Hotcoin Research｜The AI Revolution Continues: Why Is the AI Stock Market Starting to Deflate?

By the end of July 2026, the ongoing global investment boom in AI suddenly cooled, revealing cracks in a market structure supported by high valuations and high leverage. On July 28, the South Korean KOSPI index fell by more than 10% in a single day, with core companies in the AI supply chain, such as Samsung Electronics and SK Hynix, becoming centers of sell-offs. During the same period, the hedge fund Situational Awareness, focused on AI stocks, sold off its public stock portfolio after incurring losses. Notably, this decline did not occur against a backdrop of stagnating AI demand, collapsing chip revenues, or regression in model capabilities. On the contrary, Microsoft Azure, Google Cloud, and AWS continued to show strong growth, and Nvidia's data center revenues also reached new highs.

This seemingly contradictory phenomenon is the starting point for understanding the AI bubble: technological revolutions and asset bubbles are not mutually exclusive. AI can continuously improve productivity and create new product entry points, while simultaneously leading to excessive capital expenditure, imbalanced financing structures, and asset prices being overdrawn. The real question to answer is not whether "AI has value," but whether the industry's revenues, capital investments, and financial pricing are still aligned. When technological advancements change on a monthly basis, industry returns on an annual basis, and market pricing on a second-by-second basis, the risk of a bubble accumulates in the misalignment of these three speeds.

1. Capital Risks in AI Are Accumulating: Industry Growth and Asset Bubbles Amplifying Simultaneously

To determine whether there is a bubble in AI, one must first distinguish between industrial facts and financial pricing. The number of AI users, model capabilities, cloud computing demand, and chip revenues are all increasing, which is enough to refute the claim that "AI has no real demand"; however, industrial reality does not equate to reasonable valuations, nor does it mean that every capital expenditure will yield sufficient returns.

1.1 AI Demand Continues to Grow; Industry Prosperity Is Not Fabricated

The Stanford University "2026 AI Index Report" shows that the adoption rate of AI among surveyed institutions has reached 88%, with generative AI reaching approximately 53% of the global population within three years. Model capabilities have also not stagnated: several cutting-edge models have approached or exceeded human benchmarks in tests such as PhD-level scientific Q&A, multimodal reasoning, and competitive mathematics.

Source: https://hai.stanford.edu/ai-index/2026-ai-index-report

Commercial data is also improving. By the second quarter of 2026, Microsoft Azure's annual revenue exceeded $100 billion for the first time, with paid seats for Microsoft 365 Copilot surpassing 30 million; Amazon AWS's second-quarter revenue grew by 37% year-on-year, reaching an annualized revenue level of $169 billion; Google Cloud's revenue increased by 82% year-on-year, with backlogged orders reaching $514 billion. Nvidia's data center revenue for the 2026 fiscal year reached $193.7 billion, a 68% year-on-year increase. These figures indicate that companies are indeed paying for computing power, cloud services, and AI tools; AI is not just a hollow shell with valuations but no revenue.

1.2 The Bubble Arises from "Investment Outpacing Returns"

A more accurate definition of the AI bubble is not an increase supported by false technology but rather the market's excessive capitalization of real technology. According to MSCI statistics, as of May 2026, capital expenditures of U.S. AI-related companies grew nearly 60% year-on-year, about ten times that of non-AI companies, but revenue growth was significantly lower than capital expenditure growth. Meanwhile, AI-related companies enjoy valuation premiums in most markets, with AI-related firms in the U.S., South Korea, and Taiwan having price-to-book ratios approximately three times that of local non-AI firms.

Market concentration further amplifies this pricing. Data from the S&P Dow Jones Indices shows that as of June 30, 2026, the combined weight of the top ten constituents of the S&P 500 index was approximately 36.4%; depending on different dates and index adjustments, this ratio once approached 40%. Except for JPMorgan Chase, the remaining positions are occupied by large technology or semiconductor companies. When tech giants simultaneously bear the roles of index weight, AI capital expenditure, and profit growth, AI is no longer just an industry track but becomes a core risk factor held by global passive funds, pension funds, and index products. When the market rises, concentration can amplify returns; when the market turns, the same structure can also amplify withdrawals.

