AI Bubble Is Bursting

链捕手Published on 2026-06-07Last updated on 2026-06-07

Abstract

The AI Bubble is Bursting: A Necessary Purge on the Path to Ubiquitous Intelligence Market volatility has reignited debates about an AI bubble, with figures like Ray Dalio pointing to high valuations. However, this parallels the dot-com bubble, which, despite its crash, laid the physical infrastructure for today's internet era. The current AI investment frenzy, with tech giants planning trillions in infrastructure spending far outstripping current AI application revenues, appears similarly imbalanced. This 'bubble' is seen as an inevitable phase for a disruptive technology, paying the "innovation tax." Critically, AI inference costs have plummeted over 99.7% since 2023, making intelligence nearly free at the margin. This hasn't reduced spending but has instead unlocked massive new demand, as seen in enterprise AI cloud expenditure tripling. This follows the Jevons Paradox: efficiency gains lead to greater total consumption. The market is now entering a cleansing phase, weeding out speculative ventures lacking real moats. The deeper shift is a move from capital expenditure (CapEx) on hardware to value creation in operational expenditure (OpEx) through AI applications that solve real industry problems. While infrastructure valuations are high, rapid earnings growth from widespread AI adoption across sectors—from manufacturing and finance to law and healthcare—may digest these valuations over time. Ultimately, this creative destruction will leave behind robust infrastructure...

Original Title: AI Bubble Is Already Bursting

Original Author: Chengbei Xugong, Gelong

In recent days, the market has experienced severe volatility, with "AI bubble theories" swirling.

Ray Dalio, founder of Bridgewater Associates, said: There is a bubble in the AI market, and the level is "relatively high."

NVIDIA CEO Jensen Huang said: AI presents a massive opportunity, and the demand for computing power has just begun to explode.

Who to believe?

Both of them are correct.

Is there a bubble in the AI industry? Inevitably, yes.

However, bubbles in the technology field are often the only way society can pay homage to a disruptive advanced productive force.

It is not a purely derogatory term.

In the long run, this is an inevitable phenomenon at the beginning of the emergence of advanced productive forces.

Many people compare the current situation to the 2000 internet bubble, feeling deeply concerned.

The dot-com bubble did cause the Nasdaq to plunge nearly 78%, evaporating over $5 trillion in wealth.

But twenty years later, which industry can operate without the internet?

Today, the value of the internet industry far exceeds that of the bubble era.

The AI bubble, at least on the surface, appears to be a similar situation.

The bubble existing in the capital markets cannot stop virtually every industry in society from actively being empowered by AI.

AI+ is an unstoppable trend.

Just as all industries today are inseparable from the internet, all industries in the future will be inseparable from AI.

01

In that era where any company with a '.com' in its name could go public and raise money, the Nasdaq soared nearly 600% between 1995 and 2000. Subsequently, a financial storm lasting two and a half years ensued.

Those prominent names back then, like software company MicroStrategy, plummeted 62% in a single day due to accounting scandals and overblown promises; Pets.com (selling pet food online) and Webvan (the pioneer of fresh grocery e-commerce) went bankrupt outright.

......

Amid the panic, almost everyone condemned the internet as a scam.

However, the physical infrastructure left behind by the reckless spending of speculative capital often nurtures the next era's supergiants at extremely low costs.

The bubble burst not because of the internet technology itself, but because the pace of physical infrastructure construction couldn't keep up with the market's rhythm.

For example, the once-dominant telecommunications companies (like WorldCom, Global Crossing) invested heavily in laying global submarine cables and dense wave division multiplexing networks. While they went bankrupt, these cheap "information superhighways" became the perfect breeding ground for the later rise of Netflix, Zoom, and mobile internet.

Without the frenzied, ahead-of-its-time global investment in telecom infrastructure around 2000, there would have been no later explosion of YouTube's video streaming or the development of cloud computing infrastructure.

Amazon is the most typical example.

Its stock price plunged from a high of $107 in 1999 to $7 in 2001, a drop of over 90%.

