OpenAI Killed Sora: $15M/Day Burn Rate, $2.1M Revenue, Disney's $1B Deal Gone

OpenAI shut down Sora on March 24, 2026. Not pivoted. Not paused. Shut down — the app, the developer API, and sora.com, all of it gone simultaneously.

The economics behind the decision are the story. Sora was burning through an estimated $15 million per day in inference costs at peak usage. Its entire lifetime revenue from in-app purchases: $2.1 million. Bill Peebles, OpenAI's own head of Sora, described the situation publicly as "completely unsustainable." That is an unusual admission from someone inside the company that built the product.

For developers who had started integrating the Sora API or building workflows around AI video generation, this is an abrupt cliff. For everyone watching the AI industry, it is a specific data point about where generative video economics actually stand in 2026.

What Exactly Was Shut Down

The shutdown was total, not partial. The Sora iOS app stopped accepting new sessions. The Sora API — which developers had been using to generate video programmatically — went offline. The sora.com domain redirected to an OpenAI landing page explaining the discontinuation.

No migration path was offered. No deprecation warning period. Developers with active integrations had approximately 24 hours notice before the API stopped responding.

OpenAI cited two reasons officially: compute reallocation to higher-priority projects, and a shift toward using the underlying video generation research for robotics world simulation rather than consumer video creation. The robotics framing is plausible — world simulation for training robot models has substantial strategic value. The compute reallocation framing is more honest about the core issue.

The Economics That Made This Inevitable

Generating video at quality levels competitive with Runway or Kling requires enormous compute per output second. The physics are unfavorable: a language model generates tokens sequentially, but a diffusion video model generates frames iteratively across temporal and spatial dimensions simultaneously. The compute per second of output video is orders of magnitude higher than compute per token of text.

At Sora's pricing, OpenAI was subsidising almost every generation. The $15 million per day figure at peak represents a structural loss, not a scaling inefficiency that gets fixed with volume. More users would have meant more losses proportionally.

The $2.1 million lifetime revenue figure is the harder number to process. Sora launched with significant press coverage and genuine excitement. The gap between awareness and monetisation — millions of people curious about AI video versus a small fraction willing to pay for it — is the same gap that has killed consumer AI products before.

The Disney Deal That Never Closed

In December 2025, Disney announced a $1 billion investment in OpenAI paired with a content licensing deal that would allow Sora users to generate video using Disney's copyrighted characters and worlds. The announcement was treated as a validation of Sora's commercial viability — Disney does not commit $1 billion to infrastructure bets casually.

The transaction never closed. The shutdown on March 24 came before the deal completed. Disney lost its primary use case for the investment, and OpenAI lost the capital injection and the content moat that would have differentiated Sora from competitors.

The collapse of that deal is the clearest single indicator of how quickly the economics shifted. A $1 billion strategic partnership was in negotiation while the product was burning $15 million a day and generating under $3 million in total revenue.

What Developers Who Built on Sora Should Do Now

If you had production integrations on the Sora API, the immediate options are the three platforms that have absorbed most of the displaced demand:

Runway Gen-4 is the current quality benchmark for professional AI video. Priced at approximately $0.05 per second of generated video, it is expensive but predictable — Runway has been profitable-unit-economics focused since 2024 and is not subsidising output. API access is available with documented rate limits.

Kling 2.0 (Kuaishou) is the best price-to-quality option for most developer use cases. Chinese-built, Western API access available, generates 5-second clips at roughly $0.02 per second. Latency is higher than Runway but acceptable for async workflows.

Google Veo 3 is in limited preview as of March 2026. Access is restricted but Google's infrastructure economics are different from OpenAI's — they can subsidise inference at scale in ways OpenAI cannot. If Veo 3 opens broad API access, it will reset market pricing.

Pika Labs remains the best option for rapid prototyping — generous free tier, simple API, shorter clips. Not production-grade for professional output but useful for development and testing.

What This Tells You About AI Product Economics

Sora is not the first AI product to collapse under inference costs and it will not be the last. The pattern is identifiable: a generative capability that costs orders of magnitude more to produce than users will pay for, subsidised through investor capital during a growth phase, discontinued when the capital cost becomes indefensible.

The models where this pattern does not apply share a characteristic: the output is tokens, not pixels or frames. Text generation has inference economics that are improving fast — smaller models, better quantisation, faster hardware. Image and video generation inference economics are improving more slowly because the fundamental compute requirements scale with output resolution and temporal length in ways that are hard to compress.

Developers building on AI video APIs in 2026 should treat any API from a company not demonstrably profitable on video inference as a medium-term risk. Runway and Kling are the two platforms where the economics are most defensible. Everything else, including the announced products from companies that have not publicly addressed unit economics, carries concentration risk.

OpenAI's Actual Strategic Direction

The robotics framing in OpenAI's shutdown announcement is worth taking seriously even if it reads partly as face-saving. OpenAI's deal with Figure Robotics and its internal work on embodied AI both require world simulation — generating synthetic environments that robots can train in. The same diffusion video models that power Sora are the technical foundation for world simulation.

Redeploying that compute from consumer video generation to robotics world simulation is a strategic bet that robotics training data is worth more than consumer video subscriptions. Given where Figure, 1X, Boston Dynamics, and Tesla Optimus are in their training pipelines, that bet is not unreasonable.

It does mean OpenAI has effectively ceded the consumer and developer AI video market to Runway, Kling, and whoever Google decides to be with Veo 3. For developers, that market is consolidating faster than the text AI market did.

Key Takeaways

• Sora shut down March 24, 2026 — app, API, and sora.com simultaneously, with approximately 24 hours developer notice
• The economics: $15M/day inference costs at peak against $2.1M lifetime revenue — Bill Peebles described it as "completely unsustainable"
• Disney's $1B deal collapsed with the shutdown — the strategic partnership that was supposed to give Sora a content moat never closed
• Replacement API options: Runway Gen-4 ($0.05/sec, quality benchmark), Kling 2.0 ($0.02/sec, best price-to-quality), Google Veo 3 (limited preview), Pika Labs (prototyping)
• The structural issue: video inference costs scale with resolution and temporal length in ways that text inference does not — the economics improve slowly
• OpenAI's real bet: redirecting the compute to robotics world simulation, where synthetic training environments may be worth more than consumer video subscriptions
• Risk signal for developers: any AI video API from a company that has not publicly addressed unit economics carries medium-term discontinuation risk