When OpenAI launched Sora in October 2025, it promised an endless feed of AI‑generated 10‑second videos you know are fake — and, perversely, might prefer that way. The app rocketed to the top of Apple’s charts within days and left technologists asking a blunt question: can anyone afford to run a platform that streams synthetic video by the billion?

Sora matters because it collides three forces at once: the exponential compute demand of video generation, the unresolved legal landscape for copyrighted and deepfaked material, and a business model that could either monetize a new attention economy or bankrupt its creator. OpenAI’s gamble is not just technical; it is economic and regulatory. How the company answers questions about cost, carbon, and copyright will shape the next phase of consumer AI.

How Sora generates a stream of short, hyperreal clips

The stakes extend beyond OpenAI. If Sora proves that endless synthetic video can attract sustained attention, it will change advertising, content moderation, and creator rights. If it fails, it will still illuminate the real cost of scaling generative video and the hard limits of current models — data, electricity, supply chains, and lawsuits included.

How Sora generates a stream of short, hyperreal clips

The math of scale: compute, cost, and carbon

Sora produces up to 10‑second clips in the style of TikTok with prompts, user cameos, and preset characters. Behind the polish are stacked models: a text‑to‑video generator that converts prompts into motion, a multimodal embedder that tracks continuity across frames, and a personalization layer that maps a user’s cameo — a photorealistic avatar of voice and face — onto generated scenes. OpenAI’s public comments confirm the app leans on high‑capacity video models that combine framewise diffusion with temporal transformers to preserve motion and audio sync.

The technical trade‑offs are immediate. Short videos allow Sora to reuse computations — generating a 10‑second clip at 24 frames per second is vastly cheaper than naive frame‑by‑frame synthesis, but still orders of magnitude costlier than text or single‑image generation. Research such as Google’s Imagen Video and related diffusion‑based work demonstrates common strategies: operate in a latent space, predict compact video tokens, and refine frames with a denoiser that enforces temporal coherence. These architectures explain how Sora can make a feed that feels continuous without generating each pixel from scratch every time.

The math of scale: compute, cost, and carbon

Copyright, deepfakes, and the coming courtroom fights

Generating video is not a line item — it’s a multiplier. Text models like GPT produce a single token stream; video requires tens to hundreds of tokens per frame across dozens of frames. OpenAI’s Sam Altman acknowledged the economics bluntly: “we are going to have to somehow make money for video generation.” Even with model distillation and caching, the marginal GPU seconds per user session balloon relative to text chat.

That matters because compute maps directly to dollars and emissions. Studies of NLP training energy — for example Strubell et al. 2019 — showed that large‑model training can emit hundreds of tons of CO2; inference at scale multiplies this footprint. While those numbers were for text and training, video inference can use 5–50× more FLOPs per user query depending on architecture and optimization. If Sora sustains tens of millions of sessions per day, the company faces kilowatt‑hour bills and carbon accounting questions that matter to investors and regulators alike.

Who wins and who loses: creators, incumbents, and attention markets

OpenAI has been building data centers and power contracts, but those are long‑lead investments. The immediate levers are throttles: generation limits, subscription paywalls, or aggressive compression. Each breaks a promise of "limitless" user experience in different ways. The company’s choices here will determine whether Sora is an ad‑funded feed, a paid creative tool, or a costly experiment.

Copyright, deepfakes, and the coming courtroom fights

Sora’s permissive creative space has already run headlong into copyright and likeness law. Users quickly churned videos that repurpose trademarked characters, music, and the voices of deceased celebrities. OpenAI’s approach — notifying rightsholders they must opt out rather than opt in — inverts normal licensing practice and has triggered warnings from rights organizations. That approach raises exposure: rights owners and music publishers have long histories of litigation against platforms that host infringing content.

Sources

• MIT Technology Review — The three big unanswered questions about Sora (2025-10-07)

• MIT Technology Review — The Download: carbon removal factories’ funding cuts, and AI toys (2025-10-08)

• ACL Anthology — Energy and Policy Considerations for Deep Learning in NLP (Strubell et al., 2019) (2019-11-01)

• arXiv / Google Research — Imagen Video: High Definition Video Generation with Diffusion Models (2022-05-24)