship for free.

"The AI graveyard is filled with companies that built features, not businesses, and mistook investor excitement for product-market fit."

What survived tells the better story:

  • Infrastructure plays with actual moats: custom chips, fine-tuned vertical models, proprietary training data
  • Enterprise tools embedded in existing software budgets: AI that lives inside Salesforce, not competing with it
  • Agents with clear ROI: customer service bots that eliminate headcount, coding assistants that ship features faster

The OpenAI super app failure is particularly instructive. They tried to become a consumer destination when their actual value is being invisible infrastructure. ChatGPT works because it's a research preview that became a household name. A super app with calendar, email, and task management? That's competing with Apple, Google, and Microsoft on their home turf, with none of their distribution advantages. Bad bet.

Apple's Siri delays reveal a different constraint: when you're already in a billion pockets, your risk tolerance drops to near zero. Shipping half-baked AI when you're a startup means some bad tweets and a pivot. Shipping it as Apple means congressional hearings and brand damage that takes years to repair. They're learning what OpenAI is learning: consumer AI has to be embarrassingly good, not just impressively better than nothing.

The Implication

If you're building in AI right now, the graveyard is your business plan review. Ask: are you infrastructure that other companies will pay to access, or are you a feature that will be free in six months? Do you have distribution, or are you hoping to get it? Is the problem you're solving expensive enough that "pretty good" is worth paying for?

The survivors won't be the ones with the best models. They'll be the ones who understood that AI is a capability, not a category, and that the real work is figuring out where to apply the capability so that someone opens their wallet. The graveyard gets bigger before the industry gets smarter. Read the headstones. They're warnings.

Sources

TechCrunch AI