The AI wave, why did it suddenly "hit the brakes"?

Over the past two years, artificial intelligence has undoubtedly been the hottest trend in global capital markets, sparking a wealth frenzy that swept across the entire industry chain.

From upstream core hardware to downstream application implementation, the AI sector has experienced an epic rally: $NVIDIA(NVDA.US) has seen its market value soar due to its monopoly on computing power; $Z.AI(02513.HK), OpenAI, Anthropic and other large model companies have seen their valuations rise steadily; $SK Hynix(SKHY.US), $Intel(INTC.US) and other semiconductor giants have risen with the tide. Niche tracks such as optical modules, optical chips, and special ceramic materials have doubled or even tripled in value. Even traditional infrastructure sectors like data center construction, power support, and storage devices have ridden the wave of AI computing power development to achieve counter-trend performance.

In an instant, capital flooded in, companies increased their layout, and industry optimism reached its peak. Everyone was convinced that AI would usher in a new round of industrial revolution and growth cycle spanning decades.

However, in the midsummer of 2026, while the heatwave in Silicon Valley persists, the AI industry is facing a chilling "capital cold front." The Philadelphia Semiconductor Index has fallen by more than 18% over the last 20 trading days. Among them, Micron Technology (MU.US), the best-performing chip stock this year, has dropped nearly 24% in the same period. Star company Marvell Technology (MRVL.US), heavily promoted by Jensen Huang, has plummeted over 37% in 20 days. The once-cherished narrative of "permanent computing power shortage" is showing cracks.

This sudden cooling of valuations may simply be a retreat in market sentiment or short-term profit-taking by funds. However, it forces everyone to face the hidden reefs in the wild rush of the AI industry wave, which may determine the sustainability of this boom.

The First Hidden Reef: The Impact of the Regulatory Iron Curtain Falling

The iteration speed of AI technology far exceeds the pace of legal and regulatory updates. Blurred boundaries of technological empowerment, data security risks, algorithmic ethical hazards, monopoly and infringement issues, etc., may force global AI regulation to move from blank exploration to comprehensive tightening, becoming an important gate blocking the disorderly expansion of AI.

Currently, the global AI regulatory system has entered a stage of intensive implementation. The EU's Artificial Intelligence Act has officially taken effect, implementing strict access controls for high-risk AI systems and proposing transparency and risk assessment requirements for general AI models. China continues to improve management measures for generative AI, clarifying hard requirements for large model filing, data compliance, and content review. The United States, through executive orders and industry standards, conducts targeted regulation on large model training, data collection, and AI monopolistic behavior. In addition, special rules around cross-border data transmission of AI, copyright of AI-generated content, and employment replacement risks are also being continuously implemented in various countries.

The comprehensive tightening of regulations directly raises the industry's entry threshold. For small and medium-sized AI enterprises, a sharp increase in compliance costs means greater survival pressure. For leading tech giants, regulatory constraints lengthen product release cycles and force adjustments to globalization deployment strategies. Regulation does not intend to hinder the commercialization process of AI, but it will change the rules of the game.

The Second Hidden Reef: Sustainability of Capital Expansion, Scrutiny of Costs and Benefits

The explosion in AI industry valuations in the first half of this year clearly focused on the upstream industry chain. Tech giants in the downstream large models and applications, facing end-users, spent huge sums on computing power construction, model training, data center building, and technology R&D. This was an important support driving the upstream industry chain's performance improvement and subsequent valuation surge. Similarly, if these tech giants, as well as sovereign entities, reduce or slow down their investment in AI, this valuation logic will be shaken.

When scrutinizing AI investments, two main points are measured: return on investment and the ability to invest. AI investment has been ongoing for several years, and the effects produced are beginning to emerge. The current evaluation of the AI industry chain carries expectations for AI monetization. Judging from the speed and volume of the recent chase-up in the upstream industry chain, the current market expectation for AI investment is very optimistic. Once AI development falls short of expectations, AI investment progress slows down, and the amount invested shrinks, the overall market value of the upstream industry will contract.