1.3 The Decline of Tech Stocks Reveals Market Structure Vulnerabilities

The adjustment of AI stocks in July 2026 shows that market pricing has shifted from "revenue growth" to "whether growth can cover investments." After Samsung Electronics and SK Hynix reported strong performances, their stock prices still fell significantly; by the end of July, the KOSPI index had dropped more than 10% in a single day. This indicates that when market expectations are already high, merely achieving growth is insufficient to support prices; companies must continuously exceed expectations and prove that new capacities will not translate into inventory, depreciation, and price competition in the future.

According to Axios, the AI hedge fund Situational Awareness, founded by former OpenAI researcher Leopold Aschenbrenner, has sold off its entire public stock portfolio. This does not prove that the AI bubble has burst but reveals a more realistic side of the capital market: even if the long-term directional judgment is correct, high concentration and high leverage may still prevent investors from waiting for long-term logic to materialize. Often, what first breaks investors is not the fundamentals of the company but the cash flow and liquidity.

AI industry growth and asset bubbles can coexist. Demand growth proves that the AI revolution is still ongoing, but capital expenditure, market concentration, and valuation premiums have upgraded the question from "Is the technology effective?" to "Can capital returns be realized?" The recent decline does not confirm the end of the industry cycle but is a clear signal that the financial structure is beginning to come under pressure.

2. How AI Bubbles Form: Simultaneous Expansion of Capital Expenditure, Depreciation Pressure, and Financial Leverage

The AI bubble is not caused by a single company or a single valuation metric but is a continuous process: insufficient computing power drives capital expenditure, capital expenditure supports upstream revenues, upstream revenues reinforce market narratives, rising stock prices lower financing costs, ultimately attracting more capital. When this cycle operates fast enough, the market easily misinterprets short-term supply shortages as long-term profits that can grow indefinitely.

2.1 The Capital Expenditures of the Four Major Tech Giants Approach Historical Extremes

According to the latest guidance for 2026, the annual capital expenditures of Microsoft, Alphabet, Amazon, and Meta total approximately $720 billion to $745 billion. If the portion of Microsoft removed from capital expenditure due to changes in lease accounting is re-included in economic input, the actual AI infrastructure investment scale of the four companies is approximately $735 billion to $760 billion.

The latest changes in the companies' capital expenditure guidance for 2026 mainly focus on:

• Microsoft: approximately $175 billion (originally about $190 billion), down due to changes in lease classification, but economic input plans remain unchanged.
• Alphabet: $195 billion to $205 billion, up by about $15 billion from previous guidance for servers, data centers, networks, and AI research and development.
• Amazon: approximately $220 billion, an increase of $20 billion from the beginning of the year for AWS, AI chips, robotics, and satellites.
• Meta: $130 billion to $145 billion, with the lower limit of guidance raised from $125 billion for AI training, inference clusters, and data centers.

Data Source: Each company's 2026 financial reports and management guidance. The statistical criteria for capital expenditures, financing leases, and other infrastructure investments differ among companies, and the total value is used to observe investment intensity, not representing fully comparable accounting standards.

The International Energy Agency (IEA) statistics show that driven by data center construction, the combined capital expenditures of Microsoft, Amazon, Alphabet (Google's parent company), Meta, and Oracle exceeded $400 billion in 2025, even surpassing the investment scale in global oil and gas production; it is expected to grow by about 75% in 2026, approaching $700 billion. Huge investments are also beginning to translate into real energy pressures: global data center electricity demand grew by 17% in 2025, far exceeding the global electricity demand growth rate of 3%, with AI data centers growing even faster. According to IEA predictions, by 2030, global data center electricity consumption will double from current levels, with AI data centers' electricity consumption potentially tripling.

However, capital expenditure has a significant time lag: the GPUs, land, and electricity purchased today require user growth over the next few years to recover. Any slight deviation in demand judgment can quickly shift from "insufficient computing power" to "capital surplus."

2.2 The Return Threshold for New Investments Is Rising

Meta's second-quarter revenue reached $60.8 billion, a year-on-year increase of 28%, but its capital expenditure for the quarter reached $31.08 billion, leaving only $784 million in free cash flow. Amazon's operating cash flow over the past 12 months reached $161.4 billion, but net spending on property and equipment rose to $169 billion, turning free cash flow from an inflow of $18.2 billion in the same period last year to an outflow of $7.6 billion. Alphabet's second-quarter capital expenditure was approximately $44.9 billion, raising its full-year guidance to $195 billion to $205 billion.

