AI video hits a unit‑economics wall: OpenAI shutters Sora as ByteDance product pulls discounts

Two shocks in 24 hours

OpenAI quietly pulled the plug on Sora — removing the app, shutting the API and stripping video generation from ChatGPT — it has been reported that the move also terminated a previously reported $1 billion IP pact with Disney. Less than a day later, JiMeng (即梦), the ByteDance (字节跳动)‑affiliated AI video product behind Seedance 2.0, cancelled its deep membership discounts, leaving many creators facing roughly 2.5× higher effective costs. Coincidence? Possibly. But the twin moves expose the same pressure point: inference compute is expensive and retention is low.

The numbers that hurt

It has been reported that Sora burned roughly $15 million a day — an annualized run rate north of $5.4 billion — while generating only a sliver of revenue and suffering very low 30‑day retention, according to industry analysis cited in reports. Reportedly, JiMeng’s discounting scheme functioned as a paid fast‑lane: the subsidised first months masked per‑inference losses. Analysts estimate a single 5‑second 720p clip costs about ¥2.3 to infer, and discounted annual subscriptions left the platform effectively losing money on high‑usage users. Users are angry. Creators feel “stabbed,” but companies argue they must prioritise compute for next‑generation models and profitable growth ahead of IPO windows — it has been reported that internal resource reallocation at OpenAI was explicitly to free capacity for a codename "Spud" model.

A broader industry truth

This is not just about one app or one product. Kuaishou (快手)’s Keling AI (可灵), Alibaba (阿里巴巴)’s Wanxiang (万相), MiniMax’s Hailuo (海螺) and Shengshu Technology (生数科技)’s Vidu are all iterating — and Google (谷歌)’s Veo 3.1 even supports 4K — but it has been reported that the combined ARR of major AI‑video vendors remains under $1 billion, a fraction of text‑model revenues at big players. Meanwhile cloud and chip economics are tightening: it has been reported that Amazon raised EC2 ML capacity prices and Google Cloud also slated price changes, while geopolitical factors and export controls continue to constrain high‑end GPU supply. If GPUs and power costs don’t fall dramatically, the subsidy era for mass‑market AI video is over. Free lunch ended. Competition remains, but the real variable is hardware — not marketing.