OpenAI Built The Future Of Video, Then Pulled The Plug. What Were They Thinking?

Cast your mind back to late 2025: OpenAI launched Sora, its text-to-video model, and the internet did what the internet does: absolutely lost its mind.

The clips were truly impressive and the hype was enormous. It topped Apple’s App Store charts within days, hit one million downloads in a matter of hours and sparked the usual cycle of predictions about what this meant for filmmakers, advertising agencies, content creators and anyone else with a vested interest in moving images.

This week, OpenAI subtly announced it’s discontinuing Sora, shutting down both the app and the API preview, and redirecting the team toward “world simulation research to advance robotics.” The future of video lasted about as long as a Vine account.

What actually happened here, and what does it tell us about how AI labs make decisions?

From App Store Chart-Topper To Discreetly Shelved

The post-launch story of Sora is a fairly compressed version of the AI hype cycle.

Downloads hit a peak, then dropped 45% month-over-month by January 2026, with spending falling by 32%. The virality dissipated, as it tends to, when the novelty wears off and users discover that generating a useful video is harder and slower than generating impressive demo clips.

Behind the scenes, the economics of it all were worse. Sora lead Bill Peebles described the unit economics as “completely unsustainable”: video generation is extraordinarily GPU-intensive, and a small number of power users were apparently burning through compute at a rate that made the whole thing difficult to justify.

OpenAI, which is actively preparing for an IPO and under significant pressure to demonstrate a path to profitability, has a new product head in Fidji Simo, who has been explicit about deprioritising “side quests,” in favour of profitable core products like ChatGPT.

There was also the small matter of companies churning out AI features that add little value being a recurring theme in this industry. Sora’s early months were peppered with content moderation problems, IP violations (Pokémon in war scenes being a particular highlight), a lawsuit from Cameo over its “cameo” feature, and a Disney partnership for character licensing that now looks somewhat academic.

The product had real promise, but the gap between impressive demo and reliable everyday tool proved wider than the launch suggested.

Killing Sora Was Actually The Smart Move

The truth is: framed as a business decision, killing Sora is completely rational.

OpenAI’s strategic priorities have always been about the models, not the apps. The real asset is the underlying capability, and video generation turns out to be a less efficient vehicle for advancing that capability than world simulation for robotics, which is where the Sora team is apparently headed. Training robots to understand physical environments is a much closer path to the general intelligence goals that OpenAI’s mission is organised around. Video as consumer entertainment is, from that perspective, a distraction.

This fits into a notable trend in how AI labs operate, that founders should be paying close attention to. Labs move fast, they launch things with significant fanfare, and then discontinue them when overall strategy changes.

That’s what ultimately happens when you’re a research organisation, that also happens to be running consumer products, while preparing for an IPO, while also trying to build AGI. Something gets deprioritised, and this time it was Sora.

So, About That Startup You Built On Sora…

This is where the focus moves away from OpenAI and onto everyone building on these platforms.

Sora had an API preview and developers were building on it: video product startups, creative tools, custom character workflows. Those products are now in a tricky position through no fault of their own. There’s no reported fallout yet, but any startup that built a product around Sora’s API just had the rug pulled without warning. That’s a pattern worth taking seriously.

The lesson isn’t that you shouldn’t build on AI APIs – that would be absurd advice in 2026. The lesson is that you should build on them with clear eyes about what you actually control.

A startup whose core value proposition is “we use Sora to do X” has a different risk profile to a startup that uses Sora as one of several possible backends for a product whose real value is in the workflow, the integrations, the customer relationships and the domain expertise.

OpenAI isn’t unique in this – every major AI lab makes strategic pivots, every platform dependency carries risk. The degree to which your business survives a pivot like this one depends entirely on how much of the value you’ve retained for yourself, versus how much you handed to the API provider.

The Lesson Nobody Wanted To Learn This Way

The convenient narrative is that OpenAI made a mistake, launched too early, got distracted by the hype and is now retreating. That’s a neat story and probably not entirely wrong.

But the more useful read is this: AI labs are making product decisions at a pace and with a pragmatism that the rest of the industry isn’t always built to accommodate. When the compute costs don’t justify the strategic benefit, products get cut; when a research direction looks more valuable than a consumer app, the team moves; and when an IPO is on the horizon, ‘side quests’ get quietly deprioritised.

None of that is surprising if you understand what these companies actually are. The surprise only comes if you mistake the consumer product for the mission, and Sora was never the mission. It was a demonstration of a capability, and the capability has now been redirected somewhere OpenAI thinks matters more.

The strategic logic is sound, the collateral damage is someone else’s problem.