But it survived because its underlying business logic, "reconstructing retail through the network," aligned with the direction of advanced productive forces.

This is a classic case of Amara's Law: overestimating the short-term impact of a new technology while severely underestimating its long-term impact.

In the early stages of a technological revolution, the frenzy of speculative capital inevitably leads to overinvestment, forming a bubble.

This is the intelligence tax that innovation must pay.

But when the bubble subsides, what remains will be a more indestructible advanced productive force.

02

Returning to 2026, the bubble in the AI industry appears even larger.

Just the five major cloud service providers—Amazon, Google, Meta, Microsoft, and Oracle—are expected to have capital expenditures of $690 billion in 2026, with total AI infrastructure investment projected to reach $5.3 trillion by 2030.

Of this, only about 25% is spent on GPUs; the remaining 75% is invested entirely in physical infrastructure: liquid cooling systems, power transmission, network switches, optical modules, and land.

On the revenue side, the combined total revenue of all leading pure-play AI vendors like OpenAI, Anthropic, Cohere, Mistral, and Perplexity in 2026 is expected not to exceed $40 billion.

Nearly $700 billion poured into the foundational layer, only hundreds of billions returned from the application layer.

This severe asymmetry—what is it if not a bubble?

One cannot simply jump to such a conclusion.

There is a crucial point that cannot be overlooked.

In March 2023, when OpenAI released GPT-4, the blended cost per million input tokens was about $30.

By April 2025, thanks to model architecture optimization and inference computing improvements, the price for models with equivalent intelligence levels plummeted to $0.1-$0.15 per million tokens.

According to Stanford's "AI Index Report" and TokenCost data: AI inference costs have dropped over 99.7% in the past two years.

Following traditional linear thinking, with costs plummeting, corporate AI spending should decrease, right?

But the reality is, enterprise AI cloud spending tripled between 2024 and 2025.

Why?

Because when the marginal cost of "intelligence" approaches zero, AI is no longer just a simple text summarizer or chatbot; it has entered a new era of agents and multimodal augmented retrieval.

Companies are starting to let AI agents automatically run tasks thousands of times—writing code, scanning millions of legal contracts, simulating biological experiments.

Cheap tokens have unlocked a massive amount of long-tail demand previously uncommercializable due to cost constraints.

This point can also be gleaned by comparing NVIDIA in 2026 and Cisco, the network hardware kingpin of 2000.

Their ecological niches are strikingly similar, but their underlying financial health is worlds apart.

This precisely validates the economic principle known as the "Jevons Paradox": technological progress improves energy efficiency, but instead of reducing energy consumption, it leads to greater demand due to cost reduction.

Even after experiencing the so-called "DeepSeek moment" early last year, the market quickly sobered up in the following months: the more optimized the algorithms, the lower the barrier for enterprise AI adoption, ultimately leading to an exponential increase in total computing power consumption.

It is precisely because of this that AI can gradually embed itself into virtually every traditional industry.

Just as all industries embraced internet+ over the past two decades.

From SaaS software to biopharmaceuticals, and then to advanced manufacturing robotics driven by embodied intelligence, in 2026, almost every industry is embracing AI+.

No one discusses "should we use AI," but rather anxiously questions "is our data properly cleaned? Do we have enough API call quota? Is our RAG architecture optimal?"

Currently, the AI industry does indeed have a bubble.

But for enterprises, if you don't embrace the bubble, you will be crushed by the times.

This has been proven over the past two decades of the internet era.

03

Currently, we are undoubtedly at an extremely critical node in the technology lifecycle: just before the "Trough of Disillusionment" on Gartner's Hype Cycle, or at a turning point in the theory of "Technological Revolutions and Financial Capital."

The AI bubble is already bursting, it's just that many haven't realized it.

Over the past few years, a large number of venture capital firms suffered from fear of missing out (FOMO).

Any new startup, with just a few dozen PPT slides, wrapping a layer of OpenAI's API, could raise money. Now, as the tide recedes, these companies with no moat, only concepts, are dying en masse.

