Altman discusses next-generation models: RSI will delay IPO; some new AI clouds are "unsustainable folly".

OpenAI CEO Sam Altman warned that Recursive Self-Improvement (RSI) may arrive sooner than expected, and this judgment will profoundly affect the direction of capital expenditure in the AI industry, the commercial priorities of cutting-edge laboratories, and even OpenAI's own timeline for going public.

In an in-depth podcast interview, Altman revealed that OpenAI has paused a cutting-edge reinforcement learning (RL) training task and reallocated a significant amount of computing power to AI alignment and security monitoring. The reason given was not a technical bottleneck, but quite the opposite – "The pace of improvement in model capabilities is awe-inspiring, and security alignment work needs time to catch up." Altman emphasized that if the Reinforcement Simulation Index (RSI) occurs earlier than expected, he would prefer to postpone the IPO – because at critical moments requiring short-term revenue-impacting decisions such as "pausing training," the stock price pressure on listed companies would fundamentally conflict with their mission. At the same time, he warned that the current AI industry is seeing a surge of "randomly emerging new cloud vendors claiming to build massive computing power without revenue support," a phenomenon he considers "unsustainable folly." Taken together, these two statements convey a clear market message: in the eyes of cutting-edge labs, the core variable determining computing power investment has never been the commercialization pace of a particular application, but rather whether the model capabilities themselves are still on a worthwhile expansion track.

This logic is gaining validation from both the investment and research communities. Renowned tech investor Gavin Baker recently stated explicitly that, based on a firm belief in scaling laws, top large-scale model companies will not pursue free cash flow in the short term; instead, they will use all of their operating cash flow to purchase more GPUs. Meanwhile, according to Sarah Guo, founder of AI venture capital firm Conviction, over the past 12 months, an increasing number of top researchers have begun to believe that once AI research models with "recursive self-improvement capabilities" emerge, humanity may be only one to two years away from some form of "exponential intelligence."

RSI: Redefining the core variables of CapEx

A prevailing narrative in the market is that AI capital expenditure (CapEx) will peak around 2028, based on the premise that, apart from coding, AI has yet to find its next trillion-dollar application. This logic implicitly assumes that CapEx should be unilaterally determined by commercialization needs.

However, the logic behind Frontier Lab's actions has never been so simple.

OpenAI, Anthropic, xAI, Meta, Google, and SSI (Safe Superintelligence Inc.), founded by Ilya Sutskever, have recently signaled their intention to expand their training infrastructure almost simultaneously. SSI recently announced a strategic partnership with NVIDIA, promising a tenfold increase in computing power over the next 12 months, and explicitly stated that "research has entered a new phase where it is worth scaling." This statement carries far more weight than a typical funding announcement— it signifies that scaling laws remain valid in the eyes of leading researchers, and that increasing computing power can still bring revolutionary improvements in model capabilities.

RSI's logic is redefining training. Previously, training was a linear process of "collecting data → training → publishing → ending." However, within the RSI framework, models can generate new algorithms, optimize training processes, and iterate continuously, forming a cycle of "model A → training model B → optimizing model B's training process → training model C…". Training has transformed from a periodic event into a continuous process, resulting in an order-of-magnitude increase in computing power requirements—laboratories with sufficient GPUs can simultaneously validate thousands of training schemes, directly translating computing power advantages into research and model advantages.

In the interview, Altman revealed that OpenAI launched the "Automated AI Researcher" framework, Anthropic heavily involved Claude in model development, and Google's AlphaEvolve has also begun using AI to find new algorithms. These all belong to automated training, which is an early form of RSI.

This switch has a structural impact on computing power requirements. Under the RSI framework, validating a new training scheme may require running hundreds or thousands of versions simultaneously, ultimately retaining only the optimal solution. Labs with ample computing power can validate tens of thousands of schemes a day, while labs with insufficient computing power can only validate a few—the computational advantage thus directly translates into a research advantage, and further into a model advantage .

The dual sense of powerlessness of cutting-edge researchers

The intensity of this computing power race is quietly changing the mindset of the cutting-edge AI research community.

In an interview, venture capitalist Sarah Guo described a thought-provoking phenomenon: as model training budgets approach hundreds of billions of dollars and teams expand to thousands of people, more and more top researchers are experiencing two negative sentiments: "What I do is unimportant because the model will soon be able to do it itself," and "The only thing that matters is the scale of computing power; individual contributions are diluted."

Sarah Guo believes that over the past 12 months, a growing number of top researchers have come to believe that once AI research models with recursive self-improvement capabilities emerge, some form of "exponential intelligence" is only one to two years away. This belief itself is not new, but its rapid spread and growing acceptance within the research community are intensifying competition and uncertainty within the industry.

Ultraman "steps on the brakes": because he was running too fast, not too slow.

In the interview, Altman specifically clarified a misunderstanding: OpenAI slowed down training not because it encountered a bottleneck, but because the speed at which the new model's capabilities improved exceeded the coverage of the current alignment and security system.

He described the context that triggered this decision: Following the Hugging Face security incident, OpenAI observed a series of model behaviors during training that were "not entirely consistent with expectations." This, coupled with predictions about the upcoming more powerful pre-trained models, prompted management to proactively apply the brakes. "When you actually experience all of this, you realize: this is the moment we've been discussing for so long, and now it's here," Altman described the decision-making process.

It's worth noting that Altman isn't anxious about the business side. He stated that even without releasing new models, the existing models are sufficient to support products and revenue, and that corporate revenue currently exceeds consumer revenue, growing extremely rapidly. This assessment provides the confidence for his decision to "put safety above momentum."

"Unsustainable folly": Ultraman draws a line.

On the issue of computing power investment, Altman drew a clear line between OpenAI and other industry players. He stated that he was not worried about OpenAI's own computing power plans, but rather about the global computing power construction boom, especially those "randomly emerging new cloud vendors claiming to build massive computing power without revenue support." He called this phenomenon a sign of "unsustainable folly."

Gavin Baker's assessment contrasts this view: he believes that top-tier large-scale model companies, based on their firm belief in scaling laws, will not focus on generating free cash flow, but rather will invest all of their operating cash flow in purchasing more GPUs. This means that for cutting-edge labs, sacrificing short-term inference revenue and pouring computing power into training next-generation models is a rational rather than reckless choice. The difference lies in the fact that the high CapEx investment of leading labs is supported by technological logic and revenue, while a lack of basic computing power accumulation is another matter entirely.

Conditions for IPO Delay: The Speed of RSI Arrival

When asked about IPO plans, Altman unusually linked the company's governance structure directly to its technological roadmap. He stated that becoming a publicly traded company would subject the company to immense pressure on its stock price should it need to make short-term revenue-impacting decisions, such as suspending training. He initially thought the arrival of superintelligence was still a long way off, but now believes it could happen very soon.

"The mission is far more important than going public." This statement clearly expresses Altman's current priorities: if the RSI arrives as expected, maintaining absolute priority on security under the constraints of a listed company will be much more difficult than imagined, making postponing the IPO a more prudent choice.

This stance aligns with Altman's repositioning of the "superintelligence" concept. In the interview, he stated that "AGI" has become a vaguely defined marketing term, and OpenAI's internal focus has shifted to infinitely scalable superintelligence. As the pace and uncertainty of technological evolution intensify simultaneously, maintaining flexibility regarding the IPO timeline may be one way the company is addressing the "future that has already arrived."

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.