OpenAI Kills Sora to Focus on Work AI, Ending an Era

OpenAI just did something I haven't seen in thirty years of covering Silicon Valley: they publicly killed a flagship product, not because it failed technically, but because they needed the electricity for something else.

The company announced this week that Sora, their AI video generation app launched with considerable fanfare last October, is being sunset. All products using their video models are being discontinued. The compute resources—which turned out to be far more substantial than expected—are being redirected to a new model codenamed Spud. In an internal memo, CEO Sam Altman told staff the company expects Spud will "really accelerate the economy."

That phrase is doing considerable work. It's either a genuine breakthrough or the kind of internal rallying cry that sounds better before it leaks. But the broader shift is unmistakable: OpenAI is abandoning consumer experiments to focus entirely on work automation.

The Compute Constraint Gets Real

The Wall Street Journal reported that some OpenAI staff were surprised by how much computing power Sora consumed relative to user demand. That's the polite version. The honest version is that running AI video generation at scale burns through GPUs faster than the business case can justify it.

This marks the first time OpenAI has had to make this kind of choice publicly. They've delayed models before, citing compute constraints. But shutting down an entire application they'd promoted months earlier is different. It's a tangible admission that even with billions in funding and Microsoft's infrastructure backing them, there simply isn't enough computing power to pursue every promising direction simultaneously.

Fiji Simo, OpenAI's CEO of applications, framed the decision as natural evolution: "Companies go through phases of exploration and phases of refocus. Both are critical, but when new bets start to work, like we're seeing now with Codex, it's very important to double down on them and avoid distractions."

The consolidation extends beyond Sora. Altman is stepping back from direct oversight of safety and security teams to focus on capital raising, supply chains, and data center buildout—the infrastructure of the AI race. The product division is being renamed "AGI deployment," a shift from building interesting things to deploying them at enterprise scale.

What This Means for Work

The bet OpenAI is making is straightforward: the market for automating knowledge work is larger and less contested than the market for consumer AI applications. Simon Smith, writing on X, put it plainly: why chase a $680 billion advertising market dominated by Google and Meta when there's a "roughly $40 trillion plus market of automatable knowledge work?"

The logic is sound. Anthropic has gained ground in enterprise and coding applications by maintaining exactly this focus while OpenAI experimented with video generation, shopping features, and advertising. Now OpenAI appears to be acknowledging that scattering resources across multiple fronts wasn't working.

Codex, their coding assistant, has become a legitimate alternative to Anthropic's Claude in developer circles. The company is reportedly planning to combine ChatGPT, Codex, and their Atlas browser into a desktop "super app" for work. That's where the compute is going. That's where the talent is going. Bill Peebles, who led the Sora team, isn't being let go—he's moving to world simulation research focused on robotics and "automating the physical economy."

The Disney Postscript

One casualty of the Sora shutdown: a partnership with Disney that was supposed to involve a billion-dollar investment. Disney's statement was diplomatically bland: "We respect OpenAI's decision to exit the video generation business and to shift its priorities elsewhere."

The reaction to Sora's demise split along predictable lines. Some celebrated it as the death of "AI TikTok," another attention-harvesting machine we didn't need. Others noted that over 100 companies are now working in AI video generation—Sora's exit doesn't mean the category failed, just that OpenAI decided it wasn't their category.

Dax, a developer who commented on the shutdown, offered the most pragmatic take: "For every successful thing that exists, 100 efforts like this had to fail. And those learnings are fed into making something that ultimately does work."

The AGI Distraction

All of this has reignited the tiresome debate about whether we've achieved AGI. Jensen Huang told Lex Fridman he thinks we have, at least for simple applications. Benjamin Todd wrote an essay arguing we haven't, pointing to AI's "jagged frontier" of capabilities—superhuman at some tasks, worse than humans at others.

The more interesting observation, buried in the discussion, is that we might have what could be called "task AGI." Ask AI to do something discrete and specific, and it often succeeds. Ask it to string together a long sequence of such tasks without human intervention, and capability drops off quickly. That's the gap between impressive demos and actual economic transformation.

Ethan Mollick suggested we should just declare that GPT-4 was AGI and move on, "to drive home the lesson that AGI alone is not enough for transformation." He's right. OpenAI and Anthropic are both partnering with consulting firms and private equity because they've realized that capable models are only the beginning. Getting them to actually work inside complex organizations is the harder problem.

The Sora shutdown isn't really about video generation or compute constraints or even strategic focus. It's about OpenAI accepting that the hard part isn't building artificial general intelligence—it's making that intelligence generally useful. That work happens in spreadsheets and codebases and ERP systems, not in viral video apps.

Whether they can actually deliver on that promise is a different question. But at least now we know which question they're trying to answer.

—Bob Reynolds