Wednesday, July 15, 2026

In today's technological signals, the interaction between AI and financial capital is the core thread. Behind the collapse of the South Korean stock market, the premium of SK Hynix ADR, and the collapse of retail investors, this series of chaos is actually the impact of the new AI-driven financial hype model and capital flows. Although it seems that retail investors have been beaten down and SK Hynix has soared on the surface, it is actually the result of AI quantification and accelerated transactions by agents. Individual investors are still using the traditional "K-line + news" approach, and other hedge funds have already been put on the automated model to arbitrate, arbitrate, and re-arbitrate. Do you think you and AI are on the same starting line? No, AI has directly pulled the trading threshold to the institutional level, leaving retail investors almost nothing to play. The collapse of the South Korean market is a classic scene where retail investors are harvested by agents. Let alone "how long can it last", in fact, there is no chance to carry it at all-AI accelerates the financial cycle, and retail investors 'strategies have become history.

Let's look at the wealth differentiation in the AI industry chain. Liang Wenfeng, founder of DeepSeek, has directly become the new richest man, and for the first time, China's industrial chain has surpassed OpenAI on the global model rich list. Stop using the old mentality of "AI entrepreneurship must be done slowly". Under this wave of big models, winners are not slowly accumulating, but quickly forming groups, rapid financing, and taking off with one click of capital operation. 36 people became billionaires through big models. This is not a dream, it is a reality. The AI industry has entered the giant track. Want to rely on personal struggle and brick-moving technology to make a difference? Wake up, the thresholds and resource barriers for the big model are there. You have technology, no resources, or capital, and you can basically only work for these new rich people. Therefore, for ordinary developers or entrepreneurs, when entering AI now, they must either try their best to hold on to their thighs or honestly engage in generic applications. Don't imagine "fighting a counterattack giant alone."

At the same time, AI's penetration into traditional industries has become more intense. OpenAI's GPT-5.6 voice capabilities have exploded, Claude for Teachers has been launched, ByteDance has automated driving, and AIPC notebooks have blossomed everywhere... You will find that more and more AI models and tools are directly replacing manpower and changing the structure of the industry. Education technology company Stride was directly slapped in the face by Anthropic, and its stock price fell to the ground. This is no longer "AI-assisted", but AI directly grabbing the job. Companies that can understand trends are frantically signing self-discipline conventions and practicing regulations. In fact, they are afraid of policies and being backfired by AI. The penetration of this kind of large model + agent will definitely eliminate a large number of mid-waist companies in the next few years, leaving either giants or "ecologists" who can coexist with AI.

In terms of risk points, the biggest pitfall is that technology and capital are harvesting at an accelerated pace, and ordinary people and small and medium-sized enterprises have no time to breathe. Intelligent stock trading, AI entrepreneurship, autonomous driving, AI office... every track is accelerating its shuffling. Are you still waiting for "policy implementation" and "market maturity"? The tycoons have already gotten on the bus. The only way to survive is: Don't wait, quickly learn some AI automation skills, and jump to the "AI+ industry" to implement applications, otherwise you will only be systematically eliminated. Don't imagine that AI is a tool. It has become a new type of production relationship. Whoever can control AI is the new upper level.

In the final analysis, the current wave of technology is: AI models and capital are working together to accelerate the reshuffle, giants eat meat, retail investors drink soup, and ordinary developers and small companies are either embracing quickly or waiting to be harvested. This wave is not a slow cow, but a high-speed whirlwind. If you step on it, you will have the wind outlet, and if you don't keep up, you will have a pit.

Among today's signals, the most alarming thing is a set of paradoxes: While AI is being held to the altar, its costs are being fully exposed at three levels: finance, technology, and social trust. South Korea's Composite Index plunged from 9200 points to 6800 points in one month, and 1.2 million retail investors were forced to close. At the same time, SK Hynix ADR's premium on U.S. stocks was as high as 46%. It is not a coincidence that this set of data is put together, but that the valuation logic of global capital on the AI industry chain has been seriously divided. The South Korean market is a barometer of chip exports, and SK Hynix is a core supplier of HBM memory. However, the explosion of South Korean retail investors shows that the market's expectations for AI hardware demand have been excessively ahead of actual profits. The ADR premium suggests that the U.S. stock market is still paying for the AI hardware story, but the premium itself is a bubble signal. In addition, Liang Wenfeng has become the new richest man in AI with DeepSeek's valuation, and 36 people have become billionaires with big models. At the bottom of this wealth chain is the cruel reality that GPT-5.6 is called the "Token Assassin" by developers. The actual 3D web development task took nearly 3 hours, and users began to abandon the Skills function, indicating that the marginal cost of AI is rising rapidly and the company's profit model is far from complete. When the underlying technology has not solved the cost problem, the craze of the secondary market is like a castle on the beach.