Microsoft's situation is relatively stable. Its capital expenditure for the fourth quarter of the 2026 fiscal year was $41 billion, with free cash flow still at $19.6 billion, and Azure revenue grew by 43%. This indicates that massive AI investments have not plunged all tech giants into cash flow crises simultaneously. What truly deserves attention is not whether these companies will immediately run out of money, but how much additional revenue and free cash flow can be generated for every additional dollar of capital expenditure.

This is also a crucial difference between the current AI bubble and the internet bubble of 2000. The companies primarily bearing capital expenditures currently generally have mature cash flow sources from advertising, cloud computing, software subscriptions, and e-commerce, making them more resilient to risks. However, the fact that giants will not go bankrupt does not mean their stock prices will not be revalued, nor does it mean that the supply chain and highly valued startups can withstand a slowdown in capital expenditures.

2.3 Rapid GPU Iteration Brings Depreciation Pressure

Microsoft disclosed that about two-thirds of its capital expenditure for the fourth quarter of the 2026 fiscal year was allocated to short-cycle assets such as GPUs and CPUs. Alphabet also stated that approximately 60% of its technology infrastructure investment flows to servers, with the remainder mainly used for data centers and networks. Land and buildings can be used for decades, but GPUs, servers, and network equipment require continuous updates, with their actual economic lifespan far shorter than that of data center buildings.

This creates a paradox that is easily overlooked: the faster AI technology advances, the quicker the economic value of the previous generation of equipment declines. New chips can significantly reduce unit inference costs and promote industry adoption but may force cloud vendors to update their equipment prematurely. Technological advancements act as a deflationary force for users but can become an accelerator of depreciation for equipment holders.

Therefore, to determine whether AI capital expenditures are excessive, one must not only look at the number of data centers but also consider equipment utilization rates, revenue per unit of computing power, and the speed of replacement between old and new chips. If the newly added capacity cannot quickly convert into stable loads, GPU inventories may shift from being "scarce assets" to surplus equipment that requires depreciation.

2.4 Circular Financing May Overestimate Demand Independence

There exists a unique capital cycle in the AI industry: cloud vendors invest in model companies, which, after obtaining financing, purchase cloud computing power, and the cloud vendors then count these contracts as backlogged orders and future revenue. Microsoft disclosed that OpenAI's new Azure service commitments have reached $250 billion; as of the fourth quarter of fiscal year 2026, Microsoft's commercial remaining performance obligations reached $678 billion, an 84% year-on-year increase, and if OpenAI is excluded, the growth is 25%.

This does not equate to revenue fraud. Model training indeed consumes computing power, and cloud vendors genuinely deliver servers, electricity, and network services. However, it implies that part of the demand does not entirely come from end customers with positive cash flow but relies on the financing capabilities of model companies. When the capital markets are willing to continuously provide funds to model companies, the cycle can expand; once financing costs rise, the demand for computing power may decline faster than the apparent orders.

Notably, Microsoft's latest financial report also shows that the sequential growth of its commercial orders mainly comes from customers outside of leading model companies, with nearly 90% of annual cloud revenue coming from outside model companies. This indicates that AI demand is spreading to the enterprise market and cannot simply be attributed to circular financing. What truly needs to be observed is whether external enterprise demand can gradually replace financing-driven demand and become the long-term payers for AI infrastructure.

III. Differentiation in the AI Industry Chain: Real Revenue, Uneven Profit Distribution

The AI industry chain is not a completely unified interest group. Chip manufacturers, cloud platforms, model companies, and application developers are at different cash flow stages and bear different risks. The segment with the fastest revenue growth today may not be the one with the most stable long-term profits; the segment with the heaviest investment may not ultimately receive the largest value distribution.