The earnings results of the upcoming June fiscal quarter will become an important weather vane for predicting the direction of AI capital expenditure. Overseas tech giants such as Google (GOOG.US), Microsoft (MSFT.US), Amazon (AMZN.US), Meta (META.US), as well as domestic leading internet companies like Tencent (00700.HK) and Alibaba (09988.HK), will be closely observed under a magnifying glass. Their outlook on AI investment, the progress of AI development, and how AI applications integrate into their ecosystem services will all face stricter scrutiny. If the AI businesses of these giants fall short of expectations, the predictability of realizing positive cash flow is not high. If the cost-effectiveness of computing power investment, R&D investment, and infrastructure investment does not improve, they will inevitably actively cut back on subsequent capital expenditures, directly leading to a shrinkage in demand for the upstream AI industry chain.

At the same time, the performance and cash flow situation of these companies' existing businesses will also directly affect future AI investments. The profits and cash flows of their core businesses are the confidence behind these tech giants' long-term commitment to the tech track and AI layout.

However, the rapid penetration of AI technology is continuously disrupting the traditional industrial landscape. Google's advertising business, Amazon's e-commerce retail, Microsoft's Office suite, Meta's social media advertising... these are the true sources of capital expenditure for tech giants. Yet, AI's disruption of traditional industries is now biting back at its own "sponsors".

Taking the advertising industry as an example, generative AI has made content creation extremely cheap. A large amount of low-quality AI-generated information floods platforms, reducing the actual reach rate of ad slots. Brands are beginning to question the cost-effectiveness of digital advertising. Meanwhile, AI search is eating into the advertising share of traditional search engines—when users can get answers directly from large models, who is willing to click on those "sponsored links"?

More importantly, the revenue generated by these tech giants' core services relies on users' payment capacity and consumption demand. Human resources replaced by AI will reduce consumption due to decreased income, which will eventually feedback to these tech giants, causing significant declines in revenue for giants relying on traditional industries for profit. Under pressure on their own revenues, capitalists will naturally be unable to support the annual capital expenditures of hundreds of billions in the AI track.

The Third Hidden Reef: AI Commercialization "Praised but Not Popular"

Whether the AI wave can continue ultimately depends on a simple question: Can it actually make money?

Large models, as the most core monetization front-end of the AI industry chain, are the core benchmark for testing the industry's commercialization capabilities. Currently, domestic Zhipu, Baidu (09888.HK)'s Wenxin Yiyan, Alibaba's Tongyi Qianwen, and overseas GPT series, Claude, and other mainstream large models all face similar dilemmas: Single training costs reach tens of millions to hundreds of millions of dollars, daily inference computing power consumption, data iteration, and operation and maintenance costs continue to rise, but monetization models are highly concentrated in three paths: API calls, enterprise custom services, and member subscriptions.

More critically, most enterprises find it difficult to quantify their economic returns in the short term after introducing AI. For individual users, tolerance for monthly fees is limited, and conversion rates are generally low. While enterprise customers have more generous budgets, high budgets also imply high returns, and they often hope to accelerate the transformation of AI investment from a "cost item" to a "revenue item," which may be difficult to achieve in the short term.

Meanwhile, the cost of large model inference is dropping rapidly. This is both a blessing and a curse: Low costs make more application scenarios feasible, but it also continuously weakens the pricing power of large model enterprises, increasing the difficulty of their commercialization.

The Fourth Hidden Reef: The Interest Rate Hiking Cycle Makes AI Investment Expensive

The rapid expansion of the AI industry cannot be separated from the nourishment of the global low-interest-rate environment. Over the past few years, global loose monetary policy allowed tech companies to issue bonds, take loans, and raise funds at extremely low costs, recklessly adding to AI computing power construction and technology R&D. But as major central banks worldwide enter an interest rate hiking cycle, the dividend of looseness ends. This will cause the investment and financing costs of the entire AI industry chain to rise, and the return requirements for AI projects will also increase significantly.

Currently, the global bond market continues to be under pressure. High fiscal deficits in developed countries are pushing government bond yields to soar, and the overall financing interest rate in the market is rising comprehensively. At the same time, tech giants like Google, Amazon, and Meta are issuing large amounts of bonds to raise funds for continuous AI layout, which may further push up market interest rates. For AI enterprises, rising financing costs mean that the return requirements for every bit of computing power investment, R&D investment, and infrastructure investment will also be higher. Investors will also be more stringent when scrutinizing these projects.