This is the market self-cleansing, a manifestation of the bubble bursting.

But this is only the surface phenomenon.

Three profound evolutions are taking place in the market's underlying logic:

First, the shift of value from CapEx to OpEx

Currently, the money is being made by the shovel sellers—NVIDIA, TSMC, and those selling optical modules and server liquid cooling equipment have captured most of the gains.

But as computing power gradually becomes "infrastructural," like water and electricity, the real excess profits will gradually shift to the application layer.

Namely, those AI-native enterprises that can use extremely low-cost tokens to genuinely solve pain points in vertical industries and reshape business processes (OpEx optimization).

Second, valuation multiple compression and earnings digestion

The market's high valuation for AI infrastructure does not necessarily mean a crash.

In many cases, the high-speed growth of corporate earnings will gradually digest lofty valuations through a "time-for-space" approach.

As long as the revenue growth of cloud computing giants keeps pace with the depreciation rate of capital expenditures, this game of pass-the-parcel can evolve into an unprecedented industrial upgrade.

For example, global automotive manufacturing giants and chip giants, by introducing end-to-end AI twin technology, have shortened the R&D-to-mass-production cycle for new products by 35% and improved overall equipment effectiveness (OEE) by 18%.

Another example: in the financial industry, by 2026, quantitative trading, risk control, and credit assessment are fully dominated by multimodal agents. AI is not only processing macro expectations with microsecond timestamps but also deeply involved in micro-level asset pricing for each transaction.

In highly knowledge-intensive industries like law, healthcare, and auditing, AI has completed its evolution from "junior assistant" to "partner-level expert."

Among the over 1 billion active users of ChatGPT, Gemini, Claude, a significant portion uses it as a substitute tool for daily high-intensity intellectual labor.

Including you and me.

All of the above are tangible developments, visible to everyone.

04

Looking back at the grand history of technology, Schumpeter's "creative destruction" is always unfolding.

Capital markets are always impatient, hoping that $1 invested today will return $10 tomorrow.

When nearly $700 billion in infrastructure investment cannot be fully transformed into application-layer profits in the short term, the market is bound to experience a brutal reshuffle.

Eliminating those speculative, shell companies that rely solely on PPT presentations, and retaining those with real technological depth and viable application scenarios.

After the reshuffle, those cheap and massive computing centers, highly optimized model algorithms, will serve thousands of industries at extremely low cost.

After 2000, humanity entered a digital age where no industry could function without the internet.

Today, we are also irreversibly heading towards a fully intelligent era where all industries are governed and empowered by AI.

Amidst the noise of the bubble, the underlying momentum of productive force contains not a drop of water.

Related Questions

QThe article mentions that AI market bubbles are a recurring phenomenon in technological revolutions. What specific historical example does it draw a parallel to, and what was the long-term outcome?

AThe article draws a direct parallel to the dot-com bubble of 2000. While it caused a massive market crash and wealth destruction in the short term, the long-term outcome was that the internet's value far surpassed its pre-bubble valuation, becoming indispensable to almost every industry. The author suggests AI is following a similar pattern: short-term speculative excess (the bubble) funding the infrastructure for long-term, transformative adoption across all sectors.

QAccording to the article, what is the "Jevons Paradox" in the context of AI, and how does it explain the surge in corporate AI spending despite falling costs?

AThe Jevons Paradox states that technological progress that increases the efficiency of using a resource (like AI computing power) often leads to an *increase*, not a decrease, in the total consumption of that resource. The article applies this to AI: as the cost per token (the unit of AI computation) plummets (over 99.7%), it unlocks vast new, previously cost-prohibitive use cases. Companies don't just do the same tasks cheaper; they deploy AI for massively scaled, complex applications (like automated agents running thousands of tasks), leading to a net explosion in total AI spending and compute demand.

QThe article describes a significant asymmetry between AI infrastructure investment and application-layer revenue. What are the figures cited for 2026, and what broader point does this contrast illustrate?

