Roko's Basilisk
Posts
Why Sora Had To Die

Why Sora Had To Die

Plus: AI music overload, the energy warning, and NVIDIA's China premium.

Roko's Basilisk
May 06, 2026

Here’s what’s on our plate today:

🧪 Why OpenAI had to kill Sora to save its future.
📰 AI music's quiet flood, the gigawatt warning, NVIDIA's $1M black market.
🛠️ Three tools worth trying: Dust, Krea, Lavender.
🗳️ Poll: What's the real lesson from Sora's shutdown?

Let’s dive in. No floaties needed…

Teach your CRM how you sell. It does the rest.

Lightfield, an AI-native CRM, just shipped Skills.

Describe any workflow in plain English — account research, deal scoring, call prep, pipeline forecasting — and your CRM learns it. Trigger it with one sentence.
"Prep me for my call with Acme." It pulls every email, transcript, and deal note. Writes your brief.
"Score every deal in my pipeline." Done.

Teach it how you sell and watch it go to work for you. 2,500+ startups already have.

_{*This is sponsored content}

The Laboratory

TL;DR

$15M/day for vibes: Sora burned $5.4B annualized in compute while generating just $2.1M in total revenue. Its entire lifespan earned less than three hours of peak operating costs.
Anthropic forced the math: While OpenAI poured GPUs into video clips, Anthropic's Claude Code hit $2.5B annualized revenue and overtook OpenAI’s overall run rate, prompting leadership to call Sora a ‘side quest.’
Disney got ghosted: The proposed $1B equity deal collapsed alongside the product, barely three months after the announcement.
Compute is zero-sum now: Every Sora clip rendered was a Codex session that didn’t happen. So, if the best-funded AI company has to pick winners, every enterprise building on AI platforms should be demanding discontinuation clauses yesterday.

Why OpenAI had to kill Sora to save its future

No matter how ambitious the vision, an enterprise cannot survive without a sound business plan, one that allows both the company and its investors to generate returns. In artificial intelligence, the vision is clear, investors are willing, and the infrastructure is being built rapidly. Yet one critical piece remains unresolved: a durable business model, especially for companies like OpenAI.

In the short time since it reshaped the public imagination of what AI is and what it might become, OpenAI has explored multiple paths to turn that momentum into a sustainable business. So far, however, the difficulty of doing so has remained evident, and Sora has become one of the clearest examples.

On March 24, 2026, OpenAI abruptly shut down Sora, its high-profile text-to-video product, just three months after announcing a proposed partnership with The Walt Disney Company that included a planned $1B equity investment and a multi-year licensing arrangement covering major franchises. According to Reuters, no definitive agreements had been signed, and the deal collapsed alongside the product.

OpenAI shut down Sora just three months after unveiling a proposed $1 billion partnership with The Walt Disney Company, underscoring how quickly strategy can change in the high-stakes race to build a profitable AI business. Photo Credit: Bloomberg via Getty Images.

The sudden reversal raised a larger question: why would one of the world’s most valuable AI companies walk away from a marquee product and a headline partnership at the same time? The answer goes beyond Sora itself. It lies in the harsh economics of generative AI, where compute costs remain immense, competition is intensifying, and capital markets increasingly demand a credible path to profitability.

Sora served as an early measure of whether frontier AI products could translate technical spectacle into scalable business models. Its shutdown suggested that OpenAI is confronting a harsher reality: in a capital-intensive market defined by rising compute costs and intense competition, even impressive products may be difficult to justify if they cannot support the company’s larger path to sustainability.

As OpenAI pursued consumer-facing products like Sora, Anthropic rapidly expanded its enterprise business, showing how coding tools and software customers have become one of the most valuable revenue engines in the AI race. Photo Credit: Reuters.

The math that broke Sora

For many, Sora represented OpenAI’s expansion beyond chatbots and into media creation. However, despite its popularity, it was never a viable business, and the numbers make that clear without much interpretation required.

OpenAI’s video-generation model used a diffusion-transformer architecture (a system that processes video as a continuous structure across space and time, rather than generating one frame at a time). And while the approach produced impressive results, it consumed GPU resources at a rate that dwarfed every other product in OpenAI’s portfolio.

According to Forbes estimates cited across multiple outlets, Sora was burning through roughly $15M/day in compute costs at peak usage, or about $5.4B annualized. Cantor Fitzgerald analyst Deepak Mathivanan broke down the per-clip economics: each 10-second video costs approximately $1.30 to generate, requiring about 40 minutes of total GPU time across four parallel processors.

Against that cost structure, the revenue was negligible: mobile intelligence firm Appfigures estimated Sora’s total lifetime in-app purchase revenue at roughly $2.1M, meaning the product’s entire commercial lifespan generated less revenue than roughly three hours of its peak operating costs.

The user trajectory also told the same story. The standalone Sora 2 app launched in September 2025 and initially performed well, surpassing 1M downloads within five days and peaking at roughly 3.3M monthly downloads in November 2025.

But the growth did not sustain, and engagement soon collapsed. Downloads fell 32% in December, another 45% in January, and reached about 1.1M by February 2026, a roughly two-thirds decline from the peak. To make matters worse, the app’s active users are estimated to have peaked at around 1M and have since dropped below 500k.

