In the span of one week in late March 2026, the two most prominent AI safety and capability labs both stumbled publicly. OpenAI shut down its flagship video product after burning an estimated $1 million per day. Anthropic, the company founded explicitly on the premise that AI safety requires institutional discipline, exposed its most powerful unreleased model through a basic infrastructure misconfiguration. These are not isolated incidents. They are data points in a pattern.
Key Highlights
- OpenAI shut down Sora on March 24 after approximately $1 million per day in operating losses
- Anthropic exposed Claude Mythos details through an unsecured public data store on March 26
- A $1 billion Disney partnership collapsed with Sora’s shutdown
- Anthropic is simultaneously warning governments about Mythos cybersecurity risks while failing basic operational security
- OpenAI’s $840 billion valuation requires sustained revenue growth that products like Sora cannot deliver
The Capability Viability Gap
The AI industry’s central challenge is not building capable models. It is building capable models that can sustain themselves commercially. Sora could generate impressive video. It could not generate enough revenue to cover the compute cost of doing so. This is the same challenge facing every frontier AI product that requires intensive inference at consumer price points.
OpenAI’s core business, ChatGPT and the API, is growing rapidly. But the company’s ambition extends far beyond text. Sora was supposed to prove that generative AI could dominate video, audio, and eventually physical world simulation. Its failure does not invalidate the technology. It invalidates the business model of deploying compute heavy AI products at consumer pricing before inference costs come down.
The Safety Credibility Problem
Anthropic’s Mythos leak is a different kind of failure, but it points to the same structural issue: institutional capacity has not kept pace with technical ambition. Anthropic was founded specifically to build AI responsibly. Its entire brand is predicated on safety research, careful deployment, and institutional discipline.
Exposing nearly 3,000 unpublished assets, including capability assessments of your most powerful unreleased model, through a publicly accessible data store is not a sophisticated breach. It is a configuration error. The kind of error that a well resourced security team should catch in a routine audit. The kind of error that becomes more likely when an organisation is scaling faster than its operational processes can keep up.
The irony is sharp. Anthropic is privately briefing government officials on the cybersecurity risks of Mythos while failing to secure its own blog infrastructure. The model might indeed be dangerous. But the institution deploying it just demonstrated that it cannot reliably manage its own data hygiene. That should concern the same government officials receiving those briefings.
What the Market Is Pricing In
The public AI market is pricing in perfection. OpenAI’s $840 billion valuation assumes sustained revenue growth, successful product expansion, and no major operational failures that damage enterprise trust. The Sora shutdown and the Disney partnership collapse are exactly the kind of events that make enterprise procurement teams cautious.
SoftBank just took a $40 billion loan to fund its OpenAI commitment, expecting an IPO exit within 12 months. The IPO thesis depends on the narrative that OpenAI is an unstoppable growth machine. Every product failure, every partnership collapse, every Sora sized write off makes that narrative slightly harder to sustain.
This does not mean the IPO will not happen. It means the valuation conversation at IPO will be more nuanced than it would have been a month ago.
The Pattern Across the Industry
Sora and Mythos are the most visible examples, but the pattern extends across the AI industry. Companies are shipping capabilities faster than they can commercialise them, secure them, or govern them. The inference cost curve is improving but not fast enough. The safety frameworks are developing but not fast enough. The operational security practices are maturing but not fast enough.
The gap between what AI can do and what AI companies can sustainably do with it is the defining tension of 2026. The companies that close this gap, that match capability with commercial viability and operational discipline, will be the ones that survive the inevitable consolidation. The ones that keep pushing capability forward without solving the business and governance side will burn through capital and credibility at a rate that no amount of fundraising can sustain.
The TCB View
The AI industry is not in a bubble. The technology is real and the applications are genuine. But the industry is in a gap, a gap between what the technology can demonstrate and what the institutions behind it can reliably deliver as sustainable businesses.
Sora proved that generating impressive AI video is possible. It also proved that generating it profitably is not, at current costs. Anthropic proved that building the most powerful AI model is within reach. It also proved that even the most safety conscious lab can fail at basic infrastructure security. Both lessons are important. Both suggest that the AI industry’s greatest risks are not technical. They are institutional.
The next 12 months will separate the AI companies that can build great products from the ones that can build great businesses. Those are not the same thing. The market is starting to notice.
The Daily Brief by TCB
Crypto, AI & finance intelligence in 5 minutes. Every weekday morning. Free.