Another signal worthy of attention is that the crisis of trust in AI is spreading from entertainment to productive scenarios. The first prize in the Hohhot Photography Competition was found to be generated by AI, and the garbled text on the sanitation workers 'clothes exposed the image fraud. This is not an isolated incident. It contrasts with the news that Claude for Teachers entered the education market and that China mathematicians won the Fields Medal for the first time, but the award list was leaked in advance. The stronger AI's capabilities in the field of knowledge production, the deeper society's anxiety about "what is real." When teachers can use Claude to prepare lessons for free and students can use AI to complete homework, the judgment standard itself will collapse. The popularity of the article "Your app could have been a webpage" on Hacker News just shows that the developer community has begun to reflect on the stupidity of "excessive AI." People no longer blindly worship AI, but begin to ask: How much real efficiency does it bring, or does it create more noise and illusions?

The third trend is that AI agents are moving from concept to infrastructure, but the path is steeper than imagined. ByteDance quietly deployed autonomous driving through the Seed World Model team. Anthropic launched the teacher version of Claude directly impacted the share price of education technology company Stride. OpenAI canceled Codex's 5-hour quota limit but overturned resource consumption. These events point to the same underlying logic: the implementation of AI Agents requires complete infrastructure support, including browsers, office software, and financial channels. The OfficeCLI released today specializes in reading and writing Word and Excel for AI Agents, Claude Video allows models to "watch" videos, and Tilion is making the Agent's browser infrastructure. These products are building a proxy operating system that is parallel to the human operating system. But the problem is that when the consumption of AI Agents far exceeds expectations, when educational institutions are crowded with free tools, and when autonomous driving has not yet passed legal and ethical barriers, early builders of this new infrastructure will face huge cost pressures. Half a year ago, we were still discussing whether AI could replace humans; half a year later, the core of the issue has become "Can AI support itself?"

South Korea's Composite Index plunged 26% in a month, and 1.2 million retail investors sold out. This figure is reminiscent of Bill Hwang's Archegos sold out in 2021, or earlier Long-Term Capital Management. Every time a bubble bursts, retail investors are always forced out at the last minute. The question is not "why are South Korean retail investors easily beaten?" but "Who is pushing retail investors to leverage and buy AI concept stocks?" South Korean retail investors have always been known for their high leverage, but this time they are not buying semiconductors, but "the belief of the AI bull market." SK Hynix ADR has a 46% premium on U.S. stocks, a figure that in itself shows that the market has deviated from fundamentals_the same company, another market transaction can cost half as much, this is not rational pricing, this is emotion-driven. When South Korean retail investors burst positions, SK Hynix ADR premium narrowed, and outstanding funds began to emphasize "anchor industry trends" rather than "AI themes", the correction of this round of AI concept stocks may have just begun.

On the other hand, GPT-5.6 was criticized by developers as "Token Assassin", which is very interesting. After OpenAI lifted Codex's 5-hour limit, it took nearly 3 hours to develop 3D web pages, and users found that "the quota consumption skyrocketed." In other words, OpenAI uses "unlimited" to attract users, but actual cost control is not done well at all. This is not a technical issue, it is a business model issue-if each user consumes 3 hours of computing resources to run a 3D task, OpenAI's gross profit margin will quickly be eroded. Historically, similar situations have occurred in the cloud computing bubble. AWS and Azure also relied on burning money to seize the market in the early days, but that was based on the scale effect of infrastructure. However, the cost of reasoning for large models has not dropped as fast as the growth of GPU computing power. Narratives that claim that "AI costs will quickly approach zero" are being slapped in the face by actual usage data of GPT-5.6.

Looking at the first prize work of the Hohhot Photography Competition, the organizer said it was "too fake." On the surface, this incident is a mistake, but it reflects a deeper problem: AI-generated content is permeating into various fields, but review mechanisms, regulatory rules, and even the most basic identification capabilities have not kept up. When AI-generated pictures can deceive judges of photography competitions, it means that the trust crisis caused by AI content has begun. This is not just a problem for the photography industry. The news industry, academic research, and legal documents will all face the same dilemma. The Ministry of Human Resources and Social Security mentioned the "Self-discipline Convention on the Protection of Personal Information by Agents" and the "AI Ecological Code" in the plan, but the problem is that these self-discipline conventions may not even be able to achieve "post-event remedies" in the face of the proliferation of AI content. Historical experience tells us that every large-scale popularization of technical tools will first go through a period of chaos, and supervision will not really intervene until a major accident occurs. AI is no exception.

Finally, if you add up the three incidents of South Korean retail investors, the Token consumption problem of GPT-5.6, and the proliferation of AI content, what I see is not the golden age of AI, but a precursor to the beginning of the bubble to burst. Capital markets are ebbing, technology costs are exposed, and trust mechanisms are collapsing. Those narratives that claim that "AI will replace everything" are being tested by reality-not that technology is not working, but that technology is overvalued too quickly, and the gap between investment and use is much wider than imagined. Today, there are still people rushing to buy SK Hynix ADR at a premium of 46%. Tomorrow, someone may ask,"Why did I spend 3 hours and burned hundreds of Tokens to only generate a semi-finished 3D webpage?" The real danger of a bubble is not the moment it bursts, but that before it bursts, everyone thinks they are the exception.