OpenAI's Most Expensive AI Project Stuck in the Mud: GPT-5 Development Costs Exceed $1 Billion but Still Unproductive

News Summary

OpenAI's development of its next-generation artificial intelligence model, GPT-5 (code-named "Orion"), has encountered significant setbacks, with the project severely over budget and delayed by several months. The original mid-2024 release plan has now become a pipe dream. This highly anticipated AI breakthrough product is currently facing multiple challenges, including a severe shortage of high-quality training data, soaring development costs reaching billions of dollars, and internal management turmoil.

Core Challenges: Data Scarcity and Cost Crisis

According to informed sources, during the 18-month development of the GPT-5 project, OpenAI has completed at least two rounds of large-scale training runs, each lasting several months, but the results have not met researchers' expectations. OpenAI conducted at least two massive training runs, each involving months of processing vast amounts of data, with the goal of making Orion smarter. However, each test run presented new problems, and the software's performance fell short of researchers' expectations.

The most critical obstacle is the scarcity of high-quality training data. The public internet, once a rich resource, has been thoroughly mined by OpenAI and other AI developers, leaving little fresh and diverse content to train the next generation of models. To address this challenge, OpenAI has begun hiring experts such as software engineers, mathematicians, and physicists to create original training materials and explore synthetic data generation techniques.

Astronomical Development Costs

The sharp increase in training costs has become another major hurdle. A six-month training run alone could cost approximately $500 million, and Altman stated that training runs for models on the scale of GPT-5 could cost up to $1 billion. This figure does not include the capital expenditure for building data centers. Reportedly, OpenAI's first Stargate supercomputing cluster is expected to contain up to 2 million GPUs, with a construction cost of $100 billion and annual power consumption equivalent to that of a large U.S. city.

Internal Turmoil Worsens Matters

Internal management turmoil within the company has further exacerbated development difficulties. In 2024, over 24 senior leaders departed, including Chief Scientist Ilya Sutskever and Chief Technology Officer Mira Murati. The departure of these key personnel has created leadership vacuums and led to a loss of institutional knowledge and expertise.

Technical Bottlenecks and Practical Dilemmas

Despite massive investments, GPT-5's performance improvements have been unsatisfactory. According to sources cited by The Information, as of June 2025, the models under development at OpenAI are not powerful enough to be called GPT-5. Industry experts point out that the current large language model architecture may be approaching its fundamental limits, with these models having limited ability to distinguish fact from fiction and prone to "hallucinations," i.e., confidently outputting fabricated or incorrect information.

Industry Impact and Competitive Landscape

OpenAI's predicament reflects the challenges facing the entire AI industry. The traditional "scaling law"—the method of improving model performance by adding more data and computational resources—appears to be reaching a point of diminishing returns. This forces AI companies, including OpenAI, to rethink their development strategies, shifting towards more efficient algorithmic improvements and specialized model development.

Meanwhile, competitors such as Google's Gemini series, Anthropic's Claude models, and China's DeepSeek are actively advancing their respective AI projects, putting additional market pressure on OpenAI.

Future Outlook: Strategic Transformation

Facing these challenges, OpenAI is adjusting its development strategy. CEO Sam Altman's roadmap, announced in February 2025, indicates that the company will first release GPT-4.5 (internal code-named Orion), which will be the last non-chain-of-thought reasoning model. Subsequently, GPT-5 will integrate the strengths of reasoning models (such as the o3 series) to form a unified AI system.

Despite numerous difficulties, OpenAI still plans to release GPT-5 in August 2025. However, industry experts remain cautious about its ability to achieve the anticipated breakthrough improvements. This tortuous development journey not only tests OpenAI's technical capabilities and financial resilience but may also redefine the trajectory of the entire AI industry, marking a strategic shift from mere scale expansion to more refined technological innovation.

Market Reaction and Investor Confidence

OpenAI is currently valued at $157 billion, a valuation largely based on the expectation that GPT-5 will bring a "significant leap." The project's delays and technical challenges will undoubtedly affect investor confidence, especially given that Microsoft, the largest investor, had previously expected to see GPT-5 released by mid-2024.

This development highlights the complexity and uncertainty of innovation in the AI field, and it reminds market participants that even industry leaders face challenges related to technological boundaries and resource limitations. GPT-5's ultimate performance will largely determine whether OpenAI can maintain its leading position in the AI race and the future direction of large language model development as a whole.