OpenAI Killed a $1 Billion Disney Deal to Free Up GPUs. The Math Says It Was a Bargain.

One thousand three hundred and four to one.

That is the ratio of cost to revenue for Sora, the video generation app OpenAI killed on March 24, 2026. At its peak, the service burned an estimated $15 million per day in inference costs, according to Cantor Fitzgerald analyst Deepak Mathivanan via Forbes. Over its roughly six-month lifespan, Sora generated $2.1 million in total in-app purchases. Divide one by the other and you get a product that spent $1,304 for every $1 it earned.

Sam Altman reportedly told employees there would be "no more side quests." He was not speaking in metaphor.

What Sora Cost, Line by Line

Mathivanan's breakdown of the per-video compute cost is instructive. Each 10-second Sora clip required approximately 40 minutes of GPU time spread across four parallel GPUs, at a rental-equivalent cost of roughly $2 per GPU-hour. That works out to $1.30 per 10-second clip in raw compute. At peak usage, with millions of users generating multiple clips daily, the aggregate hit $15 million per day, or $5.4 billion annualized.

Revenue told a different story. Sora launched in November 2025 to 3.33 million monthly downloads. By February 2026, downloads had collapsed 66% to 1.1 million. US App Store downloads fell 32% month-over-month in December, then another 45% in January. Lifetime in-app purchases totaled $2.1 million across the entire run.

Here is how that decomposes daily:

Sora daily revenue: $2,100,000 / 182 days = $11,538 per day.
Sora daily cost at peak: $15,000,000 per day.
Cost-to-revenue ratio: 1,304:1.

For every dollar Sora earned, OpenAI spent $1,304 in compute. The $15 million daily figure represents peak burn, not steady-state, so the actual lifetime ratio is likely lower. But even at a conservative $1 million per day (a figure cited by industry analysts), the cost-to-revenue ratio was still 87:1.

The GPU Triage Calculation

Nobody has published a cross-product GPU opportunity cost analysis for OpenAI. Everyone reports the Sora numbers in isolation. But the real question is not whether Sora lost money. It is how much money the same GPUs would have generated doing something else.

OpenAI's annualized revenue run rate surpassed $25 billion by early 2026, driven overwhelmingly by ChatGPT subscriptions and enterprise API access. At a daily run rate of roughly $68.5 million, OpenAI targets approximately 50% gross margins on its core language model products, implying roughly $34 million in daily gross profit from text and code inference.

Compare revenue per compute-dollar:

Sora: $11,538 revenue / $15,000,000 compute cost = $0.00077 per compute-dollar.
ChatGPT (blended): $68,500,000 revenue / $34,250,000 COGS (at 50% gross margin) = ~$2.00 per compute-dollar.

That is a 2,600:1 gap. Every GPU-hour spent on Sora generated roughly 2,600 times less revenue than the same GPU-hour serving the company's core language model products.

At the $15M/day peak, redirecting Sora's compute to enterprise workloads at the company's blended 2:1 revenue-to-COGS ratio would have generated approximately $30 million per day in additional revenue, or $10.9 billion annualized. Even using the conservative $1M/day Sora cost estimate, the opportunity cost was still $730 million in annual enterprise revenue. Both figures dwarf the $1 billion Disney deal OpenAI walked away from.

Disney Got 57 Minutes of Notice

Variety reported that Disney had negotiated a $1 billion, three-year licensing deal granting Sora access to over 200 Disney characters. OpenAI cancelled with less than 60 minutes' notice before the planned public announcement. From a relationship standpoint, this was scorched earth. From a GPU triage standpoint, it was rational.

A $1 billion deal over three years is $333 million annually. If that deal required dedicating a meaningful fraction of GPU capacity to Disney's video generation workloads at similar cost-per-clip economics, the net contribution margin would have been negative. Disney was effectively buying access to subsidized compute at below-cost rates, and OpenAI's finance team apparently realized it before the ink dried.

The cancellation makes sense only if you believe GPUs are the binding constraint and enterprise inference is the highest-value use. Both premises appear to be true.

The Industry Pattern

OpenAI is not the only lab making this trade. Anthropic has deliberately avoided shipping image or video generation, keeping its compute focused on text and code inference where margins are defensible. Google's Gemini 3.1 Flash-Lite launched at $0.25 per million input tokens, priced for enterprise batch processing. NVIDIA's GTC 2026 keynote centered on the Agent Toolkit for enterprise autonomous AI, not consumer creative tools.

OpenAI's Sora team was redirected to "world simulation research to advance robotics," according to the official statement. Translation: the underlying model still has value as a physics simulator for robotics training, where compute costs are a research expense rather than a consumer service obligation.

This pattern has a name in resource economics: triage. When capacity is finite and demand exceeds supply, you serve the highest-value use case first and kill everything below the threshold. For frontier AI labs in 2026, that threshold sits somewhere between enterprise API margins (50%+) and consumer media generation (negative 1,300%).

Strongest Counterargument

Video generation was never meant to be a standalone revenue product. Sora was a world-model research testbed that happened to ship as a consumer app. The real value was not $2.1 million in App Store purchases but the petabytes of physics-simulation training data now powering OpenAI's robotics pivot. Killing the consumer product while keeping the model available inside ChatGPT suggests the technology succeeded and the business model was wrong. The 1,304:1 ratio mischaracterizes a research program as a failed product launch.

This is a reasonable framing, and partially correct. But it does not explain the Disney deal. You don't negotiate a $1 billion three-year licensing arrangement for a research testbed. You don't hire a product team, build an app, launch on iOS, and market to consumers if the goal was always research. OpenAI tried to make Sora a product. The product failed. The research value survived, which is a lucky outcome, not the plan.

Limitations

Several caveats constrain this analysis. First, the $15 million daily inference cost is an analyst estimate, not an OpenAI disclosure. Actual costs depend on hardware configurations, batch optimization, and internal transfer pricing. The conservative $1M/day estimate from other analysts produces a materially different picture (87:1 vs. 1,304:1).

Second, the GPU opportunity cost calculation assumes fungible compute. In practice, video generation requires different hardware configurations (more VRAM, different parallelism patterns) than language model inference. Both workloads compete for the same H100/B200 clusters, but reallocation is not instantaneous or costless.

Third, OpenAI does not disclose enterprise revenue by product or GPU allocation by workload. The $2.00 revenue-per-compute-dollar estimate uses the company's blended financials (total revenue divided by estimated COGS at 50% gross margins), not product-specific data. If enterprise inference has higher margins than the company average, the gap widens further. If consumer subscriptions have lower margins, the blended figure overstates enterprise efficiency.

Fourth, the Disney deal terms come from press reporting. Neither company has confirmed the exact figures or timeline. The $1 billion figure may include non-compute components (licensing fees, content production) that would alter the margin analysis.

The Bottom Line

GPU hours are the scarcest resource in AI, and OpenAI just demonstrated what happens when a company starts treating them that way. Sora earned $0.00077 per compute-dollar. ChatGPT earns roughly $2.00. At those ratios, killing a billion-dollar Disney deal was not reckless. It was arithmetic. Every frontier lab is now running the same calculation against every product in its portfolio, and consumer video generation loses that math every time. If your favorite AI feature disappears this year, check whether it was competing with an enterprise API for the same rack of H100s. The answer will explain why.