The RSI era is coming: Google, OpenAI, and SSI are sending the same signal.

A new change is emerging in the AI industry that deserves the attention of the capital market.

Over the past two years, discussions about AI have mainly revolved around several questions: whether model capabilities can continue to improve, whether inference costs can decrease, whether new trillion-dollar applications can emerge beyond coding, and whether there will be an overcapacity in data center construction.

However, with Google releasing Gemini 3.8 Flash, SSI announcing a tenfold increase in computing power over the next 12 months, and OpenAI publicly discussing Recursive Self-Improvement (RSI) and training resource allocation, cutting-edge labs are shifting the focus of competition in another direction:

AI is not only serving users, but also beginning to participate in training, evaluating, and optimizing the next generation of AI.

RSI, or Recursive Self-Improvement, is often translated as recursive self-improvement. In simple terms, it doesn't mean the model suddenly gains "self-awareness," but rather that it involves the model in more aspects of model development: generating algorithms, writing code, designing experiments, using training tools, evaluating results, fixing errors, and feeding back valuable experience to the next round of development.

If this closed loop continues to shorten, the pace of AI research and development may change.

RSI is not a gamble by any single company. Google, OpenAI, SSI, Anthropic—almost all cutting-edge labs are moving in the same direction.

Gemini 3.8 Flash, RSI signal emitted by Google

On September 2, Google officially released Gemini 3.8 Flash. This is the third Flash model released by Google in six weeks.

The performance data itself is impressive enough: in the high inference mode of independent testing agency Artificial Analysis, the 3.8 Flash Intelligence Index reached 59 points, close to top flagship models such as GPT-5.6 Sol and Grok 4.6, with a single task cost of only $0.58 and an input price of $0.75 per million tokens.

But what's even more noteworthy is the qualitative assessment of this model by Yao Shunyu, a researcher at Google DeepMind:

For the model, this is just a small step; but for RSI, it is a giant leap.

This statement reveals what Google is really doing.

An internal Google memo explicitly stated that the goal was to "force recursive self-evolution, transforming programming models into fully automated AI researchers, and completely opening up the entire R&D loop." To this end, Google assembled a code task force led by DeepMind CTO and directly supervised by co-founder Sergey Brin, and recruited Barret Zoph, head of post-training at OpenAI, as VP of Research, responsible for reinforcement learning and post-training.

3.8 The core upgrade of Flash is precisely the product embodiment of this strategy: the model is designed to perform more inference steps on complex tasks, repeatedly call tools, and evaluate and optimize its own results through long-running agent loops. In other words, the model begins to take on the complete process of "planning-execution-checking-repairing" .

It's worth noting that Google initially had several 3.5 Pro candidate models, but all were ultimately rejected—because they weren't significantly better than the Flash series. Meanwhile, the next-generation flagship Gemini 4 was still stuck in the post-training phase. All these unexpected events ironically led to the rapid iteration of the Flash series.

OpenAI's Braking and SSI's Acceleration: The Resonance of Giants

Around the time Google released its new model, OpenAI CEO Sam Altman made a rare statement in a public interview.

He acknowledged that OpenAI delayed a frontier reinforcement learning (RL) training run. This is the first time in OpenAI's history that it has proactively paused frontier training.

The reason was not a single "smoking gun" incident; during training, the team observed that the speed at which the model's capabilities improved was astonishing.

Altman's original words were: "The speed at which capabilities are improving... I can only describe it as 'awe-inspiring.' We need more time for safety, alignment, and security to catch up."

He further stated: "A year ago, I didn't think superintelligence would arrive anytime soon. Now I think it's possible. I'm not saying it definitely will, but we're moving at an extremely fast pace."

Altman also stated that this pause specifically targets cutting-edge RL training runs, not all training has stopped, and the clusters are not idle. He emphasized that OpenAI currently enjoys strong commercial momentum, with enterprise revenue exceeding consumer revenue, and existing models still have significant commercial value to be explored.

But the significance of this statement goes far beyond the commercial level: even OpenAI itself is putting on the brakes and waiting for safety to catch up, which is the most direct public signal that RSI is approaching.

In addition, Nvidia recently announced a major strategic investment in SSI (Safe Superintelligence Inc.), and SSI also announced that its computing power will expand tenfold in the next 12 months.

SSI was founded in June 2024 by Ilya Sutskever, co-founder and former chief scientist of OpenAI, and its "sole goal and sole product" is secure superintelligence.

The most noteworthy sentence in this announcement is:

Our research is worth scaling up.

According to an article on Wall Street News citing "minority viewpoints," a prevailing narrative in the market is that, apart from coding, AI has not yet found its next trillion-dollar application, and therefore, the growth rate of AI capital expenditure is likely to peak around 2028.

However, the SSI incident revealed that this narrative may have focused on the wrong variables.

The primary variable determining the success of the CapEx frontier lab has never been a particular application area, but rather: whether the next generation of models is still worth scaling up the training scale.

Over the past six months, almost all cutting-edge laboratories have released highly consistent signals: OpenAI continues to expand its training cluster, Anthropic continues to raise funds to build AI infrastructure, xAI continues to expand Colossus, Meta continues to build a GW-level AI park, Google continues to expand its TPU deployment, and SSI announced a tenfold expansion of computing power.

