OpenAI Kills Sora: The IPO Clock, Compute Scarcity, and Why the Coding Race Won
OpenAI killed Sora on Tuesday. The announcement landed like a product decision. It was an economic one.
Six months after launch, Sora had 1.1 million monthly active users — down from a November 2025 peak of 3.3 million downloads across iOS and Android, according to Appfigures data cited by Wired. The app never had a revenue model. Video generation is among the most compute-intensive tasks in AI. Bill Peebles, Sora's head, said it plainly in October: the economics were "completely unsustainable." On Tuesday, OpenAI agreed.
The company confirmed it would discontinue Sora's consumer app and API, redirecting the GPUs and researchers to "world simulation research" for robotics. The Sora research team, as a distinct unit, is done. What they learned building a video model that impressed the world and failed to find a market will now be applied somewhere else.
The Disney fallout was immediate. Disney had pledged $1 billion to OpenAI in December 2025, a deal that included licensing beloved characters for AI-generated content. The company had no warning before Tuesday's announcement, per the LA Times. Disney said it no longer plans to invest. No money had changed hands — the deal was still being formalized — but the relationship is over.
This is what an IPO looks like from the inside.
OpenAI's CFO Sarah Friar told CNBC the company needs to be "ready to be a public company." Fidji Simo, OpenAI's CEO of Applications — a role modeled, one source noted, on her background taking Instacart public — told employees at an all-hands meeting: "We cannot miss this moment because we are distracted by side quests." Sora was a side quest. So, apparently, was whatever the company was building in consumer hardware, the browser project, and the broad agent ambitions that never gained traction.
The focus areas are now explicit: a "super app" combining ChatGPT, Codex, and Atlas into a unified consumer interface, and enterprise coding tools where the revenue actually is.
Here is where the Anthropic framing cuts deepest. Anthropic's Claude Code reached $2.5 billion in annualized revenue by February 2026 — roughly a fifth of Anthropic's total business, per the company's own figures. OpenAI's Codex crossed $1 billion in annualized revenue in January, per a person with direct knowledge. The gap matters. Anthropic had a dedicated coding product eight months before OpenAI did. By the time Codex existed as a standalone product, Cursor — founded by a group of twentysomethings who declined an acquisition offer from OpenAI — was already the default tool for developers who wanted AI-native coding. OpenAI had to catch up in a market Anthropic was already dominating.
This was not inevitable. OpenAI demoed Codex in 2021. GitHub Copilot, powered by OpenAI's models, launched in June 2022 and had hundreds of thousands of users within months. But after ChatGPT hit 100 million users in two months, the company's focus shifted entirely to consumer products. The Codex team disbanded. Engineers moved to image generation and GPT-4. OpenAI spent 2023 and 2024 building multimodal agents while Anthropic quietly trained on messy code repositories and shipped Claude Code.
Greg Brockman, OpenAI's president, acknowledged the gap on a recent podcast: Anthropic was "focused very hard on coding from an early stage," he said, and trained on real-world code — not just academic benchmarks. "That was a lesson we were delayed on." The delay cost something measurable: Claude Code's $2.5 billion annualized revenue run rate versus Codex's $1 billion.
The compute crunch accelerated the reckoning. Data from OpenRouter shows AI usage more than tripled in 2.5 months as inference demand surged across the industry. Building new data centers has run into local opposition, energy constraints, and memory chip shortages. OpenAI has committed hundreds of billions to secured capacity and still doesn't have enough. Every GPU allocation is now a strategic decision, not a research decision.
That dynamic is not unique to OpenAI — it's playing out across the industry. But the IPO timeline makes it acute. OpenAI is losing billions annually. A company preparing to go public cannot sustain projects that burn compute without producing revenue. Sora burned a lot of both.
There is an open question about what this means for OpenAI's research culture. The company has run with a "bottom-up" resource allocation model — researchers pursued promising ideas and competed for compute internally. In January, OpenAI's VP of Research, Jerry Tworek, left after struggling to get resources for his next project. Other researchers are reportedly energized by the focus shift. Whether a company built for open-ended exploration can execute a disciplined product roadmap is a different question from whether the economics of Sora made sense. Both are real.
The strongest signal in this story is the revenue data: $2.5 billion versus $1 billion, Anthropic versus OpenAI, in annualized coding-agent revenue. That gap is the reason Sora died. Not because it failed technically, but because compute that could have sustained it was redeployed to a race OpenAI is currently losing by $1.5 billion a year.
The IPO math is simple. Focus is not a strategy — it's what happens when you're outrun in the race you actually care about and can't afford the side quests anymore.
Sources:
Wired: OpenAI Enters Its Focus Era by Killing Sora — https://www.wired.com/story/openai-shuts-down-sora-ipo-ai-superapp/
Wired: Inside OpenAI's Race to Catch Up to Claude Code — https://www.wired.com/story/openai-codex-race-claude-code/
Business Insider: OpenAI kills Sora app as AI compute crunch forces hard choices — https://www.businessinsider.com/openai-kills-sora-app-ai-compute-crunch-forces-hard-choices-2026-3
LA Times: OpenAI will shut down Sora — https://www.latimes.com/entertainment-arts/business/story/2026-03-24/openai-will-shut-down-sora-why-what-to-know