3.1 Chip and Equipment Vendors Realize Revenue First but Are Most Vulnerable to Order Reversal Impacts

AI capital expenditures have begun to translate into actual revenue for upstream companies. NVIDIA's revenue for fiscal year 2026 reached $215.9 billion, a 65% year-on-year increase, with data center revenue at $193.7 billion, up 68%. Chip and infrastructure software company Broadcom also maintained rapid growth, with quarterly revenue reaching $22.2 billion as of May 3, 2026, a 48% year-on-year increase. This indicates that AI infrastructure construction has generated real procurement demand; however, whether rapid revenue growth can long-term convert into stable profits still depends on customer capital expenditures, product prices, and supply chain conditions.

However, the high prosperity upstream comes from concentrated customer investments. When Microsoft, Amazon, Meta, and Alphabet simultaneously expand data centers, chips, HBM, advanced packaging, optical modules, and power equipment will all benefit; once two or three of them cut capital expenditures, the order changes for supply chain companies will be amplified. Chip companies face not only demand cycles but also technological iterations, inventory, and competition from customers developing their own chips.

Thus, the main risk for current AI hardware lies in the market extrapolating temporary high profit margins as a long-term norm. Hardware companies may continue to grow but find it difficult to permanently maintain a state where demand exceeds supply, prices rise, and customers repeatedly purchase simultaneously.

3.2 Cloud Vendors Control Computing Power Access but Bear the Heaviest Capital Expenditures

Cloud vendors are the segment in the AI industry chain closest to a "toll road." Microsoft, Google, and Amazon not only lease GPUs but also provide databases, storage, security, identity, model deployment, and enterprise software. Once customers migrate their data and business processes to cloud platforms, the switching costs increase, allowing cloud vendors to extend from pure computing power sales to a complete AI technology stack.

Recent data has already shown this capability: Azure's annual revenue exceeds $100 billion, AWS's operating profit for the second quarter reached $16.6 billion, with an operating profit margin rising to 39.4%; Google Cloud's revenue grew by 82%, with backlogged orders reaching $514 billion. Compared to single model companies, cloud vendors have a broader customer base and more revenue sources.

However, the issue for cloud vendors is that they must purchase equipment and electricity in advance for future demand. When application demand is insufficient, model companies can reduce usage, and enterprise customers can cut back on trial projects, but cloud vendors still have to bear data center leasing, depreciation, and energy costs. Therefore, cloud platforms have stronger long-term bargaining power but also bear greater balance sheet risks.

3.3 Model Capabilities Continue to Improve, but Models Themselves Are Becoming Easier to Replace

OpenAI disclosed that its enterprise business has contributed over 40% of revenue, with the API processing over 15 billion tokens per minute; the Stanford AI Index shows that the gaps in comprehensive capability rankings among several leading models have significantly narrowed. Model demand continues to grow, but the competition among leading models is shifting from "can it complete the task" to "who is cheaper, more stable, and more suitable for specific scenarios."

This will have a bidirectional impact. The decline in model prices and the improvement in inference efficiency are conducive to expanding the scale of AI use but will compress the unit revenue of model companies. Open models, cloud vendors' self-developed models, and vertical models are continuously entering the competition, and enterprise customers are also beginning to adopt multi-model architectures to avoid being locked in by a single supplier. Microsoft disclosed that since 2026, the number of customers using multiple model suppliers has increased fivefold, indicating that model replacement has shifted from a technical possibility to an enterprise procurement strategy.

Model companies may still build strong brands, data, and user networks, but establishing a long-term monopoly solely based on leading model parameters is becoming increasingly difficult. The real barriers in the future may come from enterprise data, distribution channels, user workflows, and unit inference costs, rather than a single benchmark test lead.

3.4 The Application Layer Determines Final Demand but Has Yet to Prove Profit Contribution Universally

AI truly creates value when enterprises redesign processes rather than simply adding an AI tool for employees. Quantifiable cases have emerged in customer service, programming, financial analysis, medical records, and supply chain management, but there is still a gap between improving efficiency at a single point and increasing profits across the entire company, which involves data governance, system integration, employee training, and responsibility allocation.

The capital expenditures at the application layer are relatively low and have the greatest opportunity to benefit from declining model costs; however, the competitive threshold may also decrease. Many AI applications rely on the same foundational models, and functionalities are easy to replicate, leading to profits potentially flowing back to traditional software companies that have customer relationships and business data. The long-term winners in the AI industry may not be the companies with the strongest models but rather those that best integrate models into paid processes.

