Original Title: "Generative AI, No Fourth Bubble?"
Original Author: Xiaojing, Tencent Technology

After three and a half years of explosive growth in generative AI, the market has entered a new point of divergence: optimism is still accelerating, while skepticism is also accumulating. Judging whether a "bubble" has arrived is not enough to explain the current complexity. The series "AI Belief and Bubbles" will explore key variables from different perspectives—market, technology, industry, and companies. This article is the first in that series.

On June 9, the Korean KOSPI index rebounded sharply, with intraday gains approaching 5%; KOSPI 200 futures rose by 5%, triggering a sidecar buy, and program trading paused for 5 minutes. On the previous trading day (June 8), KOSPI had once fallen over 8%, breaking below 8,000 points.

Over the past two years, South Korea has almost been the most sensitive amplifier of global AI trading: Nvidia's rise, HBM's rise; AI server expansion, SK Hynix's rise; storage price hikes, Samsung and Micron's valuations rewritten. It has both carried the imagination of global AI infrastructure expansion and reflected market doubts about whether this expansion is overheating.

Therefore, the Korean stock market repeatedly triggers trading cooling mechanisms between rises and falls, reflecting that global capital markets' divergence on AI is also widening.

Chart: The financing balance of the Korean stock market has risen to a historic high

On one hand, AI remains the most certain investment mainline. From chips, storage, cloud computing to large model companies, almost all core assets are being revalued within the "AI infrastructure" framework.

As long as computing power demand continues to grow, today’s capital expenditure, supply chain price increases, and high valuations can all be interpreted as pre-investments for future growth.

On the other hand, skepticism is also accumulating.

AI is becoming increasingly expensive. Giants’ capital expenditures continue to be revised upward, valuations of large model companies keep rising, and new AI startups are lining up for IPO.

Since the explosion of generative AI three and a half years ago, serious discussions about bubbles have gone through three rounds. Each round has clear trigger events, logical chains, and seemingly fatal doubts, yet each time the market finds new hope from cracks and continues to pour in.

This time, from the market’s reaction, we are already at the center of the fourth divergence.

Many investors say, "It’s still far from the time to talk about a bubble. Although it’s already eating into the cash flow of giants, their investments are still very determined and accelerating. Only when you see giants slowing down their investments should you be cautious."

Zhang Yidong, member of the Executive Committee, head of research, and chief economist at Haitong International, said: "This wave of AI is greater than the internet wave from 1993 to 2000. Under the tide of AI, there is no high or low, only diffusion."

This is the core contradiction of the current AI market: everyone knows prices are getting higher and higher; everyone believes that the growth rate will smooth out valuation bubbles; no one dares to be the first to sell.

01 Two and a Half Years, Three "Bubble Theories" and a Dangerous Consensus

The bubble theory of AI has lasted two years. Each "bubble theory" corresponds to a paradigm shift in the AI industry, as well as capital frenzy and faith shake-ups around that shift.

The first was in June 2024. Sequoia Capital published the famous "AI's $600B Question," raising doubts about huge capital expenditures for the first time. Sequoia’s question was: based on Nvidia’s data center revenue and total GPU ownership costs at the time, the AI industry would need about $600 billion in annual revenue to support this round of infrastructure investment.

At that time, the AI paradigm was the pretraining Scaling Law: bigger models, more data, more GPUs.

From early 2024 until Sequoia’s doubts, during this frenzy, Super Micro’s stock rose 217%, Nvidia rose 150%. The market’s belief was anchored by a simple equation: AI = computing power = Nvidia.

Sequoia’s doubts lasted less than three months.

In September 2024, OpenAI released GPT-1, where inference computation paradigms emerged—no longer relying on larger models, but on longer reasoning, with post-training + RL unlocking model capabilities. A new AI capability growth curve opened, and the market saw a second growth pole in computing power demand.

But the new paradigm soon revealed cracks—DeepSeek R1 was released, optimizing inference efficiency: training costs under $6 million, achieving near state-of-the-art inference capabilities.

On January 27, 2025, Nvidia’s market value evaporated by $593 billion in a single day. The second bubble theory began to surface.