AThe article states that for 2026, major cloud service providers are projected to have capital expenditures of $690 billion on AI infrastructure. In contrast, the combined revenue of all leading pure-play AI companies (like OpenAI, Anthropic) is projected to be under $40 billion. This massive asymmetry (nearly 700B in vs. ~40B out) illustrates the current 'bubble' characteristic of over-investment in anticipation of future demand. However, the author argues this infrastructure buildup is necessary, akin to the telecom overbuild during the dot-com era, which later enabled the rise of streaming and cloud computing.

QWhat are the three 'profound evolutions' in the market's deep logic that the article identifies as happening beneath the surface of the apparent bubble burst?

A1. Value shift from CapEx to OpEx: Profits will gradually move from sellers of infrastructure (hardware like GPUs) to AI-native application companies that use cheap compute to solve real business problems and optimize operational expenses. 2. Valuation compression and earnings digestion: High valuations for infrastructure may be sustained not by perpetual hype, but by companies growing their earnings fast enough to eventually justify the high prices ('time for space'). 3. The article heavily implies a third evolution (though not explicitly numbered as such in the provided text): The widespread, tangible integration of AI into core industries (automotive, finance, law, medicine), moving from concept to essential, productivity-enhancing tool, which validates the underlying technological shift.

QHow does the article characterize the nature of the current AI bubble's burst, and what is the primary mechanism of this 'self-purification' of the market?

AThe article characterizes the current bubble burst as a necessary 'market self-purification' or 'creative destruction.' The primary mechanism is the widespread failure and weeding out of speculative, low-substance companies that lacked a true technological moat—specifically, those that were merely 'wrapping' existing AI APIs (like OpenAI's) in a presentation without solving a unique problem or building defensible technology. This clears the field for serious companies with real technology and viable business models.

Related Reads

Has the 'Digital Gold' Narrative for BTC Failed?

**Title: Has the "Digital Gold" Narrative for Bitcoin Failed?** The article argues that Bitcoin's "digital gold" narrative remains valid despite a recent sharp price decline (from a peak near $126k in Oct 2025 to briefly under $61k in Feb 2026). It presents a long-term investment framework based on three core points: **1. Viewing Bitcoin as an Asset:** Bitcoin is presented as a superior potential store of value compared to gold. Key arguments are its absolute scarcity (21 million cap), superior portability, and transparent auditability via its public ledger. While acknowledging its current use in early, volatile stages (~3-4% global adoption), the author draws parallels to the early, disruptive phases of the internet and e-commerce. **2. Understanding the Recent Downturn:** The current ~50% correction is framed as a predictable, consensus-driven cycle following its post-halving peak (the 2024 halving preceded the Oct 2025 high). A crucial factor is a historic "changing of hands": the influx of new institutional buyers via ETFs allowed early, low-cost holders (miners, OG believers) to take profits. The author notes that while severe, Bitcoin's historical drawdowns (e.g., 93% in 2011, 77% in 2021-22) have been progressively smaller, suggesting maturing holder structure and decreasing volatility over time. **3. The Long-Term Perspective:** The long-term thesis hinges on Bitcoin capturing a portion of gold's market value. With Bitcoin's market cap at ~$1.4 trillion (at $70k) versus gold's ~$20 trillion, significant upside potential exists if the "digital gold" narrative is partially realized. However, the author strongly cautions that short-term risks remain, the bottom is unpredictable, and high volatility is inherent. The real risk is not Bitcoin failing but poor personal position management (over-leverage, wrong capital) and a lack of deep understanding, which can force investors out during severe downturns. The conclusion uses Amazon's 95% crash post-2000 dot-com bubble and subsequent 42x recovery as an analogy. The ultimate question is not if Bitcoin's price will rise, but if an investor's strategy and conviction can withstand the volatility to see the long-term play out. The recent divergence (gold up, Bitcoin down) is posed not as a narrative failure, but as potential evidence of this ongoing, painful transition from a speculative asset to a mainstream allocation.

marsbit1h ago

Has the 'Digital Gold' Narrative for BTC Failed?

marsbit1h ago

Has BTC's 'Digital Gold' Narrative Failed?

