AI 墓地中的三種消亡

The AI graveyard is not a list of failures. It is a list of absorptions — and the platforms doing the absorbing are not exempt from dying themselves.

TechCrunch's AI graveyard keeps growing: Relay, Character.ai, ChatGPT Atlas, Apple's delayed Siri, Microsoft's Recall, the Humane AI Pin, the Rabbit R1, and more. Read it as a scorecard and you miss the story. Sort it, and three different kinds of death fall out — and the honest answer is that they may not add up to one pattern at all.

The ones the platform ate

Relay built AI workflow automation as a Zapier alternative. It shut down entirely in August, because OpenAI, Google and the other platforms simply built the same automation into tools users already had. TechCrunch's verdict was brutal and accurate: it "struggled to hold on to a reason to exist as a stand-alone product."

That is the purest form of displacement: the feature survived, the product did not. Huxe went the same way in May 2026 — an AI audio app from former NotebookLM developers, squeezed as Spotify expanded its own AI audio tools to an audience Huxe could never match. If your entire moat is a feature a platform can ship in a quarter, you do not have a company. You have a demo with a pricing page.

The ones that were folded in

The second kind of disappearance is stranger, because nothing actually died. ChatGPT Atlas, the stand-alone AI browser, lasted less than a year before being discontinued on August 9 — its best features absorbed into ChatGPT. Operator, the agent that browsed the web for you, was folded in the same way. Image generation moved inside ChatGPT, shrinking DALL-E from a destination to a button.

Call this consolidation, not failure. A folded-in feature reaches more users, not fewer — which is exactly why the graveyard misleads if you read it as a flop list. The standalone app died; the capability lived on inside the flagship. Platforms love to call that success, and in this case they are not entirely wrong.

The flagship stumbles

Then there is the half of the graveyard nobody at the platforms wants to talk about: their own bets, failing in public. OpenAI tried to turn ChatGPT into a "super app" on July 9, mashing Chat, Codex and Work together and renaming the familiar version "ChatGPT Classic." Users called it confusing and cluttered, and OpenAI rolled it back eight days later. President Greg Brockman admitted it plainly: "It's kind of a mess. We agree."

Apple's improved Siri, promised with Apple Intelligence in 2024, was delayed repeatedly over engineering issues — a delay that contributed to a $250 million settlement over iPhone 16 AI marketing — before finally appearing in the iOS 27 beta in July 2026. Microsoft's Recall, the "photographic memory" that screenshotted your PC, triggered an immediate privacy backlash, was delayed nearly a year, and is still haunted by a researcher's tool that could extract what Recall captured. Even Sora, OpenAI's video platform, shut down in March 2026 under operating costs and retention problems.

Nobody is exempt — not even the companies doing the absorbing.

The $20 million opening bid

Character.ai belongs to none of the three kinds neatly, which is why it is the graveyard's most dramatic entry. It raised $150 million at a $1 billion valuation in 2023. In 2024, Google paid a reported $2.7 billion to license its technology and bring its founders back. Then lawsuits over teen safety piled up, Disney sent a cease-and-desist over its characters, and the company filed for bankruptcy in early October. Disney is now the opening bidder — offering $20 million for a license to some of its intellectual property.

A $1 billion valuation, a $2.7 billion licensing deal, and a $20 million opening bid are three different kinds of figure, not one company's falling price. Conflating them is exactly the sloppy reading the graveyard invites.

The pattern is still a question

Zoom out and the scale sharpens: S&P Global Market Intelligence found 42% of companies abandoned the majority of their AI initiatives in 2025, more than double the 17% of the year before, with organisations scrapping 46% of projects between proof of concept and broader adoption. That is context, not a verdict — most of those initiatives never reached the public, while the graveyard is a public sample.

So here is the take: do displacement, consolidation and flagship failure add up to one story — the platform absorbing the feature — or are they three separate stories sharing a list? The three kinds of disappearance are confirmed. The single pattern is not. And that unresolved question is more useful than any scorecard, because it forces the question every AI builder should be asking: are you building something a platform cannot eat, or are you just waiting for your turn in the graveyard?

Go deeper

Read the full analysis on Agentic Affairs: The AI Graveyard Has Three Kinds of Dead

Originally published on Agentic Affairs (https://agentic-affairs.com)