Three and a half years after the outbreak of Generative AI, the market has reached a new inflection point: optimism is still accelerating, while skepticism is also accumulating. Judging whether a "bubble" has arrived is not enough to explain the current complexity. The "AI Faith and Bubble" series will explore key variables from different perspectives of the market, technology, industry, and companies. This article is the first in this series.
On June 9, the South Korean KOSPI index staged a sharp rebound, with an intraday gain of nearly 5%; KOSPI 200 futures surged by 5%, triggering a sidecar buy, and program trading was suspended for 5 minutes. The previous trading day (June 8), the KOSPI had once fallen by over 8%, dropping below 8000 points.
Over the past two years, South Korea has been one of the most AI-sensitive amplifiers in global trading: stocks like NVIDIA and HBM surged; AI server production expanded, benefiting SK Hynix; as storage prices rose, the valuation logic of Samsung and Micron was rewritten. It has both embraced the imagination of global AI infrastructure expansion and absorbed doubts about whether this expansion is overheated.
Therefore, the South Korean stock market has repeatedly triggered trading cooling-off mechanisms between rising and falling, reflecting the escalating global capital market divergence on AI.
On one hand, AI remains the most certain investment theme. From chips, storage, cloud computing to large-scale model companies, almost all core assets have been reintegrated into the valuation framework of the "AI infrastructure."
As long as the demand for computing power continues to grow, today's capital expenditures, supply chain price hikes, and high valuations can all be explained as upfront investments for future growth.
On the other hand, skepticism is also accumulating.
AI is becoming increasingly expensive. The capital expenditures of tech giants continue to rise, valuations of large-scale model companies continue to soar, and AI upstarts are queuing up for IPOs.
Three and a half years into the Generative AI boom, serious discussions about a bubble have gone through three rounds. Each round had clear trigger events, a coherent logical chain, and seemingly fatal doubts, yet each time the market found new hope from the rifts and continued to bet big.
This time, judging by the market's response, we are already at the center of the fourth round of divergence.
Several investors have indicated that "it is far from time to discuss a bubble now. Although it has begun to erode the cash flow of giants, the commitment of the giants is still very strong, accelerating. When you see the slowdown in giant investments, that is when you should be on high alert."
Zhang Yidong, Member of the Executive Committee, Head of Research, and Chief Economist of Haitong International, said: "This AI wave is even more intense than the Internet wave from 1993 to 2000. In the tide of the AI era, there is no division between high and low, only diffusion."
This is the core contradiction in today's AI market: everyone knows that prices are getting higher, everyone believes that the growth rate will smooth out the valuation bubble, and no one dares to get off the ride.
01 Two and a Half Years, Three "Bubble Theories," and a Dangerous Consensus
The AI bubble theory has lasted for two years, with each "bubble theory" corresponding to a paradigm leap in the AI industry and the capital frenzy and belief shake-up surrounding this leap.
The first time was in June 2024. Sequoia Capital published the renowned "AI's $600B Question," the first time questioning the massive capital expenditure. Sequoia posed the question: based on the revenue from NVIDIA's data centers and the total GPU ownership cost at that time, the AI industry would need about $600 billion in annual revenue to support this round of infrastructure investment.
The AI paradigm at that time was the Pre-training Scaling Law: the bigger the model, the better; the more data, the better; the more GPUs, the better.
From early 2024 to Sequoia's questioning, Super Micro rose by 217%, and NVIDIA rose by 150% during this frenzy. The entire market's belief anchor was a simple equation: AI = computing power = NVIDIA.
The questioning by Sequoia lasted for less than three months.
In September 2024, OpenAI released o1, the reasoning-time computing paradigm emerged, not relying on larger models but on longer thinking, using post-training + RL to lift the ceiling of model capabilities. A new AI capability growth curve opened up, and the market saw the second growth pole of computing power demand.
However, the new paradigm quickly developed new cracks—DeepSeek R1 was released, pushing the efficiency of reasoning-time computing to the extreme: achieving close to cutting-edge model inference capabilities with less than $6 million in training costs.
On January 27, 2025, NVIDIA's market capitalization evaporated by $593 billion in a single day. The second bubble theory began to emerge.
