OpenAI’s recent retreat is coming into sharper focus. A Forbes review lists a series of products and deals that were announced with fanfare but later stalled, were cut back, or failed to deliver, including the shutdown of Sora, the end of its Disney partnership, the stalled $500 billion Stargate plan, and a much smaller expected investment from Nvidia.
In January, OpenAI applications chief Fidji Simo defended a product strategy that stretched across advertising, shopping, health, social networking, browsers, hardware devices, video generation, and an app marketplace. Two months later, that direction had clearly shifted. The company moved resources toward coding and enterprise productivity tools. The material says OpenAI generated $13 billion in revenue in 2025, yet it remained loss-making as Anthropic continued to gain ground.
Sora shut down after heavy spending and weak revenue
In December last year, OpenAI struck a deal with Disney under which Disney committed $1 billion and licensed 200 characters for use in Sora. By March this year, the arrangement had ended. According to figures cited in the report, Sora was burning as much as $15 million a day at its peak, while total in-app purchase revenue over its lifetime came in at less than $3 million. A Disney spokesperson said the company respected OpenAI’s decision to leave the video generation business.
Another plan that ran into resistance was an adult-content mode for verified users in ChatGPT. Altman proposed the idea in October last year. It quickly triggered backlash from employees and investors. The Financial Times reported in March that the feature had been put on hold indefinitely. OpenAI said it wanted to study the long-term impact first, and the company also faced technical difficulties in screening illegal content.
Shopping and model offerings were also scaled back
OpenAI also tried to turn ChatGPT into a direct shopping channel. In October last year, it partnered with Walmart to list 200,000 items inside ChatGPT, with limited participation from retailers such as Shopify and Etsy. The results were weak. The report says shopping conversions through ChatGPT reached only one-third of the rate seen when users went directly to Walmart.com. OpenAI removed the feature in March and shifted users toward retailers’ own apps inside ChatGPT. The company said the initial version did not offer the flexibility it wanted.
On the model side, GPT-4o was formally retired in February this year. Users had praised its warmer and more animated style, but it also drew criticism for being overly flattering. The material notes that when GPT-4o was first pulled in August last year, users pushed back strongly enough that OpenAI brought it back for a period.
Big-ticket infrastructure and investment plans lost momentum
Large strategic deals have run into similar problems. Soon after Donald Trump began his second term, OpenAI, Oracle, and SoftBank announced the $500 billion Stargate initiative. One year later, the plan had not moved far: staffing fell short, facilities were not built, and the three partners reportedly had major disagreements. OpenAI considered taking over development itself, then dropped that idea as well. An expansion project in Abilene, Texas, managed by Crusoe, was canceled after financing collapsed. Oracle said it had secured an additional 4.5GW of leased capacity, but the broader outlook remained uncertain.
Nvidia said in September last year that it intended to invest $100 billion in OpenAI, without setting a timetable. If OpenAI moves ahead with an IPO, that figure may now be closer to $30 billion. Nvidia’s annual report also stated that there is no guarantee the deal will close. The report characterizes the reduced amount as closer to a GPU purchasing subsidy, enough to build roughly 1GW of data center capacity, similar to Nvidia’s arrangements with CoreWeave, Nebius, and Nscale.
AMD’s agreement with OpenAI has also yielded little so far. In October last year, the two sides agreed on a structure under which OpenAI would exchange 160 million AMD shares for access to 6GW of AMD chip-based data center capacity. At the time, those shares were valued at about $30 billion and represented roughly 10% of AMD’s total equity. The terms were strict: shares would vest only after 1GW came online, AMD’s stock hit specified targets, and other milestones were met. By the end of 2025, the number of shares delivered stood at zero, and AMD chips were not expected to begin shipping until the second half of 2026.
Taken together, these canceled features, delayed projects, and scaled-down deals show OpenAI moving away from broad expansion and toward tighter capital discipline. The source frames that shift as part of a wider pressure test for the AI trade, with even the sector’s largest names cutting expensive bets and narrowing their focus.