The end of OpenAI's Sora video AI – "Spud" is coming: When computing power is more important than visions

Xpert Pre-Release

Available in 27 languages 📢

Published on: March 25, 2026 / Updated on: March 25, 2026 – Author: Konrad Wolfenstein

The end of OpenAI's Sora video AI – "Spud" is coming: When computing power is more important than visions – Image: Xpert.Digital

$15 million in losses per day: The real reason for the sudden end of OpenAI Sora

Billion-dollar deal with Disney falls through: Why OpenAI is pulling the plug on its flagship project

Too expensive, too hungry: The brutal truth about the demise of the revolutionary video AI Sora

It was a bombshell that caught the technology and entertainment worlds completely off guard: Just 15 months after its spectacular launch and shortly after announcing a historic billion-dollar deal with Disney, OpenAI has surprisingly shelved its revolutionary video model, Sora. What at first glance appears to be an inexplicable setback in the rapid development of AI, on closer inspection reveals itself to be a harsh admission of economic reality. Exorbitant operating costs of an estimated $15 million per day, insufficient user revenue, and a massive global shortage of high-performance chips forced the company, under Sam Altman, to radically prioritize. Shortly before a planned gigantic IPO, OpenAI is now parting ways with its most expensive cost drain. The freed-up resources will instead be invested in a new, top-secret super-language model called "Spud," as well as in the ultimate path to generalized artificial intelligence (AGI). Read here why the video revolution failed for now due to the mundane arithmetic of server costs, what the abrupt end means for the entire AI industry, and how OpenAI is completely realigning its future.

How OpenAI buried its most expensive mistake – and what that reveals about the limits of the AI industry

March 24, 2026, marks one of the most unusual strategic decisions in the recent history of the technology industry: OpenAI, the company that describes itself as committed to the well-being of all humankind, completely shut down its much-lauded AI video generator, Sora—a mere 15 months after its public launch and only three months after signing a multi-billion-dollar licensing agreement with Walt Disney. The terse statement on the social network X simply read: “We are saying goodbye to Sora.” What sounds like a matter-of-fact press release is, in reality, an admission of fundamental economic contradictions that call into question the entire business model of the AI industry.

From showcase technology to cost-cutter: Sora's brief history

When OpenAI released the first demonstration videos of Sora in February 2024, the tech world reacted with almost incredulous amazement. The ability to generate flowing, physically plausible video clips from just a few words of text was considered a technological quantum leap. Directors, advertising agencies, and content creators sensed a revolution. The entertainment industry was abuzz. The model seemed not only to simulate video production but also to actually understand fundamental laws of physics—objects behaved coherently in space, shadows fell correctly, and water flowed naturally.

This claim was not mere marketing rhetoric. OpenAI itself described Sora's architecture as a step towards a "World Simulator"—a system that not only reproduces visual patterns but also internalizes the physical logic of the world. Bill Peebles, the Sora group's technical lead, who co-developed the underlying diffusion transformer architecture principle, spoke of a technology that could eventually replace biological laboratory experiments in digital environments.

The commercial launch as a standalone iOS app in September 2025 seemed to confirm this euphoria: Within a week, Sora surpassed the one million download mark, even overtaking ChatGPT in the speed of its early distribution. By Halloween 2025, 4.5 million downloads had been recorded, and the platform climbed to the top of the US App Store charts. The signing of a three-year licensing agreement with The Walt Disney Company in December 2025—including a planned equity investment of one billion dollars and the rights to use over 200 characters from the Disney, Marvel, Pixar, and Star Wars franchises—seemed like the final confirmation that Sora would help shape the future of audiovisual storytelling.

But hidden behind the shiny facade lay a problem that was known early on and that no amount of positive user numbers could solve: the sheer cost structure.

The brutal economics of video generation: Why Sora was doomed to fail financially

The fundamental economic challenge of AI-generated video lies in the information density of the output. While a language model produces text consisting of discrete tokens, a video model must coherently generate thousands of image pixels across multiple frames for every second of film, while maintaining physical continuity. This means that each generated clip requires computational operations on a scale that far exceeds those required for text or image generation by orders of magnitude.

Analysts at the financial firm Cantor Fitzgerald calculated that producing a single ten-second Sora video cost OpenAI around $1.30 in computing power—an amount that experts at SemiAnalysis considered a conservative estimate. Extrapolating this to actual usage, the picture was one that caused a stir even in Silicon Valley: According to Forbes estimates, OpenAI was burning through roughly $15 million daily just to operate its Sora infrastructure, equivalent to over $5.4 billion annually. This sum represents more than a quarter of the company's total annual revenue.