The core of the market's doubt was: Does achieving equivalent AI capability really require this much computing power? This round of panic came on strong but was resolved more quickly. One month later, NVIDIA released its financial report, with Blackwell's quarterly revenue of $11 billion far exceeding expectations. The market used performance to prove that the new inference paradigm would create more inference demand, leading to a total increase in computing power requirements.
Driving the inference paradigm, OpenAI became the absolute center of this wave, with anyone signing with it experiencing a surge. CoreWeave completed an IPO with a $11.9 billion five-year contract, Oracle set a new high with a $300 billion agreement through the "Stargate Plan," and Broadcom secured a billion-dollar custom chip order.
This time, the market's confidence wavered the least. OpenAI mass-produced "concept stocks," and the belief anchor shifted from "training arms race" to "large-scale deployment of inference."
The third time was in October to November 2025.
Goldman Sachs released a report listing five bubble signs for AI: CapEx peaking, slowing enterprise profit growth, rising tech company debt, the start of a Fed rate cut cycle, and widening credit spreads, explicitly drawing an analogy to the eve of the 1997 internet bubble. Bank of America fund managers' survey saw the first-ever "overinvestment" judgment in 20 years. Both Wired and The Atlantic published in-depth investigations in the same week, pointing to the same finding: 95% of enterprise AI investments had not yielded tangible returns.
The narrative of massive AI investment reaching a climax in the AI industry chain was pushed by NVIDIA's revenue from cloud providers, the cloud providers' AI revenue growth from model company expansion, the model companies' valuation from investors, and the investors' returns from the reassessment of model companies on paper.
But who is actually paying for AI upstream?
In the Q3 2025 earnings calls, nearly all major U.S. stock giants responded to analysts' inquiries with almost the same sentence, also countering Goldman Sachs' CapEx peaking judgment: "We'd rather overinvest than lose the future." Goldman Sachs left a small tail in the report, suggesting the present is more like 1997 than 1999, implying that signs of a bubble had appeared but a burst was still a distance away.
In November 2025, the Fed cut interest rates by another 25 basis points, and liquidity continued to support valuations. The Nasdaq hit new highs amidst skepticism. The market's consensus became: I know there may be a bubble, but getting off now is more dangerous than staying on the ride.
But what truly resolved this round of skepticism was the advent of a new paradigm.
In the second half of 2025, Agentic AI centralization erupted, transforming AI from dialogue to autonomous digital employees capable of self-planning, execution, and iteration. AI directly replaced workflows, shifting the income ceiling from "search replacement" to "labor replacement." More importantly, Agents naturally consumed tens of times more tokens than dialogue, and the demand for computing power not only did not diminish but rather opened up an order-of-magnitude growth space.
Looking back at the three waves of the "bubble theory," three clues have always run through.
How long will the demand for computing power last; who will bear the enormous AI expenses and their return on investment; and whether large models will usher in a new paradigm breakthrough.
After the three waves, the market formed a dangerous consensus: "Doubts will always be quickly refuted."
Behavioral finance experience shows that investors who re-enter the market after experiencing panic often have a higher risk appetite because they have already "validated" that the panic was wrong.
02 The Fourth AI Bubble and Three Rifts
Over the past few trading days, the market has been violently shaking.
The first rift appeared between the "profit" and "cash flow" of tech giants.
In the 2026 Q1 earnings season, the giants hit record profits while their cash flows approached zero. Amazon's free cash flow plummeted by 95% year-on-year, and the big four tech giants collectively burned $20 billion a day. Some giants' "net profit growth" was interpreted as a paper revaluation of the investment in AI companies, using the returns from investing in AI to prove that investing in AI is right, circular reasoning.
In April, Goldman Sachs provided a figure: about 40% of the expected 2026 S&P 500 earnings growth comes from the industrial chain transmission effect of AI-related capital investment. Several investment banks and media estimates show that the AI-related capital expenditures of the several major hyperscale cloud providers in 2026 have already reached the range of $600 billion to over $700 billion.
Some media commented: "Silicon Valley tech giants are left with only profits." This also means that the overall growth expectations of US listed companies are built on the same building block, where AI CapEx can only increase, not decrease, pulling the whole system along.
Morgan Stanley pointed out: By 2026, the CapEx-to-revenue ratio of large-scale enterprises will reach 34%, rising to 37% by 2028, formally surpassing the historical peak of 32% during the 2000 Internet bubble. Over the three years from 2026 to 2028, the total AI infrastructure expenditure of just the top five giants will reach $20 trillion.