The Fifth Hidden Reef: Worries about Overcapacity

During the 狂热 investment period, giants often fall into "scale worship," ignoring supply and demand laws. Core tracks such as memory chips, computing chips, optical modules, and servers have welcomed trillion-level capacity investments. This blind expansion is creating a serious crisis of supply-demand imbalance, which could become an important factor hindering the industry's continued upward movement.

Taking memory chips as an example, the trillions of dollars invested by giants are likely to evolve into a severe overcapacity problem in a few years. When the market is flooded with homogenized large models and excess hardware, price wars are inevitable, and profit margins will be extremely compressed.

The Sixth Hidden Reef: "Economic Backlash" Triggered by Job Replacement

The popularization of AI technology is progressing from replacing repetitive labor to replacing mid-to-high-end intellectual labor, triggering large-scale employment restructuring. Previously, the public generally believed that AI would only replace assembly line workers and basic service personnel. But now, code generation models can replace junior and intermediate programmers to complete programming work. AI office tools can replace copywriters, operations, finance, basic legal affairs, and other positions. Large tech companies batching layoffs has become the norm, with many high-paying intellectual jobs being replaced by AI.

This job replacement is not an individual phenomenon but a group-based, industry-wide structural change. The displaced workforce mostly bears rigid expenses such as mortgages, consumer credit, and family elderly care and child-rearing. Large-scale unemployment means income 断层 (breaks), directly triggering a series of economic chain reactions: Declining resident incomes lead to consumption contraction, rising mortgage default risks, and continuously rising credit default rates, thereby suppressing real estate and consumer market vitality and 催生 (generating) local deflationary pressures.

From the perspective of industry chain transmission, the continuous weakness in the downstream consumer market will gradually transmit upwards to the physical industry and the tech industry. Enterprises in all walks of life will face pressure on profitability and will weigh the cost-effectiveness of AI investments more cautiously, no longer blindly following the trend to add to AI layouts. The overall economic deflationary pressure will also force passive adjustments in global monetary policy, further affecting capital market liquidity, and constraining the continued expansion of the AI industry from the bottom layer.

Physical Limits and Abstract Barriers

Besides economic and social constraints, AI itself faces insurmountable physical and cognitive ceilings.

First is the rigid bottleneck of physical infrastructure. Training and inference of large models are truly "electric tigers." It is estimated that by 2030, global data center electricity consumption will double. Future constraints may no longer be chip supply, but gigawatt-level power to supply them.

Additionally, there is an uneliminable gap between AI's virtual simulation and the real material world. Digital models can only fit correlations and cannot fully replicate physical causality such as thermal deformation and material stress. In fields with extremely low fault tolerance like manufacturing and aviation, AI's "hallucinations" could bring fatal risks.

At the cognitive level, current AI systems essentially still only work within human cognitive frameworks, reciting text patterns, but lack understanding of causal logic in the real world. They cannot invent new concepts and create new things through interaction with the physical world like humans do. Without a continuous stream of fresh real data, even expanding model parameters might trap them in a dead end of repeated fitting.

Insights: Returning from Fanaticism to Reverence

What blocks the AI wave is never a single technical bottleneck, but a giant web woven from regulation, capital, economy, society, physics, and cognition. This brings us profound insights: The key inflection point of a technological revolution is often not the moment of technological enhancement, but when society and enterprises finally find new organizational forms compatible with the new technology.

Half a century after the invention of the steam engine, the factory system truly released its productivity; decades after the advent of electricity, the assembly line completely changed manufacturing. AI is the same—its true inflection point may not be the day of computing power breakthrough, but the day we finally learn how to organize talent, reconstruct processes, and design incentive mechanisms, allowing AI and humans to perform their respective duties and co-evolve.

Regulation must keep up, capital must calculate accounts, business models must be verified, employment must transform, and physical bottlenecks must be broken through—none of these are accomplished overnight. But precisely these "resistances" allow AI's development to return from restlessness to steadiness, and from speculation to value.

By: Wu Yan