No company's actions indicated that "the training is over."

RSI is reshaping the logic of computing power: training is changing from a "one-time" process to an "never-ending" one.

The key to all this may lie in the fact that RSI (Recursive Self-Improvement) is changing the very nature of training.

The traditional training model was: collect data → train → publish → end.

The training model in the RSI era was: Model A generates a new algorithm → train Model B → Model B optimizes the training process → train Model C → Model C discovers a better RL strategy → the cycle continues.

This means that training has become continuous training—it will never stop.

To validate a new algorithm, where previously it was trained once, now it might run 100, 1000, or even 10000 versions simultaneously, ultimately keeping only the best one.

According to an analysis cited by Wall Street Insights, this leads to a direct conclusion: computing power advantage directly translates into research advantage, and research advantage directly translates into model advantage. Labs with sufficient GPUs can complete all verifications in a day, while labs lacking GPUs can only complete 5 or 10 per day, immediately widening the speed gap.

OpenAI introduced the concept of Automated AI Researcher; Anthropic heavily involved Claude in model development; and Google's AlphaEvolve has begun using AI to find new algorithms. These are all real-world applications of RSI.

Gavin Baker's warning: Top labs may voluntarily cut revenue to fund training.

What does this mean for investors?

In a recent public dialogue, renowned technology investor Gavin Baker provided a specific calculation:

Assuming a laboratory has 10GW of computing power, of which 8GW is used for inference, at $60 billion per GW, the annualized revenue would be approximately $480 billion.

Assuming they achieve a major research breakthrough and decide to reduce inference allocation from 8GW to 2GW while expanding training from 2GW to 8GW, their annualized revenue would plummet from $480 billion to $120 billion. I think they might actually make such a decision, and that's something the public market has to get used to.

Gavin Baker also stated that, based on his firm belief in Scaling Laws, top large-scale model companies will not pursue free cash flow in the short term. Instead, they will reinvest all profits and funds into purchasing computing power and model training.

He also pointed out that this is different from internet companies like Meta and Google—whose fundamentals are quite stable, and there is no huge trade-off between infrastructure costs and revenue. Frontier Labs, on the other hand, may at any time sacrifice huge amounts of short-term commercial revenue for long-term technological advantages.

This is a significant risk warning for the valuation of AI stocks.

The mindset of a cutting-edge researcher: the sense of powerlessness brought about by overwhelming computing power.

This competition is also changing the mindset of cutting-edge researchers.

In a video blog interview, Sarah Guo, founder of Conviction and an AI investor, shared her observations of about 250 leading AI builders. Over the past 12 months, more and more top researchers have begun to believe that once AI research models with recursive self-improvement capabilities emerge, humanity may only be one to two years away from exponential intelligence.

However, at the same time, as model training budgets approached hundreds of billions or even trillions of dollars and teams expanded to thousands of people, some top researchers developed two negative sentiments:

What I do is not important, because the model can do it itself very quickly.

The only thing that matters is the scale of computing power; individual contributions are diluted.

This psychological shift itself is also a testament to the growing acceptance of RSI logic within the industry—when researchers begin to feel that "only computing power is important," it precisely demonstrates that the connection between computing power and research capabilities is already close enough.

With RSI approaching, market narratives need updating.

A conclusion with direct implications for investors is emerging:

The market narrative that "AI capital expenditure will peak in 2028" is based on the premise that CapEx is driven unilaterally by the demand for commercial applications.

However, RSI offers a different framework for judgment: what determines future CapEx is not only inference demand and application revenue, but also whether the model's capabilities are still worth continuing to expand training.

As long as cutting-edge laboratories continue to believe that "our research is worth scaling," investment in training will not cease. And judging from all publicly available signals so far, this premise remains unshaken.

Will RSI become a key variable in the next round of AI capital expenditure?

It's too early to draw conclusions now.

RSI has not yet been proven to consistently deliver exponential capability leaps. Improvements in benchmark tests do not equate to continuous, autonomous improvement in real-world development environments. Higher agent capabilities also lead to higher token consumption, more complex systems engineering, and greater security risks.

However, judging from the signals released by Google, SSI and OpenAI, the frontier labs clearly no longer regard RSI as a distant theoretical concept.

Google is trying to enable low-cost models to perform longer, more complex, and more verifiable agent tasks; SSI is using a 10x scaling up of computing power to verify whether "research is worth scaling up on a large scale"; and OpenAI is discussing how to reallocate resources between security, commercialization, and cutting-edge training.

These changes all point to one trend:

The next stage of competition in the large model industry may not just be about "whose model is better at answering questions," but more importantly, "who can enable AI to participate in AI research and development more quickly and complete self-iteration within safe boundaries."

If this trend continues, the core issues of the AI industry chain will also change.

The market should not only ask: Is the revenue from AI applications enough to support the construction of data centers?

Further questions remain: Is the capability curve of cutting-edge models still trending upwards? Can AI truly shorten the R&D cycle? Can training clusters translate computing power into research results? And, can the security system keep pace with the speed of model self-iteration?

This may be the most important impact of RSI on AI capital expenditure, model competition, and technology investment.

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.