OpenAI's Sora decision is all about AI compute - Business Insider

Five months before OpenAI shut down Sora, the head of the project posted publicly that the economics were broken.

Bill Peebles, who led Sora at OpenAI, wrote on Oct. 31, 2025: "we have been quite amazed by how much our power users want to use sora, and the economics are currently completely unsustainable." He also warned that the team would eventually need to reduce free generations of video "because we won't have enough gpus to do it otherwise." He followed that with a deadpan addendum: "video models really are expensive."

On March 24, 2026, OpenAI confirmed it was discontinuing Sora in its consumer app and API, shifting the research team toward robotics. The official statement — delivered to CNET by an OpenAI spokesperson — made no explicit mention of compute constraints, instead framing the decision as a pivot toward "world simulation research to advance robotics that will help people solve real-world, physical tasks." The New York Times reported that OpenAI gave no reason for the shutdown.

But the gap between what OpenAI said and what its own employee said five months earlier tells the story. Even with more than $20 billion in annualized revenue — confirmed by CFO Sarah Friar in January 2026 reporting — and hundreds of billions in committed chip and data center spending, OpenAI apparently could not keep Sora running alongside frontier model training. Business Insider was first to report the compute-crunch framing, describing how GPU scarcity was forcing tradeoffs across the company. Bernard Golden, CEO of the tech analysis firm Navica, put it plainly to Business Insider: "Given the frantic search for more compute across the industry, OpenAI is prioritizing its greatest growth engine — ChatGPT."

The shutdown of a flagship consumer product that launched in September 2025, briefly topped the Apple App Store charts, and had secured a landmark deal with Disney — which invested $1 billion in OpenAI and licensed more than 200 characters from its studios — raises a straightforward question: if Sora could not survive at a company burning through tens of billions annually, what does that say about the compute economics of any AI company at the frontier?

The answer, according to multiple supply-chain data points, is that the constraints are physical. Lead times for data-center GPUs from distributors BCD and Sourceability stretch to 36-52 weeks, according to an analysis by the AI platform Clarifai. Token consumption across AI systems has grown roughly 50-fold in a few years per that same analysis. AI token usage on the model-routing platform OpenRouter more than tripled in 2.5 months, Business Insider reported, suggesting demand is still accelerating faster than supply.

Sora's trajectory illustrates the bind. The product launched in September 2025 and quickly hit the top spot on the Apple U.S. App Store, according to PCMag. That same month, Peebles was already telling users the free tier could not be sustained. He introduced a paid option for power users — a step toward what he described as a "new sora economy" with rightsholder monetization — while explicitly acknowledging the GPU ceiling was real.

Disney, which partnered with OpenAI in December 2025 in a deal that included a $1 billion equity investment and three-year character licensing agreement, offered a diplomatic response. "As the nascent AI field advances rapidly, we respect OpenAI's decision to exit the video generation business and to shift its priorities elsewhere," a Disney spokesperson told CNET. The phrasing is notable — Disney characterized it as an exit from video generation, even as OpenAI said the underlying technology would continue in robotics research.

The strategic reorientation has a name inside OpenAI: no side quests. Fidji Simo, hired as CEO of Applications in May 2025 — announced by Sam Altman in a blog post that week, according to the New York Times and CNBC — told employees at an all-hands meeting in March 2026 that the company needed to stop being distracted by peripheral projects and focus on high-productivity use cases, Business Insider reported in a March 21 profile. Simo, who spent years running Instacart and previously held product leadership roles at Meta, oversees roughly two-thirds of OpenAI. Her mandate, per that profile: make OpenAI's products profitable before an eventual IPO. Deutsche Bank estimated OpenAI would accumulate $143 billion in losses between 2024 and 2029; OpenAI disputes that figure, pointing to its own projections showing cash burn of $111 billion by 2030 per The Information.

Sora was a side quest by that definition — compute-intensive, not yet monetized at scale, and dependent on the same GPU infrastructure that frontier training requires. Bernard Golden's framing suggests the math was straightforward: if compute is scarce and ChatGPT is the proven revenue driver, compute goes to ChatGPT.

What's striking is that even hundreds of billions in committed chip spending — OpenAI has locked in supply deals with Nvidia, AMD, and custom silicon vendors — was not enough buffer to keep Sora alive as a consumer product while simultaneously training the next generation of frontier models. That contrast — between a company making bold claims about frontier progress and one that could not keep a flagship consumer product alive — is the part worth sitting with.

That is the real window this story opens: not a product obituary, but a view into the operational reality of frontier AI labs. Training runs at the frontier consume enormous, unpredictable amounts of GPU time. Consumer products — even ones generating billions of video frames — compete for the same pool. When a company at $20 billion-plus in annualized revenue cannot sustain a flagship consumer product, the implication is not that Sora failed. It is that the frontier itself is extraordinarily expensive to maintain, and the tradeoffs are only going to get sharper as token consumption continues to climb.