The AI industry chain has already generated real revenue, but profit distribution is unstable. Hardware benefits first but also faces order cycles; cloud vendors control access but bear capital expenditures; model companies grow rapidly but face price competition; the application layer is closest to final value but still needs to prove scalable profits.

IV. The Cryptocurrency Market Becomes a 24/7 Leverage Transmission Layer for AI Risks

Traditional stock markets primarily trade during fixed hours, but cryptocurrency platforms are transforming expectations for NVIDIA, Microsoft, AI indices, leveraged ETFs, and AI startups into all-day contracts, allowing users to conveniently participate in global AI asset trading through platforms like Binance and Hotcoin. The cryptocurrency market has not created an AI bubble but has changed the way risks are traded, amplified, and transmitted. The cryptocurrency market extends AI risks from stock trading hours to an all-day market and increases the levels of risk through perpetual contracts, leveraged ETFs, and AI-related tokens. It enhances global participation and price discovery speed while allowing oracle deviations, funding rates, and forced liquidations to participate in asset pricing more quickly.

4.1 Perpetual Stocks Are Extending AI Trading Beyond Traditional Market Closures

CoinGecko data shows that the monthly trading volume of RWA perpetual contracts grew from $230 million at the beginning of 2025 to $34.717 billion in May 2026, with a cumulative trading volume of $1.32 trillion in the first five months of 2026. The monthly trading volume of perpetual stocks grew from $831 million in July 2025 to $34 billion in May 2026, an increase of nearly 40 times.

The AI industry chain is a significant driving force behind the growth of perpetual stocks. In May 2026, the trading volume of Micron-related perpetual stocks increased from $736 million the previous month to $13.16 billion, an increase of about 17 times; NVIDIA, Tesla, and Circle also ranked among the high trading volume targets. By June 2026, the monthly trading volume of RWA perpetual contracts further approached $47 billion.

Most of these products are not stock tokens and do not mean that users hold corresponding company stocks. They are typically synthetic contracts referenced by oracle prices, maintained through margin and funding rates to anchor prices. Users trade stock price changes rather than voting rights, dividends, or residual rights of the company.

4.2 Leveraged ETFs Add Double Leverage, Turning Path Loss into Liquidation Risk

Perpetual stocks already include margin leverage. If the contract target is a two or three times leveraged ETF, the risk structure becomes further complicated: the first layer comes from the daily leverage reset within the ETF, and the second layer comes from the margin and forced liquidation of the perpetual contract account.

Leveraged ETFs do not guarantee a fixed multiple of cumulative returns for the underlying within a week or a month. The U.S. Securities and Exchange Commission has provided actual cases: one index rose 2% over four months, while the corresponding two times leveraged ETF fell 6%; another index rose about 8%, while its three times leveraged ETF fell 53%. The reason lies in daily resets, compounding, and volatility loss.

If users further leverage the leveraged ETF through perpetual contracts, the underlying's path loss, perpetual funding rates, and account liquidation lines will simultaneously act. During continuous market declines, leveraged ETFs need to reduce risk exposure, and perpetual accounts may also be forcibly liquidated, forming an amplification chain of "underlying stock decline --- ETF reduction --- perpetual liquidation --- liquidity continues to decline."

4.3 AI Tokens Carry the Tech Narrative but Often Do Not Capture the Same Cash Flow

In addition to perpetual stocks, the cryptocurrency market also carries the tech stock narrative through AI tokens, AI Agents, and DePIN projects. However, these tokens do not share the same equity structure as publicly listed AI companies. Holding NVIDIA stock represents a claim on the company's profits and residual assets, while holding a certain AI token may not grant rights to protocol revenue, computing power equipment, or legal rights of model companies.

This means that the same "AI growth" narrative may be further abstracted when entering the cryptocurrency market. Projects having users does not mean that tokens capture revenue, and network-generated revenue does not guarantee distribution to token holders. The valuation of some projects derives more from future adoption expectations, token liquidity, and market sentiment rather than verifiable cash flow.

What truly deserves attention is not whether the project uses the AI concept but the function of the token within the system: is it used to pay for real computing power, obtain service discounts, bear network security, or share verifiable revenue? If the token is merely a narrative carrier, then during tech stock adjustments, it often suffers greater drawdowns than publicly listed companies with mature cash flows.