Bill Peebles, head of the Sora team, acknowledged the problem publicly in an October 2025 social media post, stating that the economics were “completely unsustainable.” That was five months before the shutdown.

The competition OpenAI couldn’t ignore

Despite the unsustainable economics, one should remember that OpenAI has a long history of running products at a loss while building market position. So why did it kill Sora?

Beyond the direct economic burden, which OpenAI might have been willing to tolerate for longer, Sora also carried a substantial opportunity cost that became harder to ignore in an increasingly competitive market. What changed was the rapidly shifting landscape, where Anthropic was moving quickly to capture the enterprise revenue streams OpenAI needed most.

The Wall Street Journal investigation into the shutdown reported that while OpenAI’s Sora team was focused on video generation, Anthropic was winning over the software engineers and enterprise customers that drive durable revenue. Claude Code, Anthropic’s coding tool, reached $1B in annualized revenue within six months of its public launch in mid-2025 and hit $2.5B by February 2026.

Anthropic’s overall annualized revenue run rate surpassed $30B by April 2026, overtaking OpenAI’s approximately $25B.

The company’s internal response was blunt. At an all-hands meeting on March 16, OpenAI’s applications CEO Fidji Simo told employees that Anthropic was a ‘wake-up call.’ In a subsequent internal memo, she wrote that the company was ‘spreading its energy across too many applications and technology stacks’ and needed to simplify and focus. Sora, she reportedly told staff, was a ‘side quest,’ and eight days later, the product was dead.

What the GPU dashboard reveals

Inside OpenAI, researchers could track chip allocation through an internal dashboard. Some were surprised by the resources devoted to Sora, given its limited revenue and lack of contribution to the company’s core language model capabilities. The project’s secrecy led former employees to describe it as a “startup within a startup.”

This is the part of the story that extends beyond OpenAI. Every frontier AI company faces the same fundamental constraint: GPU capacity is finite, making product strategy a zero-sum game. A compute resource directed to one service cannot be used elsewhere. In such a scenario, every Sora clip that rendered consumed resources that could otherwise have powered a Codex coding session, an API request, or a ChatGPT query. For a company generating roughly $2B in monthly revenue primarily from text-based products, allocating scarce compute to a video app used by fewer than 500k people carried an enormous opportunity cost.

The resources freed by Sora’s shutdown are being redirected to higher-value targets identified by OpenAI. Even the Sora team itself is not being disbanded.

According to OpenAI, it will continue as a research unit focused on world simulation for robotics, repurposing the model’s understanding of physics and motion for applications with clearer commercial paths. The company has also announced a unified ‘superapp’ combining ChatGPT, Codex, browsing, and agentic capabilities into a single platform.

The collateral damage

The shutdown of Sora, while beneficial for OpenAI, did have wider repercussions, with the most immediate casualty being the Disney partnership.

A legal analysis from law firm Ropes & Gray noted that the licensing terms had been agreed before the equity investment closed, meaning Disney likely had a contractual right to walk away once Sora was killed. The entertainment company’s public statement was measured but unmistakable: “We respect OpenAI’s decision to exit the video generation business and to shift its priorities elsewhere.”

Beyond Disney, Sora left behind a difficult ethical record. The app’s ‘Characters’ feature, which allowed users to scan their faces into videos, produced deepfakes of deceased public figures, including Martin Luther King Jr. and Robin Williams. This led to complaints from their families. Users generated videos of copyrighted characters in inappropriate situations, testing legal boundaries that OpenAI’s moderation systems were not equipped to enforce.

A Brookings Institution analysis estimated that more than 100k U.S. entertainment jobs could be disrupted by generative AI tools by 2026, and that 75% of film companies that adopted AI had already reduced or eliminated jobs.

Meanwhile, for enterprise decision-makers, the Disney episode introduces a new risk category. The abruptness of Sora’s end, a product killed three months after anchoring a billion-dollar partnership, will likely push enterprise buyers to demand contractual guarantees against platform discontinuation and to build more flexible, multi-vendor AI strategies.

The era where AI companies could pursue every frontier simultaneously appears to be ending as compute scarcity is imposing a discipline that the market, with its appetite for spectacle and its tolerance for losses, would not impose on its own.

OpenAI’s decision to kill Sora, then, appears to be a recognition that vision without prioritization becomes a liability. If the most well-funded AI company in the world has to choose which products deserve to exist, the question for every company building on or with AI is whether they have made that choice yet, and whether they will get enough notice when someone else makes it for them.

Quick Bits, No Fluff

AI music flooding streaming services who want it: AI-generated music is flooding Spotify and other platforms, but listener demand is unclear.
Data center expert warns of gigawatt-scale energy demands: A data center expert warns about the energy demands of AI infrastructure.
NVIDIA B300 server $1M China US curbs: NVIDIA's B300 servers are selling for $1M in China due to US export controls.

Outperform the competition.

Business is hard. And sometimes you don’t really have the tools you need to be great at your job. Well, Open Source CEO is here to change that.

Tools & resources, ranging from playbooks, databases, courses, and more.
Deep dives on famous visionary leaders.
Interviews with entrepreneurs and playbook breakdowns.

Are you ready to see what’s all about?