OpenAI Killed Sora — Is the AI Video Generation Path a Dead End?

Remember the scene when Sora was released in early 2024?

The entire internet exploded. “Hollywood is finished,” “Video creators will lose their jobs,” “The film industry will be completely disrupted” — apocalyptic narratives were everywhere. I wrote an article at the time saying “Don’t rush, let’s see how it actually works,” and got roasted in the comments.

Two years later, OpenAI itself killed Sora.

The official statement was “unsustainable computational costs.” In plain English: burning cash. Generating a 30-second video consumes hundreds or thousands of times more compute than text conversations, and what users are willing to pay? Nowhere near covering the costs.

Honestly, this outcome doesn’t surprise me.

AI video generation has a fundamental awkwardness — it’s stuck in the middle ground of “demo is stunning, product is鸡肋.” Ask Sora to generate a short clip of sea turtles swimming, it’s beautiful. But ask it to precisely control character expressions, camera movements, scene transitions, and lip-syncing according to a director’s vision? Sorry, can’t do it.

The applicable scenarios are too narrow; the scenarios users are willing to pay for are even narrower.

OpenAI’s strategic shift this time is quite interesting — kill Sora, go all-in on programming tools and enterprise services. Codex CLI just launched; GPT-5 series reasoning capabilities are surging. Sam Altman’s calculation is clear: AI products that make money are productivity tools, not toys.

Someone might ask: What about Google’s Veo 3.1? Kuaishou’s Kling? Aren’t these AI video tools doing well?

Doing well doesn’t mean making money. Kling’s 7.8 million monthly active users is impressive, but MAU and profitability are two different things. Veo 3.1 Lite’s slogan of “costs halved” precisely shows cost is this track’s biggest pain point.

From my project experience, currently AI video has only two scenarios with clear ROI: first, batch generation of e-commerce product display videos (but low quality requirements); second, short video platform special effects filters (essentially image processing, not video generation). Other scenarios? Still burning cash to explore.

But I don’t think AI video generation is a “pseudo-demand.” It’s more like a “timing not yet right” track — wait for compute costs to drop another order of magnitude, wait for model controllability to rise another level, and this market will explode again. However, the frontrunners may not be the current players.

The Sora story taught all AI entrepreneurs a lesson: technical moats don’t support business models.

Your model can be awesome, but if every user you serve loses money, that’s not a product — it’s charity.