4.4 Oracles and Continuous Liquidation Shorten Risk Transmission Time

In March 2026, S&P Dow Jones Indices officially authorized Trade[XYZ] to launch perpetual contracts based on the S&P 500 on Hyperliquid, indicating that traditional index providers are beginning to recognize the commercial value of crypto perpetuals as a new distribution channel.

However, 24/7 trading also brings new pricing issues. After U.S. stock markets close, the underlying stocks do not continue to trade, and stock perpetuals can only be priced based on index futures, related assets, market maker quotes, and changes in market expectations. At this time, the contract price reflects the expectation of the next opening rather than an immediately arbitrageable spot price. Significant events over the weekend may cause contracts to deviate significantly from the previous trading day's closing price.

When traditional markets reopen, spot prices, oracle data, and perpetual positions need to converge again. If the discrepancies are too large, highly leveraged positions may be liquidated before the spot market regains liquidity. Therefore, while the crypto market provides a more continuous expression of risk, it does not necessarily offer a more accurate fundamental price. It is primarily a continuous trading layer of sentiment, leverage, and expectations.

Outlook and Conclusion: AI Will Not Disappear Due to Bubbles, Markets Will Return to Cash Flow Pricing

The AI market is facing two opposing forces simultaneously. On one hand, the latest growth data from Microsoft, AWS, and Google Cloud indicate that demand for computing power and enterprise AI remains strong; on the other hand, the cash flow pressures on Meta, Alphabet, and Amazon suggest that capital expenditures have entered a phase where returns need to be justified.

This combination means that the market is more likely to experience scenarios where "strong earnings reports coexist with weak stock prices" or "high investment leads to strong rebounds." Investors are no longer trading on whether there is demand for AI, but rather on how much better the actual results are compared to high expectations. Once expectations are sufficiently high, even if revenues continue to grow, a slowdown in growth or continued increases in capital expenditures may trigger price adjustments. The focus of the market is shifting from "how many GPUs are owned" to "how much cash flow each dollar invested can generate."

In the next one to two years, the AI market may see three outcomes: if enterprise demand grows rapidly and absorbs the new computing power, capital expenditures will gradually convert into revenue and profits, leading to a relatively mild market digestion; if demand continues to grow but does not keep pace with investment speed, the industry will not collapse but will experience valuation adjustments and profit redistribution, with clearer differentiation among chip, cloud platform, model, and application companies; if financing conditions tighten, model companies cut back on computing power purchases, and cloud vendors reduce capital expenditures, upstream orders may weaken in tandem, leading the market from valuation adjustments to further downward revisions of profit expectations.

To assess whether risks are escalating, five signals need to be monitored: whether capital expenditures continue to outpace AI revenues, whether free cash flow continues to deteriorate, whether depreciation pressures are significantly increasing, whether cloud orders can spread from model companies to enterprise customers with independent payment capabilities, and whether enterprise-level AI can genuinely improve renewals, costs, and profits. If these indicators weaken simultaneously, combined with rising market concentration, margin financing, and leverage product scales, the originally limited fundamental changes may evolve into more severe price fluctuations.

Conclusion

The AI revolution is not over, yet the bubble alarm has already been sounded; these two realities are not contradictory. Technological advancements will expand applications, lower costs, and create new industrial opportunities, while also stimulating enterprises to reinvest, accelerate equipment depreciation, and raise market expectations. The faster the technological progress, the easier it becomes for capital to factor in distant future revenues into today's prices.

Historical examples of railroads, the internet, and fiber optics demonstrate that technological revolutions can change the world while causing significant losses for many investors. The true long-term value of AI will not be negated by a single adjustment in tech stocks, but not all AI assets will automatically become winners simply because the industry direction is correct.

The crypto market further extends this pricing experiment into 24/7 uninterrupted trading. Stock perpetuals, leveraged ETF perpetuals, and AI tokens provide more users with exposure to AI, allowing valuations, leverage, and liquidations to interact at a faster pace. Ultimately, what will determine whether AI assets can survive the bubble is not how long the story can continue, but whether cash flow can catch up with investments before capital patience runs out.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.