The moment when hype meets reality
For three years, generative AI was sold as the revolution that would transform everything. Capital poured in, valuations soared, on-stage demos turned heads. And then, from late 2024 onwards, the first major products began to die. Not obscure startups lost in the noise. Products backed by the biggest tech companies in the world, presented as the future, and which crashed publicly.
Five of these failures deserve closer scrutiny, because they tell us something precise about the real state of the AI market in 2026. The technology is advancing. Users are curious. And yet the majority of major industrial bets are failing. Here is why.
Sora (OpenAI): $15 million a day, lost
On 26 April 2026, OpenAI announced the shutdown of Sora, its flagship video generation app. The app had launched in September 2025 amid a wave of spectacular enthusiasm, becoming number one on the App Store within days. Six months later, it was dead.
The numbers explain everything. According to several corroborating sources, Sora was costing OpenAI $15 million a day in compute (the heavy calculations needed to generate each video). Over its entire commercial life, it generated $2.1 million in revenue. In total. Not per day, not per month, over the full six months. A 10-second video requires generating 240 to 300 separate frames, each costing compute power. The economics were fundamentally broken from day one, and nobody knew how to save it.
Bill Peebles, who led Sora's development, had warned internally as early as October 2025 that the product's economics were "completely unsustainable". OpenAI carried on regardless for another six months. Sam Altman eventually justified the shutdown by talking about "reallocating resources towards automated researchers and enterprises", a polite way of saying that Sora was burning money intended for more profitable projects.
The most spectacular collateral effect: Disney had signed a billion-dollar deal with OpenAI, including the integration of Sora into its production pipelines. The studio learned of Sora's shutdown less than an hour before the public announcement. The partnership collapsed on the spot. Disney simply responded: "We respect OpenAI's decision." It doesn't get more glacial than that.
Apple Intelligence: the promise with a $250 million fine
At WWDC in June 2024, Apple presented Apple Intelligence as the revolution that would modernise Siri. Adverts saturated the internet, television and the airwaves. iPhone 16s were sold with the promise of this new Siri capable of understanding personal context, acting across apps, transforming the iPhone experience.
>>Nearly two years later, in May 2026, nothing has been delivered. Apple has repeatedly delayed the new Siri: first planned for spring 2025, then 2026, and now iOS 27, expected in June 2026. And in the meantime, the company did something that would once have seemed impossible: it pays Google $1 billion a year to use Gemini instead of its own models. The deal was finalised in January 2026.
The legal outcome landed in May 2026. Apple agreed to pay $250 million to settle a consumer class action brought by users who felt they had been misled. Eligible iPhones (bought between 10 June 2024 and 29 March 2025) are entitled to compensation of between $25 and $95 per device. The complaint accused Apple of having sold "AI capabilities that did not exist at the time, still do not exist, and will not exist for two years or more".
| Metric | Siri (Apple) | Google Assistant |
|---|---|---|
| Answer accuracy | 83.1% | 92.9% |
| Error rate | 1 answer in 6 | 1 answer in 14 |
| Delay against the announced schedule | 24 months and more | Benchmark |
For a company worth $3.8 trillion at the launch of that campaign, and which presents itself as synonymous with excellence, it is a hard blow. And it serves as a reminder of something fundamental: AI is no more magical for Apple than for anyone else. When the product does not work, the courts eventually take care of it.
Microsoft Designer: quietly buried in Teams
Microsoft Designer was pitched as the product that would democratise graphic design for Microsoft 365's 350 million users. Launched in 2024 with native integration in Word, PowerPoint, Outlook and Teams, it was meant to be Microsoft's answer to Canva.
On 27 February 2026, Microsoft quietly removed Designer from Teams. No triumphant press release, no keynote. Just a notification in the Microsoft 365 admin centre under ID MC1197104. The Designer bot and Designer banners were replaced by Copilot in Teams. The rollout of the removal had begun silently in mid-January 2026.
Before burying it, Microsoft pulled a commercial manoeuvre that drove users away. Designer launched with free unlimited generations in 2024. Once people had become dependent, Microsoft capped it at 15 generations per day, then turned Designer into a paid Microsoft 365 feature (60 AI credits per month). User forums erupted: "Like a dealer who gets his customers hooked and then makes them pay." Designer did not die by accident; it died of its vendor's greed.
The most ironic part: according to Microsoft's internal figures, only 3.3% of Microsoft 365 users who interact with Copilot Chat pay to use it, despite the $37.5 billion the company invested in AI last quarter. Microsoft is therefore replacing a specialised AI product that worked with a general-purpose AI product that almost nobody wants to pay for. Corporate logic, illogical for the user.
