What Truly Determines 2028 Capex Is Not Coding, But RSI

It is evident that the following narrative is gaining increasing influence, logically profoundly affecting and even dominating the operations of the US, Chinese, South Korean, and Japanese markets:

Apart from Coding, AI has not yet found the next trillion-dollar application; therefore, around 2028, the growth rate of AI capital expenditure (Capex) is highly likely to peak.

The market is always right, and this logic appears very smooth:

AI companies have burned hundreds of billions of dollars, and the only major commercial direction that has truly succeeded so far is Coding; if no new applications are seen generating revenue, the market no longer wants to wait. The market believes that cloud providers and Frontier Labs will naturally cut back on investment, and the AI infrastructure cycle will end, with the timing being 2028.

However, just a few days ago, an announcement jointly made by NVIDIA and SSI (Safe Superintelligence Inc.) precisely illustrates:

This narrative may have misidentified the true variable.

For context, SSI was founded in June 2024 by Ilya Sutskever, co-founder and former Chief Scientist of OpenAI. The company states that its "sole goal and sole product" is safe superintelligence (SSI), making this its mission, name, and entire product roadmap.

In this announcement, SSI declared:

NVIDIA will make a substantial investment in SSI, and the two parties will establish a long-term strategic partnership; within the next 12 months, SSI's compute power will increase tenfold (10× Compute); SSI believes its research has entered a new phase that is "worth scaling."

Many people merely interpret the above event as just another round of financing. However, when viewed within the broader development context of Frontier Labs, it is actually a key industrial signal:

What truly drives the next round of Capex may not be just new AI applications, but a return to training, specifically training at a certain critical point.

Do you remember?

In last week's viral meeting minutes, there was no concern for AI applications. The exact words were: "Because we can see that there is really a bigger watermelon ahead, while what's in front is really just some small sesame seeds... because the AGI opportunity ahead should be very large, the AGI opportunity ahead is always very large..."

Therefore, the narrative that "apart from Coding, AI has not yet found the next trillion-dollar application" may have been incomplete from the start. The reason the market believes Capex will peak in 2028 is essentially based on an assumption:

Capex should be determined solely by commercial demand.

However, you must understand that Frontier Labs in the AI era are never ordinary companies! OpenAI, Anthropic, DeepMind, SSI, DeepSeek, and others are companies representing humanity in the pursuit of higher human intelligence.

The biggest competition among these Labs today is not who produces the second GitHub Copilot first, but who first creates the next-generation leading model that, in SSI's terminology, is "worth scaling."

In other words, the primary purpose of Frontier Labs purchasing GPUs and storage today is not to meet existing Inference demand, but to revolutionarily enhance the capabilities of their future models once again.

Those meeting minutes also showed that for Frontier Labs, the primary variable determining Capex has never been any specific application sector, but whether the next-generation model can continue to revolutionarily improve its capabilities.

So, why is NVIDIA investing in SSI?

The most noteworthy sentence in the SSI announcement is actually not the tenfold expansion of compute power, but:

Our research is worth scaling.

You should understand the weight of this statement.

If Scaling had become ineffective, if adding more GPUs could no longer revolutionarily improve model capabilities, then no Lab would continue to invest billions of dollars to expand compute capacity. This point is more worthy of attention than any market opinion.

More importantly, SSI's announcement of large-scale training expansion is not an isolated case. Over the past six months, almost all Frontier Labs have released highly consistent signals.

• OpenAI continues to expand its training clusters.

• Anthropic continues to raise funds to build new AI infrastructure.

• xAI is constantly expanding Colossus.

• Meta continues to build GW-level AI campuses.

• Google continues to expand TPU deployment.

• Now SSI announces a tenfold increase in compute power over the next year.

These events share a common characteristic:

None of the companies' actions reflect that "training is over"; on the contrary, they are all increasing training capabilities.

In fact, one of the most discussed directions among Frontier Labs now is:

Recursive Self-Improvement (RSI)

If we divide the technical roadmap of Frontier Labs over the past two years into several stages, it can be roughly understood as follows:

• 2024–2025: Pretraining → Reasoning

• 2025–2026: Reasoning → AI Agent → Monetization

• Now (starting from the second half of 2026): Reasoning has not ended, but has begun to serve the next larger Training cycle, with the goal being RSI (Recursive Self-Improvement).

There was a misconception in the past that the era of reasoning meant a decline in the importance of Training?

The answer is now becoming increasingly clear: No.

The reason is that previously, Training involved humans training models, whereas in the future, Training will become AI training AI, which represents a completely different order of magnitude in requirements for AI infrastructure. Therefore, OpenAI proposed the Automated AI Researcher, Anthropic has heavily involved Claude in model development, and Google's AlphaEvolve has already begun using AI to find new algorithms.

These all belong to Training Automation, which is RSI.

In the past, training a GPT model followed the process of: collect data → train → release → end. However, RSI is:

Model A → generates new algorithm → trains Model B → Model B continues to optimize the training process → trains Model C → Model C continues to discover better RL strategies → ... → ...

The entire process repeats itself, turning the training process into:

Continuous Training, not training once every month or two, but training continuously, with training never stopping.

For example, to verify a new algorithm, previously it involved training once, whereas now it might involve running 100 versions, 1,000 versions, or 10,000 versions simultaneously, keeping only the best one and discarding the rest.

The reason for doing this is simple.

If AI can propose 100 new training schemes daily, then only Frontier Labs with sufficient GPUs can complete all validations within a day or two; whereas Frontier Labs lacking GPUs can only complete 5 or 10 validations a day, immediately widening the speed gap. Thus:

Compute Advantage becomes Research Advantage, and finally becomes Model Advantage. Is this not a dimensional strike?

After entering the RSI stage, the concept of cluster sizes with millions of cards or more is not an exaggeration. Because continuous training and automated experiments both require massive amounts of Training Compute.

This will significantly boost demand for ultra-large-scale, high-bandwidth training clusters, creating greater rigid demand for the Rubin and Rubin Ultra generations of products than during the Blackwell era.

Why is the SSI event worthy of attention?

Because it clearly tells the market for the first time:

Large model research has entered the "Worth Scaling" phase, which is a critical point.

In NVIDIA's view, the most important customers in the future are not just those buying GPUs, but the Frontier Labs that will continue to conduct ultra-large-scale training.

Of course, this does not mean that AI commercialization is unimportant.

However, directly deducing that "Capex will inevitably peak in 2028" from the fact that "there are temporarily no new applications of comparable scale outside of Coding" misses a key premise:

Frontier model capabilities have stopped improving, and Frontier Labs no longer believe that expanding training is worth continued investment.

Judging from currently public facts, this premise has not been verified at all.

On the contrary, what we see is that almost all Frontier Labs are continuing to expand their training infrastructure. These facts indicate that the main participants in the industry chain still believe that larger training scales, more frequent model iterations, and continuous R&D investment still hold revolutionary value.

It still comes down to that saying:

"Because we can see that there is really a bigger watermelon ahead, while what's in front is really just some small sesame seeds."

Source: Shaoshupai Viewpoint Bureau

Risk Warning and Disclaimer

The market carries risks; investment requires caution. This article does not constitute personal investment advice, nor does it take into account the specific investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment decisions made based on this content are the sole responsibility of the investor.