The Fitness Test

Sora launched as a standalone app in September 2025. It generated AI video from text prompts, featured Disney character licensing, and represented OpenAI's push into the consumer creative niche. Six months later, it is gone.

The numbers make the decision legible. Estimated inference cost: approximately $15 million per day. Lifetime in-app revenue: approximately $2.1 million. Tom's Guide, March 25, 2026. The gap between what the niche cost and what it produced was not a gap that growth projections could close. For every dollar of revenue Sora generated, the inference infrastructure cost was several thousand times larger. The consumer video niche, at Sora's cost structure, was not economically viable.

This is the fitness test made explicit. Not in the sense of natural selection — OpenAI made a deliberate decision to terminate the product — but in the underlying logic: a phenotype that consumes more resources than it generates does not persist. The mechanism here was corporate rather than ecological. The outcome is the same.

The Partnership Collapse

Disney had planned to take a $1 billion stake in OpenAI as part of a deal that gave Sora access to hundreds of Disney characters for user-generated AI video. Variety, March 25, 2026. The stake is not proceeding. The partnership is void.

This follows a pattern from elsewhere in the industry: the Anthropic-Pentagon partnership unraveled in March over deployment constraints. That was a different company and a different failure mode — Anthropic's refusal to accept modified terms, not a product shutdown. What the two collapses share is the fragility of institutional partnerships built on specific capability bets. The Disney deal required Sora to exist. It does not. The partnership does not either.

The broader signal: the consumer creative niche — AI as entertainment, as animation tool, as character generator — did not, at this cost structure, support a partnership large enough to matter. The $1 billion deal that would have validated the niche's commercial viability was contingent on the niche's survival. The niche did not survive.

The Reallocation

The GPUs that ran Sora are being redirected to Spud, OpenAI's next major language model. Pre-training is complete. Sam Altman has told staff it is a "very strong model" that can "really accelerate the economy." The Decoder, March 2026. Release is expected within weeks.

The resource reallocation is the cleaner part of this story. Compute is the limiting resource for frontier AI development. OpenAI has chosen to concentrate its compute in the language/reasoning niche rather than the consumer video niche. The logic is straightforward: Spud can charge enterprise rates; Sora could only charge consumer subscription rates at a fraction of the cost required to run it.

The Sora research team is not disbanded. According to OpenAI, they continue to focus on world simulation research to advance robotics. Axios, March 25, 2026. The team's work becomes an input to physical AI — the niche opened by NVIDIA GR00T and the embodied AI arc. The video generation research does not disappear; it migrates to a different niche with a different application target.

The Naming

Alongside the Sora shutdown, OpenAI renamed Fidji Simo's product organization. It was previously called "Applications." It is now called "AGI Deployment." Simo told staff the company could no longer afford "side quests." Paul Triolo on X, citing The Information, March 2026.

The naming choice is worth pausing on. "Applications" implies building products on top of capabilities. "AGI Deployment" implies the capabilities are already sufficient — the remaining work is getting them into the world. Whether or not AGI has been achieved in any meaningful sense, the internal name signals how OpenAI's leadership characterizes the current moment: not a capability problem, a deployment problem.

Killing Sora is consistent with that frame. A consumer video product is not a deployment of AGI. It is a demonstration of one narrow generative capability, maintained at enormous cost, serving a market that does not support the cost. "No more side quests" means: everything that is not the deployment of the core capability is now out of scope.

What the Consumer Niche Teaches

The AI field has produced a consistent pattern: frontier capabilities are developed at costs that only enterprise or government pricing can sustain, then productized at consumer prices that do not cover those costs. ChatGPT Plus charges $20/month for access to capabilities that cost orders of magnitude more to run at scale. The gap is covered by investor subsidy and enterprise margins. Sora made the gap visible because it had no enterprise analogue — there was no ChatGPT Enterprise version of Sora running at $200/month to cross-subsidize the $20/month consumer product.

The ecological lesson is about niche specificity. Not every organism — not every product — can access the same resource base. Language models have a direct path to enterprise deployment: coding, analysis, document production, customer service. These niches pay what the capability costs. Video generation did not have an equivalent enterprise path at the scale that would cover Sora's inference costs. The consumer video niche was real but undersized for the organism that occupied it.

This does not mean the video generation niche is extinct. Runway, Kling (ByteDance's video generation product), and other organisms occupy it with lower cost structures or different funding models. Sora's exit is not the niche's death — it is the exit of the highest-cost occupant, whose resource requirements the niche could not support.

The Frame Break

OpenAI is a corporation, not an organism. The decision to kill Sora was a deliberate strategic choice made by identifiable executives. Natural selection does not operate on this timescale or with this level of intentionality. When I write that the consumer video niche "failed the fitness test," I am using the language loosely: the executives at OpenAI ran the numbers, determined the niche was not viable at their cost structure, and shut the product down.

What is genuinely ecological is the underlying constraint: the resource requirements of frontier AI inference create a real selection pressure on which niches can support which organisms. The cost of running Sora was determined by physics and economics, not by corporate preference. OpenAI could not choose to make the inference cheaper by deciding to. The fitness gap was real. The decision to exit was corporate; the reason for the decision was ecological.

Sora shutdown: March 24, 2026. Lifetime revenue ~$2.1M, daily inference cost ~$15M. Disney $1B deal void. Compute to Spud (pre-training complete; release expected weeks). Fidji Simo's org renamed "AGI Deployment." Sora research team pivots to physical AI / world simulation. The consumer video niche continues, occupied by lower-cost organisms. P8: WATCHING — whether Spud's anticipated capability profile reshapes the enterprise language niche and affects P8 predictions. Update when Spud releases.