The Sora Shutdown: AI Enters the "Brutal Reckoning" | Yunqi Tech π

On March 25, 2026, OpenAI officially announced the shutdown of Sora, the AI video generation tool that once captivated the world. Its iOS app, API, and ChatGPT-embedded features were all taken offline. From its dazzling debut in February 2024 to its abrupt exit, Sora's entire lifespan lasted just 25 short months. The night before the shutdown, Disney was still advancing a collaboration project with OpenAI, and the company had even posted guidelines on safe usage.

As a VC firm that has long tracked frontier technology and the AI ecosystem, we were stunned. But we also tried to extract a "pitfall avoidance guide" for AI entrepreneurship and investment from the premature death of this star product. Based on industry discussions from the past two days, this issue of "Yunqi Tech π" shares some takeaways with you.

Key Takeaways

• Facing its Q4 2026 IPO target, OpenAI had to demonstrate a clear path to profitability for capital markets. Against this backdrop, it's hardly surprising that a business burning $5 billion annually with no visible "money scene" was discarded. Technical leadership does not equal commercial success. Compute consumption without a commercial闭环 is more likely to drag a company down.
• Sora was caught in a fatal paradox — too expensive and uncontrollable as a consumer product, too crude and imprecise as a professional tool. AI startups need to truly切入 and deeply cultivate actual industrial production workflows.
• Even with abundant resources, major tech companies will actively abandon star projects. Companies must know how to subtract, betting core resources on the most deterministic scenarios that can form revenue闭环. This battle is shifting from the old moat (dazzling models) to the new moat (enterprise efficiency and commercial delivery).

This article is republished from WaveGlocal

Original title: Who "Killed" Sora? From Deification to Shutdown in 25 Months — OpenAI's "Strategic Abandonment" and the Crossroads of AI Video

A Star Crushed by Compute Costs

The most direct culprit in Sora's death is compute.

Video generation and text generation are completely different leagues when it comes to compute consumption. According to estimates by Cantor Fitzgerald analyst Deepak Mathivanan, OpenAI's cost to generate a 10-second video runs about $1.30. That may seem modest, but multiplied across massive user requests, the numbers become staggering.

Analysis firm SemiAnalysis estimates that Sora's daily operating costs hovered around $15 million, with an annual burn rate of $5.4 billion. This includes GPU rentals, electricity, inference costs, and other expenses. In other words, OpenAI was burning $15 million per day on Sora — and that doesn't even include R&D team salaries.

More fatally, Sora's model performance was nowhere near worth its burn rate. Multiple review bloggers noted that only 5% to 10% of videos generated by Sora reached publishable quality. Users had to "gacha" their way through repeated attempts, spending half an hour just to land a few usable seconds of video. This experience was torture for free users, a dealbreaker for paying users, and a double whammy for OpenAI — bearing high compute costs while failing to achieve satisfactory user conversion.

To control costs, OpenAI had no choice but to slash free users' daily generation quota from 30 to 6. This cut struck directly at Sora's vital point. Already fragile user experience became even worse after the quota reduction.

And the compute cost dilemma was intimately tied to OpenAI's IPO timeline. The company plans to launch its initial public offering in Q4 2026, and before that must demonstrate a clear path to profitability for capital markets. With every dollar of compute now subject to investor scrutiny, a business burning $5 billion annually with no "money scene" in sight was destined to be abandoned.

The Commercial Dead End

What 1% retention means

If compute costs were Sora's "internal injury," then its dismal commercial performance was the "last straw" that broke it.

a16z partner Olivia Moore once posted a SensorTower monitoring screenshot on social media with shocking data: Sora's 30-day retention rate was just 1%, and its 60-day retention rate flatlined at zero. This means that out of 100 users who downloaded Sora, only one would still be using it after a month, and after two months virtually no one remained.

Why? Because Sora was more of a "one-time toy" than a "persistent tool."

What was the most eye-catching feature when Sora 2 launched? The ability to put yourself in a movie. Users could generate videos placing themselves in various IP universes using their own portraits. The feature was genuinely stunning, but most users, after trying it once, had no idea what else to do with it. Once the novelty wore off, retention naturally plummeted.

The more practical problem was that many users simply weren't willing to hand their portraits over to an AI app to process — there was a trust issue here as well. Without celebrity faces or IP加持, the content Sora could generate quickly hit a ceiling of "interesting but useless."