Bill Peebles, the head of the Sora team, summed up the situation with remarkable candor on October 30, 2025: “The economics are completely unsustainable right now.” This statement from a senior executive about his own product is virtually unparalleled in its stark honesty in Silicon Valley corporate history. OpenAI was caught in a two-sided trap: the platform was undervalued for users to cover costs, but any significant price increase would have caused its user base to collapse immediately.

To make matters worse, usage growth had already peaked by fall 2025. Monthly downloads plummeted by 32 percent in December 2025—a month typically boosted by new smartphones and the holiday season for app platforms. In January 2026, installations dropped another 45 percent to 1.2 million, while consumer revenue also fell by 32 percent. The platform's total lifetime revenue amounted to a mere $1.4 million from a total of 9.6 million downloads—a coverage rate that, even under the most optimistic scenario, was less than one percent of actual operating costs.

These figures highlight a structural problem that extends beyond Sora: AI products based on particularly computationally intensive modalities cannot support consumer-oriented business models in their current technological development phase. The cost argument, as one analyst put it, is pure arithmetic.

GPU scarcity as a strategic bottleneck resource: The core of the decision

OpenAI's decision to shut down Sora cannot be fully understood without considering the overarching resource crisis the company faces. Computing power in the form of graphics processing units (GPUs), particularly NVIDIA's high-performance processors like the H100 and the new Blackwell chip, is the critical bottleneck in the entire race for AI dominance. This bottleneck is not only financial but also physical: even well-funded companies cannot simply purchase unlimited quantities of GPUs, as delivery times are frozen due to contractually agreed-upon quotas.

Sam Altman had already publicly admitted in February 2025 that OpenAI "had run out of GPUs" and therefore had to delay the rollout of GPT-4.5. This statement was not a PR stunt, but rather an expression of a real shortage that has since reached systemic proportions. DDR5 memory kits, which cost around $90 in 2025, have risen to $240 or more. Data centers are facing long lead times for new GPU orders, and new ordering capacity is reserved exclusively for existing partners.

Against this backdrop, shutting down Sora becomes a necessary resource decision. Every GPU rendering video frames is unavailable to answer speech model requests—and speech models generate the overwhelming majority of revenue. ChatGPT, with over 800 million weekly active users, its enterprise API, Codex for software development, and the planned "super app" that combines ChatGPT, Codex, and an AI browser under one roof—all these products have direct monetization that exceeded Sora's revenue structure a hundredfold.

In a company projecting losses of around $14 billion for fiscal year 2026 and which, according to its own forecasts, will not reach profitability until 2029 or 2030 at the earliest, the question of which product to allocate scarce computing time to is not a strategic one, but a matter of survival. HSBC analysts estimate that OpenAI may need over $207 billion in additional funding by 2030.

The Disney debacle: Collateral damage of a strategic shift

The shutdown of Sora left not only technical scars – it also tore apart one of the most sensational corporate partnerships in recent technological history. The agreement signed with Walt Disney in December 2025 was considered a game-changer in the industry, one that was meant to catapult AI-generated entertainment content into the mainstream. Under then-Disney CEO Bob Iger and Sam Altman, a three-year licensing agreement for over 200 characters and a planned equity investment of one billion dollars were agreed upon.

This agreement has now completely collapsed. According to information from The Hollywood Reporter, Disney is also withdrawing from the broader partnership with OpenAI, which included the equity investment and the API licensing agreement. Because the investment was structured as stock options and not as a cash payment, no capital changed hands. The financial damage for both parties is therefore limited, but the reputational damage for OpenAI is considerable: A highly publicized deal has been completely reversed within a few months.

For the entertainment industry, this development sends a mixed signal. On the one hand, it emboldens critics who have warned against over-reliance on individual AI providers. On the other hand, the industrial need for long-term AI imaging solutions is not diminishing – it is simply shifting to other providers. Competitors such as Runway, Google with Veo, and Chinese companies like Kuaishou will benefit from OpenAI's strategic withdrawal from the video sector. Disney stated that it respects OpenAI's decision to withdraw from video generation and will continue to explore collaborations with other AI platforms.

🎯🎯🎯 Data-driven B2B industry hub as a quasi-in-house solution

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business

Why OpenAI is sacrificing Sora – and focusing on productivity with Spud

Codename Spud: What's behind the next language model?

The freed-up computing resources are not flowing into a vacuum, but into a concrete project: the language model developed under the internal codename "Spud," whose pre-training, according to "The Information," has already been completed. The model's exact capabilities remain undisclosed, but Sam Altman is said to have told employees that the model could "really accelerate the economy." This statement is noteworthy and fits into a larger narrative strategy: moving away from the fascination with visual AI effects and toward measurable, productive impacts.