More covertly, the five companies also have nearly $1 trillion in off-balance-sheet lease commitments, which are long-term contracts for data centers that have not yet been completed and do not appear on any balance sheet.
Global AI usage is skyrocketing, with many companies proclaiming "tokenmaxing." Fomo sentiment is spreading, CEOs are afraid of being left behind by AI, and are rushing to swipe their cards to "tokenize" their employees.
However, in the "tokenmaxing" movement, a large part of the consumption comes from systemic redundancy within the Agent architecture, over-engineered Harnesses, leading to a massive bubble in large model token usage. No institution has yet disaggregated the ratio of "effective computation" and "architectural idling."
According to foreign media reports, Uber burned through its entire annual AI coding budget in the first four months of 2026.
Engineers, spurred by the call for tokenmaxing, are starting to use tools like Claude Code as parallel labor: running multiple tasks simultaneously, opening multiple worktrees at the same time, Agents searching, generating, error-reporting, and fixing autonomously for extended periods. AI usage seems to be increasing, but the finance department finds it challenging to immediately assess how much quantifiable output these tokens ultimately bring.
Usage volume is a key metric for model company valuations, but if this metric itself is bloated, how reliable is the trillion-dollar valuation built on top of it?
Whether the Agentic paradigm is truly effective in enhancing corporate productivity is the market's second fault line.
The capital markets frenzy continues, with Anthropic reportedly nearing a $965 billion valuation and having secretly filed for an IPO; OpenAI has also secretly filed for an IPO, with a valuation of $852 billion after its latest funding round.
Clearly, the market is paying full price for an unrealized future. This does not necessarily mean a crash, but it may indicate an extremely narrow margin of error.
"All great technological revolutions create bubbles. No one can judge with complete accuracy. You either invest a lot of money to seize market share without worrying about spending too much, or you invest too little and then lose market share." On June 3, Ray Dalio, founder of Bridgewater Associates, said in an interview.
Dalio believes that a bubble burst occurs when investors try to convert paper wealth into cash, and the current AI-driven market is "progressing along this path," even though AI itself "is a fantastic technology."
From this perspective, the long-term value of the technology can coexist with a short-term valuation bubble, just as the Internet itself underwent a profound reshaping of the global economy after the bursting of the Internet bubble.
Discussing the bubble of Generative AI from a macro perspective is a topic the media loves, but it is not an "effective topic" in the eyes of investors.
Peter Thiel believes that "AI technology is real, but the market has already priced in the next 15-20 years." He fully exited his entire Nvidia position in Q3 2025, a $100 million position accounting for 40% of the fund, while cutting 76% of Tesla, reducing the total position by 65%. He accurately predicted the Internet bubble in 1999; is this prediction different? There is no answer at the moment.
But what is certain is that Peter Thiel missed out on the Agent New Paradigm frenzy that began at the end of 2025.
It's not just Thiel. Berkshire Hathaway's Q1 2026 report shows that Buffett's cash position has ballooned to $397.4 billion — a historical high, accounting for 59% of total assets.
The long-term trend of technological evolution moving in a positive direction does not mean that investors will not take profit-taking or deleveraging actions in the short term. The contradiction between the long-term trend and short-term investment strategy is the third crack in the market.
Above these three cracks, the market's nerves have begun to become sensitive. When U.S. interest rate expectations rise and the market begins to question whether AI capital spending is overheated, Korea, as one of the "purest" markets in this AI wave, sees significant volatility. Its rise comes from AI faith, and its sharp drop also stems from a loosening of AI faith.
Investment trading has thus entered a highly challenging range.
Many experienced investors still believe that "the AI bubble has not yet arrived." However, determining whether AI is a bubble or undervalued, and whether it is possible to establish an effective investment system, are two completely different things. The former is a directional judgment, while the latter is a comprehensive test of rhythm, position, valuation, cash flow, and exit window.
In a market where faith is still present and volatility is intensifying, being right about the long-term trend does not necessarily mean that most people can withstand short-term retracements.
The fourth wave of Generative AI bubble may not have arrived yet, but the need to remain vigilant has already come. Experienced, adaptable captains can navigate through the storm to find the treasure, but ordinary sailors may perish in the storm.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
