The AI market is growing rapidly, but far from all the projects that companies and investors placed big bets on are proving viable. Some startups shut down because they cannot withstand competition from OpenAI, Google, and other giants. Others fall victim to their own technical problems, the high cost of development, or a lack of demand. According to S&P Global Market Intelligence, about 42% of corporate AI initiatives are ultimately shut down by the companies that launched them.

Against this backdrop, a kind of “graveyard” of AI projects is gradually taking shape. It includes both small startups and products from the largest technology companies. Some of them cease to exist entirely, others turn into features within larger services, and still others continue to develop despite an unsuccessful start.

OpenAI is also shutting down its own AI projects

As TechCrunch reports, even OpenAI has not managed to avoid failed launches. One of the most notable recent examples was an attempt to turn ChatGPT into a more universal app.

On July 9, the company updated the app by combining several separate modes in it — Chat, Codex, and Work. At the same time, the familiar version of ChatGPT was renamed ChatGPT Classic. Users quickly criticized the new interface for being overly complex and cluttered, after which OpenAI brought back the previous version of the app.

At the same time, the company has for some time been moving in a different direction — gradually abandoning standalone products and transferring their functions directly into ChatGPT.

This is what happened with Operator, an AI agent that could independently interact with websites and carry out tasks for the user. ChatGPT Atlas, a standalone AI browser, met a similar fate: it lasted less than a year and was shut down on August 9, while its most in-demand capabilities were integrated into ChatGPT.

DALL-E’s standalone role is also shrinking. Image generation is being integrated ever more deeply directly into ChatGPT, so a separate service for working with DALL-E is becoming less necessary.

Another example is the video service Sora. It shut down in March 2026 after problems with high operating costs and user retention.

This creates a paradoxical situation: some OpenAI products do not disappear because of a lack of technology, but simply stop existing as standalone products because it makes more sense to build their functions into the company’s core service.

Siri AI: several years of delays

Apple faced a different problem. In 2024, the company introduced Apple Intelligence, making the updated Siri one of the central parts of its new AI strategy.

Apple promised an assistant that would be able to understand context, take into account the user’s actions across different apps, and independently perform more complex tasks.

However, the launch was repeatedly postponed. Among the reasons cited were engineering problems and system malfunctions. The delays eventually led, among other things, to legal proceedings over Siri’s advertised capabilities, which ended with Apple agreeing to pay $250 million.

The new version of Siri finally appeared in the iOS 27 beta in July 2026. In August, the AI features began rolling out to English-speaking users, and Apple plans to add support for other languages later.

The story of Siri shows another side of the AI race: even a company with enormous resources can run into the fact that promised features turn out to be significantly harder to implement than originally expected.

Microsoft Recall: when AI collides with privacy

Microsoft Recall was originally conceived as a kind of “photographic memory” for a computer. The feature was supposed to periodically save screenshots so that the user could later find virtually any moment from their digital history.

The idea quickly raised security concerns. The saved screenshots could potentially include passwords, personal messages, financial information, and other sensitive data.

After the criticism, Microsoft postponed the launch for almost a year and reworked the security and data protection system.

However, the problems did not end there. A cybersecurity researcher recently created a tool capable of extracting and displaying information that Recall had saved on a computer. This once again raised questions about how reliably the feature protects the data it collects.

Recall ultimately became an example of how the problem with an AI product may lie not so much in the quality of the model itself as in the consequences of its integration into everyday devices.

Notion Mail couldn’t withstand competition from AI agents

Notion launched Notion Mail in April 2025. The service was supposed to help users organize email using AI and automate work with incoming messages.

But the company discovered an unexpected problem: its own users had started using standalone AI agents for the same tasks.

As a result, Notion decided to shut down Notion Mail. The service is scheduled to be permanently discontinued on September 22, 2026.

This is a rather telling case for the AI market: a product may turn out to be unnecessary not because it performs its function poorly, but because that same function becomes part of a more universal AI agent.

Humane AI Pin: ambitious hardware turned out to be too difficult

Humane AI Pin became one of the most famous failures among AI devices. The company wanted to replace the familiar smartphone with a small wearable device through which the user could interact with AI.

Humane raised $230 million in investment, and the product itself received enormous attention even before its release.

But after launch, it became clear that the device had serious performance problems. The situation was worsened by the company’s warning about a potential danger involving the charging case: users were advised to stop using it because of the risk of battery fire.

In February 2025, Humane shut down the business related to AI Pin. Most of the company’s assets were acquired by HP for $116 million.

The story of the Pin showed how difficult it is to turn AI capabilities into a standalone physical device capable of competing with the smartphone the user already has.

Rabbit R1: huge interest does not guarantee success

Rabbit R1 was unveiled at CES in January 2024 as a compact AI companion capable of performing various actions on behalf of its owner. The company said that it sold 100,000 devices shortly after launch.

However, sales did not guarantee the quality of the product. Early reviews pointed to the device’s immaturity, unstable performance, and a small number of genuinely useful integrations.

Unlike Humane, Rabbit has not shut down the project so far. The company continues to update R1 and is gradually positioning it as a computer controller capable of performing tasks with the help of AI agents.

In addition, Rabbit is working on a new device, Project Cyberdeck — a portable computer aimed at so-called vibe coding.

So for now, R1 is more an example of an unsuccessful launch that the developers are trying to turn into a more viable product.

Huxe: a good product turned out not to be independent enough

Huxe was created by former Google NotebookLM developers. The app turned written materials into audio in the format of a conversational podcast.

At first glance, such a service had a clear niche. However, major platforms began offering similar capabilities to their own audiences.

For example, Spotify is gradually expanding its set of AI tools for working with podcasts, including audio generation and summary features.

In May 2026, Huxe shut down. This is another example of a problem faced by small AI startups: even if they are the first to find an interesting use case for the technology, a large platform can incorporate a similar feature into an existing product fairly quickly.

Yupp: 800 models turned out to be not enough

Yupp offered an unusual approach to choosing AI models. A user could send one prompt to several models at once and compare the responses side by side.

At its peak, the platform allowed users to test more than 800 models, including developments from OpenAI, Google, and Anthropic. Users could vote for the best responses and also receive cryptocurrency for their activity.

But Yupp’s creators concluded that the service never managed to find a sufficiently sustainable development model and the necessary product-market fit.

The platform shut down in March 2026.

Figgs AI: even a million users were not enough to save it

Figgs AI existed from 2023 to 2024 and allowed users to create their own AI characters for role-playing games and stories.

According to available data, the project attracted more than a million users. However, the developers ran into a different problem: keeping the service free was becoming too expensive.

As a result, Figgs AI also had to shut down.

And this is perhaps one of the main characteristics of today’s AI market. A large number of users by itself still does not mean that a product can exist as a business. Generating content with AI requires computing resources, which means that a popular free service can at the same time turn out to be the most expensive to maintain.