Spud's position within the OpenAI model hierarchy is still unclear. Speculation within the community ranges from a classification as GPT-6, representing a complete generational shift, to an incremental iteration like GPT-5.5. It is known that OpenAI models are evaluated along an "expert parity" scale: GPT-5 reportedly achieved 38 percent parity with human experts, GPT-5.2 reached 70.9 percent, and GPT-5.4 83 percent. A model intended to "accelerate" the economy must rank significantly higher on this continuum.

Hints from internal communications suggest that Spud could be natively multimodal and particularly strong in productive, business-oriented tasks such as coding, planning, analysis, and automation. The strategic focus on a future "super app" that combines ChatGPT, Codex, and an AI browser under one roof suggests that Spud will form the backbone of this integrated platform. It would therefore be less of a research model and more of a directly commercially oriented tool—a shift symptomatic of OpenAI's overall transformation.

“AGI Deployment”: The renaming as a program

Parallel to the Sora shutdown, OpenAI underwent another, less spectacular, but at least equally significant change: The product organization, led by Fidji Simo, was renamed "AGI Deployment." This renaming is far more than a purely semantic act. It signals a fundamental shift in the company's internal self-perception: from a technology company that develops products to a company that sees itself as actively engaged in the AGI deployment phase.

This self-positioning stands in tension with simultaneous measures that demonstrate a more commercially oriented approach. In February 2026, OpenAI dissolved its "Mission Alignment" team, which had been responsible for public communication of the company's original mission. Joshua Achiam, the former team leader, was given the newly created title of "Chief Futurist"—a role that addresses forward thinking but implies no operational authority in the core business. Previously, in 2024, OpenAI had already eliminated its "Superalignment" team, which dealt with the existential risks of superintelligent AI.

The word "safely" has even been removed from the official mission statement that OpenAI submitted to the US Internal Revenue Service (IRS). This gradual abandonment of safety and mission rhetoric in favor of commercial efficiency is no accident. It reflects the increasing pressure on a company that simultaneously claims to be saving humanity and aims to go public and raise trillions of dollars.

The infrastructure question: Sam Altman and the global data center race

One of the key takeaways from the Sora episode is that the real bottleneck in the AI industry isn't algorithms or data, but physical computing infrastructure. Sam Altman has made this insight his personal mission. According to internal communications, he will now primarily focus on three tasks: raising capital, managing supply chains, and building data centers on an unprecedented scale.

The scale of these infrastructure ambitions is staggering. The Stargate project, a collaboration with Oracle and SoftBank, envisions investments of $500 billion over four years in new AI infrastructure, with $100 billion immediately allocated to US data centers. A strategic partnership with NVIDIA aims to build at least ten gigawatts of AI data centers, supported by a planned NVIDIA investment of up to $100 billion, released with every gigawatt deployed. The first phase, using NVIDIA's Vera Rubin platform, is slated to go live in the second half of 2026.

These infrastructure investments, however, are not without significant risks. Forbes analysis from December 2025 showed that a company with $20 billion in annual revenue can hardly invest $1.4 trillion in infrastructure without risking dangerous financial difficulties—a parallel to the exuberant telecom infrastructure projects before the dot-com bubble burst. Indeed, OpenAI has already scaled back: The company is abandoning ambitious construction plans and accepting its role as a major consumer of cloud services from partners like Microsoft Azure, Oracle, and Amazon Web Services, rather than operating its own data centers on a large scale. According to CNBC, OpenAI currently owns no data centers and will not do so in the foreseeable future.

World models and robotics: The true legacy of Sora

Those who reduce the story of Sora's shutdown to the story of a failed product miss the deeper strategic logic. In its statement to the New York Times, OpenAI explicitly emphasized that video generation technologies will continue to be used internally—for training robots. The Sora team itself was not disbanded, but rather redirected to a new task: research on so-called world models and physical simulation environments, primarily intended to benefit the field of robotics.

This step has its own technological logic. The key insight Sora provided was not that AI can produce good videos, but that emerging physical intelligence develops through training on video data. OpenAI found that Sora models, when scaled sufficiently, develop "emergent capabilities": three-dimensional consistency, object permanence, realistic physics—properties that are solely a "phenomenon of scaling" and were not explicitly programmed. A model that understands how a basketball rebounds from the backboard understands basic physical causality.

This very characteristic makes world models so valuable for robotics. A warehouse robot that has tested millions of package handling scenarios in simulation—including rare edge cases that hardly ever occur in physical warehouses—develops an adaptability that purely rule-based systems cannot achieve. The spectrum of potential applications extends far beyond logistics: molecular structure simulations in chemistry, the modeling of physical laws, climate system forecasting, and training for medical procedures are just a few of the application areas mentioned.