The core market doubt was: do we really need so much computing power to achieve the same AI capabilities? This panic was fierce but resolved faster. A month later, Nvidia released its earnings report, with quarterly revenue of $11 billion exceeding expectations. The market proved that the new inference paradigm would generate more inference demand, and total compute demand did not decline but increased.

OpenAI, which promoted the inference paradigm, became the absolute center of this wave. Those who signed contracts with it soared. CoreWeave completed an IPO with a five-year, $11.9 billion contract, Oracle set a new high with a $300B "Stargate" plan, and Broadcom secured a hundred-billion-dollar custom chip order.

This time, market confidence was shaken the shortest, as OpenAI mass-produced "concept stocks," shifting the belief anchor from "training arms race" to "large-scale inference deployment."

The third was from October to November 2025.

Goldman Sachs issued a report listing five signs of AI bubbles: CapEx peaking, slowing corporate profit growth, rising tech debt, the Fed’s rate cut cycle starting, and widening credit spreads, explicitly comparing to the pre-Internet bubble era of 1997. Bank of America fund managers’ survey for the first time in 20 years showed signs of "overinvestment." Deep dives by Wired and The Atlantic Weekly published the same week pointed to a common finding: 95% of corporate AI investments have not yielded tangible returns.

Huge AI investments in the AI industry chain reached a climax, with Nvidia’s revenue coming from cloud providers, whose AI revenue growth came from model companies expanding, whose valuations came from capital, and whose returns came from paper revaluations of model companies.

But who is paying for AI upstream?

In Q3 2025 earnings calls, US tech giants almost responded with the same sentence to analyst questions, also countering Goldman Sachs’ judgment of CapEx peaking: "We’d rather invest more than lose the future." Goldman left a small tail in its report, suggesting that the bubble signs are present but the burst is still some distance away—more like 1997 than 1999.

In November 2025, the Fed cut interest rates by another 25 basis points, continuing liquidity support for valuations. The Nasdaq repeatedly hit new highs amid doubts. The market’s consensus became: I know there might be a bubble, but it’s more dangerous to sell now than to stay.

But what truly resolved this wave of doubts was the arrival of a new paradigm.

In the second half of 2025, Agentic AI exploded, transforming AI from dialogue into autonomous digital employees capable of planning, executing, and iterating independently. AI directly replaced workflows, with revenue ceilings rising from "search replacement" to "labor force replacement." More importantly, agents naturally consume dozens of times more tokens than dialogue, and the demand for computing power not only did not decline but opened up space for exponential growth.

Looking back at the first three waves of "bubble theories," three clues have always run through them.

How long will compute demand last; who will bear the huge AI expenditure and what is their ROI; whether large models will usher in a new paradigm breakthrough.

After three waves, the market has formed a dangerous consensus: "Doubts will be quickly disproved."

Behavioral finance shows that investors who re-enter after panic tend to have higher risk appetite than before because they have "verified" that the panic was wrong.

02 The Fourth Bubble of AI and Three Cracks

In recent trading days, the market has experienced intense volatility.

The first crack appeared between the "profits" and "cash flow" of tech giants.

In Q1 2026 earnings season, giants’ profits hit new highs, but cash flows approached zero. Amazon’s free cash flow plummeted 95% YoY, and the four tech giants burned a total of $2 billion daily. Some giants’ "net profit growth" was interpreted as a paper revaluation of AI investments, using AI profits to justify further investment, a circular argument.

Goldman Sachs provided a figure in April: about 40% of the expected profit growth of the S&P 500 in 2026 would come from the industry chain effects of AI-related capital investments. Several investment banks and media estimated that by 2026, the AI-related capital expenditure of major cloud providers had already reached over $600 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 foundation: AI CapEx can only increase, not decrease, affecting the entire system.

Morgan Stanley pointed out: in 2026, the CapEx-to-revenue ratio of large-scale enterprises will reach 34%, and by 2028, 37%, surpassing the 2000 internet bubble peak of 32%. Between 2026 and 2028, the total AI infrastructure expenditure of just the five giants will reach $2 trillion.

