The Math on Sora Is Brutal: A $20/Month Subscriber Costs OpenAI $65 in Compute

OpenAI's Sora, the video generation model that drew widespread attention at launch for its cinematic quality, is also drawing attention for a less glamorous reason: the unit economics don't work. A detailed breakdown circulating on Hacker News this week calculates that a $20/month ChatGPT subscriber who uses Sora meaningfully costs OpenAI approximately $65 in GPU compute — a loss ratio of more than 3-to-1.

The calculation is back-of-envelope but grounded in publicly available data on GPU costs, inference throughput benchmarks for diffusion video models, and OpenAI's stated compute partnerships. It is not a precise audit, but the order of magnitude is defensible enough that it has attracted serious engagement from AI infrastructure engineers.

Why Video Is Fundamentally Different

Text generation at scale is expensive, but it has a saving grace: autoregressive inference is relatively compute-efficient per token, and the market has driven GPU costs down substantially as demand has grown. Video generation operates on different physics.

Generating a 10-second, 1080p video clip with a diffusion-based model like Sora requires iterative denoising across a high-dimensional spatiotemporal latent space. Each denoising step involves dense attention computations across the full video tensor. A single high-quality generation can require hundreds of GPU-seconds on H100-class hardware. At cloud spot rates for H100s, that adds up quickly.

The numbers are stark: even at aggressively optimized inference, a moderately active Sora user who generates 20-30 video clips per month — a conservative estimate for anyone using the tool meaningfully for creative work — can consume $40-70 in raw compute. The subscription that funds this costs $20, or $0 for users on the Plus plan who access Sora as an included feature.

The OpenAI Response

OpenAI has not publicly commented on the specific numbers. The company has acknowledged that AI video generation is "resource-intensive" and has implemented quality-tier and usage throttles that limit how much Sora access subscribers receive at the highest quality settings. These constraints are, effectively, an implicit acknowledgment that unconstrained Sora access is not economically viable at subscription prices.

The longer-term bet is that hardware cost curves and model efficiency improvements will close the gap. Diffusion model inference has historically improved in efficiency faster than initial projections, and next-generation hardware — NVIDIA's Blackwell architecture, custom Google TPUs — is expected to deliver meaningful improvements in video inference cost.

The Strategic Problem

The economic pressure on AI video creates a strategic bind. Raising prices risks losing subscribers to competitors; maintaining current prices means subsidizing usage at scale. The middle path — throttling quality and quantity — degrades the product experience that justified the subscription in the first place.

Runway, Pika, and Kling — the specialized video AI players — face the same economics. None of them have achieved the kind of scale that would make current pricing profitable. The implicit assumption underlying all of their business models is that hardware costs will fall fast enough to close the gap before runway (financial, not the company) runs out.

For OpenAI specifically, Sora's economics are a microcosm of a broader challenge: the company is simultaneously the most visible AI lab in the world, the leader in consumer AI subscription revenue, and deeply unprofitable. The video model is the most extreme expression of a pattern that runs through the entire business: capabilities are advancing faster than the economics of serving them.