In parallel, competitors are working intensively on the same topics. NVIDIA has released simulation tools for robot AI with the Cosmos framework and the Newton Physics Engine, and has made the Isaac GR00T basic models available as open source. Google DeepMind has introduced its own world model architecture, Genie 3. Yann LeCun, the prominent AI researcher, left Meta at the end of 2025 to launch his own world model startup – with a target valuation of $3.5 billion. The competition for dominance in physics-based AI simulation has thus reached a new level of intensity.

IPO preparation and the logic of portfolio streamlining

Hardly any aspect of the Sora shutdown can be completely separated from OpenAI's upcoming IPO. The company is aiming for an IPO in the fourth quarter of 2026, with a valuation between $830 billion and $1 trillion. For potential shareholders, the coherence of the business model is crucial: A company that burns through $15 million a day on a service with minimal revenue is difficult for institutional investors to value.

The shutdown of Sora, in this interpretation, sends a clear signal to the capital market: OpenAI is prioritizing capital efficiency and focus. This is all the more remarkable given that the company continues to incur massive losses – projected at $14 billion in 2026 alone. The difference lies in the nature of the losses: investments in language models that generate direct revenue are considered strategically acceptable "growth losses," while subsidizing a loss-making video platform is classified as operational inefficiency.

The situation is comparable to the classic "focus and kill" startup scenario: The company streamlines its portfolio before going public to present a clear investment thesis. ChatGPT, with 800 million weekly users and annualized revenue exceeding $20 billion by the end of 2025, is such a clear thesis. A loss-making video app with 9.6 million lifetime downloads and $1.4 million in total revenue is not.

The dissolution of the Disney partnership has an ironic twist in this context: Disney hadn't made any cash payment, but rather structured its stake entirely through stock options. Strictly speaking, the termination of the deal therefore cost OpenAI nothing – except for the publicity-generating loss of face that inevitably accompanies withdrawing from such a high-profile agreement.

Scaling limits and their strategic consequences: A comprehensive perspective

The story of Sora's rise and fall is symptomatic of a deeper crisis in the current AI paradigm: the limits of hardware scaling. Until recently, the prevailing thesis in AI research was that more computing power almost automatically leads to better models. This scaling hypothesis has dominated development since GPT-2 and has seemingly been impressively confirmed.

However, Sora demonstrates that scaling alone does not create a viable business model if the associated costs cannot be covered by adequate revenue. The structural problem is not technical, but economic: video AI is currently simply too expensive for the consumer market to be offered profitably. Analysts estimated that the cost of video generation would decrease by approximately fivefold by the end of 2026 – but this window of opportunity was no longer economically viable for OpenAI.

This has consequences far beyond OpenAI. Other players in the field of generative video AI—Runway, Pika, Stability AI, and Google DeepMind with Veo—face the same fundamental cost challenges. Those who cannot cross-subsidize the deficit from other profitable products will feel similar pressure. Compared to startups focused solely on video, OpenAI had the luxury of diversifying its business. More specialized companies without this diversified portfolio approach face a significantly greater risk of failure.

At the same time, OpenAI's withdrawal from the video sector opens up market opportunities for competitors. In particular, Chinese providers like Kuaishou (with its Kling model) and international platforms like Runway could fill the resulting gap in the professional video segment. The irony is that OpenAI, with its pioneering work, validated the entire market for AI-powered video generation – and is now exiting it just before the technology becomes commercially mature.

From flagship product to strategic scapegoat

Sora's shutdown is not a sign of failure in a technological sense. The product worked. It delighted users. It was technically groundbreaking. It failed due to the fundamental incompatibility between computing costs and willingness to pay – a gap that could not be closed in the short term. In this sense, OpenAI's decision is rational and logical.

What the episode reveals, however, are the structural tensions of a company that simultaneously proclaims its commitment to the well-being of humanity, posts massive losses, is working towards a billion-dollar IPO, is disbanding its security and mission teams, and is pursuing a trillion-dollar data center project. Renaming the product organization to "AGI Deployment" might sound like hubris to outsiders, but internally it's a deliberate strategy: all resources are focused on the core objective, whatever the next step towards AGI may be.

The codename "Spud"—ironically, the English word for potato, the most down-to-earth of all vegetables—symbolizes this pragmatism. No visions of Hollywood machines and dancing Mickey Mouse, but a linguistic model designed to accelerate the economy. It's a commitment to mundane utility rather than spectacular demonstrations, to long-term monetization rather than short-term buzz.

Whether Sam Altman is right about this, only time will tell. What is already clear is that with the shutdown of Sora, OpenAI has demonstrated a strategic discipline that is rare in the hype-driven AI industry – the willingness to bury a celebrated product when strategically necessary. This is either visionary pragmatism or an admission of having taken on too many projects at once. It's probably both.