More covertly, these five companies also have nearly $1 trillion in off-balance-sheet leasing commitments—long-term contracts for data centers not yet built, which do not appear on any balance sheet.

Global AI call volume is surging, with many companies shouting "tokenmaxing." FOMO (fear of missing out) is spreading; CEOs fear being left behind by AI, rushing to "push tokens" for employees.

However, in the "tokenmaxing" movement, much of the consumption comes from systemic redundancies within the Agent architecture, over-designed harnesses, creating huge token call bubbles. No institution has yet split the ratio of "effective computation" to "architectural idling."

According to foreign media, Uber burned through its entire annual AI coding budget in just the first four months of 2026.

Under the push for tokenmaxing, engineers began using tools like Claude Code as parallel labor: running multiple tasks simultaneously, opening multiple worktrees, agents searching, generating, erroring, and repairing over long periods. AI usage appears to have increased, but finance departments find it hard to immediately quantify the actual output of these tokens.

Call volume is a core valuation metric for model companies, but if this metric is inflated, how reliable is the trillion-dollar valuation built on it?

Whether the agentic paradigm truly enhances enterprise productivity is the second crack in the market.

The capital market’s frenzy continues: reports suggest Anthropic’s valuation has reached nearly $965 billion, and it has secretly filed for an IPO; OpenAI has also secretly filed for an IPO, with a recent valuation of $852 billion.

Clearly, the market is paying full price for a future that has yet to materialize. This doesn’t necessarily mean a crash, but it could mean very little room for error.

"All great technological revolutions generate bubbles. No one can judge perfectly. You either invest heavily to capture market share without worrying about overspending; or invest too little and lose market share," Ray Dalio, founder of Bridgewater Associates, said in an interview on June 3.

Dalio believes that bubbles burst when investors try to convert paper wealth into cash, and the current AI-driven market "is heading down that path," even though AI itself "is an incredible technology."

From this perspective, the long-term value of technology and short-term valuation bubbles can coexist, just as the internet continued to reshape the global economy after the burst of the dot-com bubble.

Discussing the AI bubble from a macro perspective is a popular media topic, but it’s not an "effective issue" in investors’ eyes.

Peter Thiel believes, "AI technology is real, but the market has already overvalued the next 15-20 years." In Q3 2025, he completely sold out his Nvidia holdings, worth $100 million, accounting for 40% of his fund, and cut 76% of his Tesla holdings, reducing total holdings by 65%. He accurately predicted the internet bubble in 1999; whether his current prediction is different remains unknown. There is no answer yet.

But what is certain is that Thiel missed the Agent paradigm frenzy that began at the end of 2025.

Not only Thiel. Berkshire Hathaway’s Q1 2026 report shows Buffett’s cash holdings have swollen to $397.4 billion—an all-time high, accounting for 59% of total assets.

The long-term trend of technological evolution remains positive, but that doesn’t mean investors won’t take profits or reduce holdings in the short term. The contradiction between long-term trends and short-term investment strategies is the third crack in the market.

Above the three cracks, the market’s nerves are already becoming sensitive. As US interest rate expectations rise and doubts about AI capital expenditure overheating grow, Korea, one of the most "pure-blood" markets in this wave of AI, has begun to fluctuate wildly. Its rise is driven by AI belief; its sharp decline also stems from the loosening of that belief.

Investment and trading have thus entered a very challenging zone.

Many experienced investors still firmly believe, "The AI bubble hasn’t arrived yet." But judging whether AI is a bubble or undervalued, and whether one can establish an effective investment system, are two entirely different matters. The former is a directional judgment; the latter involves rhythm, position, valuation, cash flow, and exit windows.

In a market where belief persists and volatility intensifies, seeing the long-term trend doesn’t mean most people can withstand short-term setbacks.

The fourth generative AI bubble may not have arrived yet, but the warning signs are already here. Experienced and adaptable captains can navigate through storms to find treasure, but ordinary sailors may be lost in the waves.

Original link

Click to learn about Rhythm BlockBeats’ job openings

Join the official Rhythm BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia