Sora video burns thousands in GPU bills

Video generation is the expensive corner of generative AI — and Sora just became the clearest example of why. OpenAI spent 2024 and 2025 turning Sora into a flagship product, with 1080p clips, editing tools, remixing, and eventually synced audio. But by late April 2026, the company had discontinued the Sora app, and its developer docs now say the Sora 2 video API is deprecated and scheduled to shut down on September 24, 2026. (openai.com) ### What actually changed? The big change is not a new launch. It is a pullback. OpenAI’s September 30, 2025 post introduced Sora 2 as a more realistic and controllable video-and-audio model, but that same page now carries a note saying the Sora product was no longer available as of April 26, 2026. The help page for users says Sora has been discontinued and warns creators to export content as soo(openai.com)ow is offered. (openai.com) ### Why does that point to a cost problem? Because high-quality video generation is one of the most compute-hungry things these models do. OpenAI’s own technical material describes Sora as a system that generates video by denoising over many steps and operating across spacetime patches of image and video latents. Basically, every upgrade people want — longer clips, higher resolution, better motion, syn(openai.com)up fast. OpenAI has not published a neat “this video costs X dollars in GPUs” breakdown, but the architecture itself tells you why the bills get ugly. (openai.com) ### Didn’t OpenAI try to productize it? Yes — and that is part of the story. When OpenAI first rolled Sora out to consumers in December 2024, access was tied to ChatGPT Plus and Pro subscriptions. Later, the standalone app pitched 1080p, 20-second clips, multiple aspect ratios, and editing workflows like storyboard, remix, and blend. That sounds like a mainstream creative tool. The catch is t(openai.com)osts do not fit together cleanly. (techcrunch.com) ### What do the current docs say? The current API docs are blunt: the Videos API and several Sora 2 model variants are deprecated, with a shutdown date of September 24, 2026. The model page still lists Sora 2 as a flagship video model with synced audio, but the broader product picture is now clearly transitional rather than expansionary. In plain English — OpenAI is not acting like a company trying to flood the market with cheap unlimited video. (developers.openai.com) ### Is this just a safety shutdown? Safety is part of it, but not the whole thing. OpenAI’s launch materials emphasized visible watermarks, C2PA metadata, reverse-search tracing, and consent-based likeness controls. Those systems add operational overhead, and video is already expensive before moderation and provenance tooling enter the picture. So the problem is(developers.openai.com)too. (openai.com) ### Why does this matter beyond OpenAI? Because Sora was supposed to be the proof that text-to-video could become a normal consumer product. Instead, it looks more like a warning that the demo curve and the business curve are diverging. AI video can impress people instantly, but sustained consumer access needs cheaper inference, tighter limits, lighter models, or much higher prices (openai.com)check moment for AI video more broadly — and that reads right. (techcrunch.com) ### So what is the real lesson? The real lesson is that video is where generative AI stops feeling like software and starts feeling like infrastructure. A chatbot can serve millions of users on tolerable economics. A high-end video model can burn through compute, safety review, storage, and product support much faster. Sora did not fail because people stopped want(techcrunch.com)d reliable enough to run as a real product at scale.