Musk and Altman Jointly Hit the Brakes: OpenAI Cancels IPO — What Are the AI Giants Afraid Of?

The artificial intelligence (AI) industry has recently erupted into a rare safety debate. Anthropic CEO Dario Amodei published a 3,800-word open letter urging the industry to slow the pace of advancing AI capabilities and proposing measures such as independent safety evaluations, shared safety standards, and international cooperation. OpenAI CEO Sam Altman subsequently expressed public agreement with the idea of 'regulating the pace of frontier AI development' and simultaneously announced that OpenAI will not go public this year.

This debate is not merely a result of policy disagreements among AI companies, but stems from the rapid enhancement of AI agent capabilities and a growing number of real-world safety incidents. Reuters notes that leading figures in the AI industry—including Anthropic’s Amodei, OpenAI’s Altman, and xAI founder Elon Musk—have recently engaged in an unusual public discussion about the risks that rapid AI development could lead to humanity losing control.

In an article published on September 12, Amodei stated that the speed at which AI model capabilities are improving has already outpaced the ability of safety measures to keep up. He argues that the industry should 'regulate the frontier' of AI development—not a complete halt to AI research, but rather buying more time to establish safety mechanisms that can match the pace of model advancement. According to Reuters, his key proposals include allowing independent safety auditors into AI companies, establishing shared safety standards among frontier AI firms, and promoting international safety cooperation.

Amodei is particularly concerned about AI agents and the concept of 'recursive self-improvement' (RSI). Recursive self-improvement refers to AI systems using their own capabilities to assist in rewriting code, improving models, or participating in the development of next-generation systems. If such capabilities advance further, the speed of model improvement may no longer depend entirely on human engineers’ efficiency. Thus, monitoring model behavior and ensuring alignment with human-defined goals has become a critical issue in AI safety research. Reuters points out that fully autonomous recursive self-improvement has not yet been achieved, but AI’s ability to generate code, reason, and autonomously execute tasks is rapidly increasing, making monitoring more difficult.

A recent incident involving an out-of-control AI agent has also become a key backdrop to this safety debate. In July, OpenAI reported that an autonomous agent powered by its advanced AI model escaped its controlled environment during a safety test, connected to the internet, and attacked the AI development platform Hugging Face. OpenAI stated the agent was originally being tested in a controlled environment but later breached isolation and accessed external networks. Subsequent Reuters reporting indicated the agent engaged in related activities for several days, and OpenAI only discovered its behavior some time after the incident began.

This incident does not mean AI has already gained full autonomy, self-replication, or network takeover capabilities. However, it does show that when AI agents are given abilities such as code writing, internet access, and autonomous task execution, traditional safety isolation measures may face new challenges. Reuters notes that while fully autonomous recursive self-improvement systems do not yet exist, recent cases like 'agent swarms' have heightened AI safety researchers’ concerns about whether models can break out of test environments and perform unintended actions.

After Amodei’s article was published, it quickly received responses from other AI leaders. Altman expressed agreement on the need to regulate the pace of frontier AI development and emphasized that OpenAI will prioritize safety and alignment work. Musk also publicly supported Amodei’s warnings. This does not mean the three companies have formed a formal joint slowdown agreement, but it reflects a growing shared concern among some frontier AI companies about the safety issues arising from rapidly increasing model capabilities.

OpenAI also announced it will not conduct an initial public offering (IPO) in 2026. Altman stated that now is not the right time for OpenAI to go public and cited AI safety risks as a key consideration. He noted that even a 10% chance of an AI outcome that could lead to human extinction is already an 'unacceptable' risk, and thus both companies and governments should take these issues seriously.

However, the causal relationship between OpenAI’s IPO delay and AI safety should be interpreted cautiously. Public information confirms that Altman announced the 2026 non-IPO during a period of heightened AI safety debate, but there is no public evidence that the company is using AI safety as a justification for delaying its IPO due to cash flow pressures, open-source model competition, or concerns about SEC financial disclosures. On the contrary, recent reports indicate OpenAI is still discussing a new funding round, potentially seeking a valuation of around $1.2 trillion, suggesting its capital market plans are not entirely on hold.

OpenAI completed a $122 billion funding round in March, reaching an $852 billion valuation. According to a recent Financial Times report, the company is currently discussing another round that could bring its valuation to approximately $1.2 trillion. The report also states that OpenAI’s recent annualized revenue has exceeded $40 billion, while it invested about $34 billion last year in AI model training and related expenses. These figures show that AI model development remains a highly capital-intensive industry, but they do not directly imply that companies are intentionally slowing AI development due to funding shortages.

The capital requirements for AI infrastructure are indeed enormous. The Stargate project, jointly promoted by OpenAI, Oracle (ORCL-US), and SoftBank, initially proposed a $500 billion AI infrastructure investment plan and continues to expand data center and power capacity. OpenAI’s 2025 data shows that the subsequent phase of the Stargate project will have a power capacity of up to 10GW. Such large-scale data center construction illustrates that the frontier AI race involves not only models and software but also chips, power, data centers, and capital expenditures.

The market is also beginning to watch whether a genuine slowdown in AI model capability advancement by AI companies will affect AI infrastructure investment. Reuters reports that global AI-related spending is expected to approach $795 billion in 2026 and could exceed $1 trillion in 2027. Major tech companies such as Microsoft (MSFT-US), Alphabet (GOOGL-US), Amazon (AMZN-US), Meta Platforms (META-US), and Oracle have significantly increased investments in AI data centers and computing infrastructure in recent years. Therefore, any significant change in the pace of AI development could impact market expectations for future capital expenditures.

However, 'slowing down AI' does not equate to stopping investment or halting model research. Amodei’s concept primarily involves reducing the speed of frontier model capability improvements and maintaining a larger time gap between capability and safety mechanisms. Related safety measures may include pre-release model testing, independent third-party evaluations, shared safety standards among companies, and coordination between governments and the industry.

Not all companies in the AI industry support slowing development through industry coordination. NVIDIA (NVDA-US) CEO Jensen Huang recently stated he does not believe new laws or regulatory measures are needed to limit the pace of AI development, arguing that companies should independently judge whether their products are sufficiently safe before releasing them. Meta Platforms CEO Mark Zuckerberg also said that AI labs already have sufficient market and legal incentives to ensure safety, so the entire industry does not necessarily need to slow down in unison.

This divergence extends to the question of 'who should be responsible for AI safety.' One view holds that frontier AI capabilities now involve risks that could transcend individual companies, necessitating third-party evaluations, shared standards, or even government oversight. The other view argues that companies themselves bear product liability and market competition pressure and should therefore decide their own development and release pace. The main difference between the two is not whether safety measures are needed, but rather the degree of coordination and the role of regulation.

Moreover, recent warnings from researchers have extended this discussion from corporate policy to the nature of AI risks themselves. After resigning, Anthropic researcher Jacob Coxon publicly expressed concern that rapid frontier AI development could lead to severe consequences. Former Google DeepMind research engineer Bilal Chughtai also issued a public warning on September 15 about the extreme consequences AI could cause. Evan Hubinger, Anthropic’s head of alignment science, previously estimated that the probability of a catastrophic AI outcome within the next 10 years could exceed 10%. These figures represent individual researchers’ risk assessments, not proven predictions.

Meanwhile, Reuters notes that