Hence, this "purely impulse-driven" product delivered miserable commercial conversion. According to media reports, Sora generated only about $2.1 million in in-app purchase revenue throughout its brief lifespan. Appfigures data shows Sora App user spending fell from a December 2025 peak of $540,000 to $367,000 in January 2026. Meanwhile, competitor Anthropic's Claude Code added $6 billion in new revenue in February 2026 alone.

OpenAI's Sora lead Bill Peebles publicly admitted that Sora 2's business model was "completely unsustainable." When a product's operating costs and revenue are completely out of proportion, its fate is sealed.

At this point someone might ask: if toC doesn't work, why not go toB? WaveGlocal's answer is that toB is even harder.

Because Sora's model was "input some text, pray it generates what you want." If the result wasn't right, users could only modify the prompt and try again — they couldn't微调 a specific局部 of the video with surgical precision like using a scalpel. This lack of "pixel-level control" black-box mechanism doomed Sora to remain at the entertainment level of social media, unable to deeply embed itself into modern industrial production workflows.

For B2B professional clients who are actually willing to pay for technology — advertisers, film studios, game developers — what they need has never been "blind box" style random video generation. The film and television industrial system operates under exacting standards: directors need characters to make specific expressions at specific frame counts, lighting designers need light sources hitting from specific angles. Sora couldn't do any of this. It could only produce something that "looks beautiful," but whether that result was controllable and predictable was entirely unknown.

This was Sora's fatal paradox: as a consumer product, it was too expensive and uncontrollable; as a professional tool, it was too crude and lacking in precision. It was stuck in an awkward middle ground, pleasing no one on either side.

The Strategic Pivot

Strategic abandonment before IPO

Currently, OpenAI stands at a critical crossroads. Pressure from competitors like Anthropic and Google, combined with its Q4 IPO target, is forcing OpenAI into a round of "strategic abandonment" — cutting unprofitable businesses and concentrating resources on the most deterministic directions.

The threat from Anthropic is particularly urgent. According to latest statistics from fintech company Ramp, a year ago only 1 in 25 enterprise clients used Anthropic; now that ratio has risen to nearly one-quarter. In new customer acquisition, Anthropic is capturing roughly 70% of first-time purchase share. And Anthropic's success formula is precisely focus — the company has yet to launch image or video generation products, instead concentrating its efforts on enterprise services and the programming market.

OpenAI's applications lead Fidji Simo said at an all-hands meeting: "We cannot let ourselves be distracted by 'side quests.' We must achieve excellence in productivity, especially enterprise productivity."

Against this backdrop of strategic focus, Sora's elimination was only a matter of time. It was an extremely compute-hungry product — the electricity and chip depreciation burned on each video generation could probably power dozens of ChatGPT responses. And if that compute were reallocated to Codex (OpenAI's AI programming assistant), the business logic would be seamless: Codex has tripled its user base this year, quintupled usage, and now has over 2 million weekly active users.

Meanwhile, OpenAI is consolidating its product lines. ChatGPT desktop, code development tool Codex, and browser are being merged into a "super app." The Sora team wasn't disbanded either — it was renamed "AGI Deployment" and will pivot to long-term world simulation for robotics technology.

For Whom the Bell Tolls

What Sora teaches us

Sora's exit doesn't mean the end of the AI video generation赛道. Quite the opposite — this赛道 is entering a new developmental stage, only the protagonist is no longer OpenAI.

Just as Sora made its quiet exit, Chinese players are rising powerfully. ByteDance's Seedance 2.0 and Kuaishou's Kling 3.0 have already far surpassed Sora in both generation speed and quality.

More importantly, these Chinese players have found sustainable business models. According to latest earnings data, as of December 2025, Kuaishou's Keling AI monthly revenue had surpassed $20 million, with annualized revenue soaring to $240 million, over 60 million global creators, and cumulative video generation exceeding 600 million clips.

Sora's exit precisely validates a simple commercial truth: technical leadership does not equal commercial success, and炫酷 demos do not equal sustainable products. In the post-AI era where capital patience has run out, every model must pass the "compute monetization" test to deserve a place in the future.

It marks the official end of the AI industry's "burn money for growth" expansion phase, and the arrival of a "compute精算" rational era. Future AI video technology will lean more toward lightweight, verticalized落地, focusing on细分 scenarios like film and television production, short video creation, and enterprise marketing, rather than pursuing the全面铺开 of general-purpose large models.

Goodbye, Sora. You once made us believe AI could create cinematic-grade video, but you also made us understand that in this commercial world, being stunning is far from enough.