Humane AI Pin: $230 million raised, remotely deactivated
Launched in 2023 by former Apple executives, funded by OpenAI, Microsoft, Salesforce and Marc Benioff personally, Humane had raised $230 million to revolutionise mobile computing. The AI Pin, a small device clipped to a jacket lapel, was meant to replace the smartphone through voice interaction and a laser projector onto the hand.
The April 2024 launch was a disaster. Tech reviewers described the device as "bad at almost everything it does". Overheating, a battery life of a few hours, sluggishness, a laser projector unreadable in daylight. The product sold for $700 plus $24 per month for a mandatory subscription, and you still had to keep your smartphone nearby for essential functions.
On 28 February 2025, Humane sold its assets to HP for $116 million (half of what it had raised) and every Pin already sold was remotely deactivated. Overnight, users were left with an inert device clipped to their jacket. Fewer than 10,000 units had been sold in total.
Rabbit R1: the demo that did not deliver
Unveiled at CES 2024, the Rabbit R1 caused a stir with its demo: a small orange gadget capable of ordering an Uber, booking a restaurant and managing the user's applications via its "Large Action Model", a revolutionary new approach to AI.
100,000 units sold on enthusiasm. And then delivery. The CES demo was largely faked: what was presented as an autonomous AI actually used pre-recorded scripts. The shipped product could do almost none of what had been shown on stage. Returns were massive. By 2026, Rabbit is struggling to pay its staff, according to several reports.
Microsoft Tay: the classic that taught nobody anything
To put all this in perspective, let us recall an older but still instructive case. In 2016, Microsoft had launched Tay on Twitter, a chatbot designed to engage 18-24 year-olds in natural conversations. In less than 24 hours, Tay had been hijacked by users and had become racist, antisemitic and a Holocaust denier. Microsoft had to pull the plug on the bot in a panic.
Ten years later, the same patterns are repeating. A large company launches an AI product to the general public, without having gauged the collision with reality. And then reality sends the bill. Tay made people laugh, Sora lost billions, Apple is paying in cash. The lesson never seems to sink in.
The five patterns that keep repeating
If you look at these failures as a whole, five patterns come back systematically. And each deserves to be named, because it probably predicts the next failures to come.
Pattern 1: the economics broken from the start. That is the Sora case. The technology works, users love it, but the product costs more to run than it brings in. Nobody knows how to fill the hole, so they cut it. This will probably happen to several video and image products in the next 12 months.
Pattern 2: the promise that exceeds delivery. That is the Apple case, Rabbit, and many others. The demo is spectacular because it is cherry-picked or faked. The shipped product does not resemble the demo. Users sue or get refunds. Apple pays 250 million, Rabbit struggles to pay its staff.
Pattern 3: hardware with no clear use. That is Humane, Rabbit, and the whole craze for the "dedicated AI device". The fundamental question never got an answer: what does this gadget do better than my smartphone, which already contains the same AI? Without an answer to that question, the product dies mechanically.
Pattern 4: the startup swallowed by the giants. That is Microsoft Designer absorbed into Copilot. It is also what happens to many third-party AI products when OpenAI, Google or Microsoft launch an equivalent feature built into their ecosystem. As we analysed in our article on the OpenAI lock-in, the consolidation of super-apps mechanically crushes specialised tools.
Pattern 5: AI released into the real world too early. That is Tay in 2016, but it is also the PocketOS case we mentioned in our article on keeping AI under control. Engineers cannot predict how their product will behave in front of real users or unexpected situations. And they launch it anyway.
What these failures really tell us about 2026
Beyond individual cases, there is something broader to be heard. For three years, the dominant narrative was: "AI will transform everything, get on board quickly or you'll be left behind". That narrative fuelled hundreds of billions of dollars of investment and dozens of products launched at breakneck speed, sometimes without a viable business model, sometimes without any clear advantage over what already existed.
In 2026, the market is starting to correct. Not because AI is disappointing in absolute terms, but because the classic economic filters (profitability, real usefulness, customer satisfaction) are starting to apply to a sector that had until then been exempt from them. Sora costs too much: it gets cut. Designer didn't find its audience: it gets buried. Apple promises and doesn't deliver: it pays 250 million.
For businesses and users, the practical lesson is clear: in 2026, treat AI announcements with more caution. A spectacular demo doesn't mean a shippable product. A shippable product doesn't mean a profitable product. And a profitable product today doesn't mean a lasting product tomorrow. The test is time passing and real users staying.
The real AI successes of 2026 are not the most hyped products. They are the ones that have been working quietly for several years and have been through the fire. ChatGPT, which despite its economic problems keeps growing. Claude Code, which has transformed the work of millions of developers. Mistral, which has established itself in Europe without fanfare. The others will probably make the headlines of the next wave of burials.