The article discusses Bitcoin's "digital gold" narrative, its recent price drop, and long-term outlook through the perspective of "Jason". It argues the narrative is not a failure but that Bitcoin represents a superior, new asset class due to its fixed supply (21 million), portability, and auditability. The piece compares its current ~3-4% global adoption rate to early internet/e-commerce, suggesting significant growth potential. Regarding the 2025-2026 price decline (from ~$126k to briefly under $61k), the author views it as a predictable, consensus-driven sell-off within Bitcoin's ~4-year cycle post-halving, exacerbated by a major "handover" from early, low-cost holders to new institutional buyers via ETFs. A key observation is that historical peak-to-trough drawdowns have lessened over time (e.g., 93% in 2011 to ~50% in 2026), indicating maturing volatility as holder structure changes. For the long term, the author uses a simple framework: Bitcoin's total market cap (~$1.4T at $70k) is only about 7% of gold's (~$20T). Even capturing 30-50% of gold's value would imply substantial upside. However, the article strongly cautions against viewing this as investment advice, emphasizing extreme volatility and the critical importance of risk management, position sizing, and deep fundamental understanding to survive severe drawdowns. It concludes by drawing a parallel to Amazon's 95% crash in 2000 and subsequent 42x recovery, stressing that the key is surviving market cycles to realize long-term potential.

链捕手2h ago

Has BTC's 'Digital Gold' Narrative Failed?

链捕手2h ago

From Code to Cognition: A Ten-Thousand-Word Guide to the Evolution of the Robot Brain

"From Code to Cognition: The Evolution of Robot Brains" The journey of robotic intelligence has shifted dramatically from manually coded systems to AI-driven brains. For decades, robots relied on layered software stacks—perception, state estimation, planning, control—each handcrafted. While predictable, they lacked adaptability. The 2010s saw deep learning revolutionize perception (e.g., object detection) and control (via reinforcement learning), but learned skills remained narrow. The arrival of Large Language Models (LLMs) marked a turning point. LLMs acted as high-level planners, interpreting natural language instructions and generating sequences of actions for traditional robotic systems to execute. However, true integration came with Visual-Language-Action (VLA) models, which fused vision, language, and motion prediction into a single network. Pioneered by models like RT-2 and open-source projects like OpenVLA, VLAs enable robots to reason and act directly from visual input and commands. The most advanced humanoid robots now employ a "dual-brain" architecture: a slow-thinking, large VLA (System 2) for reasoning and planning, and a fast-reacting, small network (System 1) for high-frequency motion control, sometimes with an even lower-level System 0 for balance. This split balances cognition with the physics of real-time movement. Computation is split between onboard hardware (e.g., NVIDIA Jetson) for safety-critical control loops and cloud/edge servers for non-critical tasks like learning and interfaces. A crucial driver is the open-source ecosystem—models like GR00T and OpenVLA allow startups to build upon pre-trained brains and fine-tune them with their own data, accelerating development. Despite progress, current systems struggle with recovery from errors, sample inefficiency, and long-horizon tasks. This has spurred the rise of **World Models**—neural networks that predict the consequences of actions. By simulating possible futures before acting (like NVIDIA Cosmos or Meta V-JEPA), robots can plan, recover, and generalize better. This represents the next frontier: shifting intelligence from learned reactions to an internal model of physics and cause-and-effect. The field is rapidly evolving. While not yet at its "ChatGPT moment," the convergence of cheaper hardware, scalable simulation, and world models points toward robots that are increasingly capable, adaptive, and useful. The question is shifting from "what can robots do?" to "what *should* they do?"

marsbit2h ago

From Code to Cognition: A Ten-Thousand-Word Guide to the Evolution of the Robot Brain

marsbit2h ago

Trading

Spot
Futures

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片