The first wave of the AI bubble has burst.

History repeats itself, especially in the technology sector.

In 1999, the dot-com bubble swept across the globe. As long as a company added ".com" to its name, its stock price would soar, as investors believed the internet would change the world.

More than 20 years later, the same scenario is repeating itself in the AI sector. Any company tangentially related to AI, regardless of whether it has core technology or profitable operations, sees its stock price double. Many companies have even overdrawn the profit valuations for the next dozen or even hundreds of years.

But no matter how large a bubble is inflated, it will eventually burst.

When the dot-com bubble burst back then, the B2B sector with no viable business model collapsed first. Today, AI concept companies with unproven narratives have also lost their halo.

"Pure AI hype companies" have long been hitting consecutive daily limit downs. Even companies that do have AI businesses or products but have unverifiable narratives and unfulfillable performance have suffered massive drawdowns. For example, Unisound has retreated 90% from its all-time high.

The bursting of the bubble for AI concept companies that "have narratives but no delivered performance" stems from a shift in valuation logic.

The market has shifted from FOMO to a cautious stance, starting to recalculate the real value of enterprises, and the valuation logic has gradually shifted from "whether a company is involved in AI" to "how much actual profit the AI business has generated".

From a quantitative perspective, this round of bubble bursting does not target Forward P/E (forward price-to-earnings ratio), but Forward EPS (forward earnings per share). Stock price = valuation × earnings. At present, valuations have only seen a correction, while the bubble of overly optimistic earnings expectations for companies with unproven business narratives is being squeezed out rapidly.

Therefore, companies like Micron and SK Hynix, which have clear paths to fulfill their order and profit targets, have staged a strong rebound after valuation corrections, while many companies driven purely by narratives with uncertain earnings are seeing their market capitalizations collapse.

Just as a bull market does not advance in one single push, a bubble does not deflate in one single drop. This round of drawdown among AI concept companies that "have narratives but no delivered performance" may only be one phase in the long industrial cycle.

/ 01 / The forced AI hype narratives no longer hold water

Let's turn back the clock to 2000. The Nasdaq peaked on March 10, and fell below its 200-day moving average on April 14, plummeting 36% in just one month.

In this typical bear market trend, the first signal that emerged was that the B2B sector without viable business models collapsed first.

Back then, the internet rally was driven by growth potential. The B2C business model was inherently logical: the more users a platform had, the more obvious the scale effect would be, and it could recover its investment through charging users in the future. But for the B2B model, no sound logic or applicable scenarios emerged during that period. Therefore, consecutive limit downs in the B2B sector became the first domino to fall in the bursting of the dot-com bubble.

Essentially, this change reflected that the market finally woke up to the fact that not all internet business models could make money. Capital began to sell off sectors that could not even tell a convincing story, and the crash spread to application companies with zero revenue.

History does not repeat itself, but it often rhymes, and a similar story is unfolding in the AI sector.

In the past few years, AI-related stocks have risen like crazy. Any company tangentially related to AI, regardless of whether it has core technology or profitable operations, saw its stock price double. As a result, under the banner of "AI technology", countless traditional enterprises rushed to attach AI labels to their businesses, and some extreme cases even emerged where "1% of the business supports 100% of the tech valuation".

The more frenzied the market trend is, the more extreme phenomena it tends to produce.

In the past, companies packaged with AI concepts met at the peak of consecutive daily limit ups; since the beginning of this year, many of them have gathered at consecutive daily limit downs.

The companies that have seen the largest drawdowns are naturally "pure AI hype companies": they have no AI technology, no AI business, no AI orders, but only fabricate AI-related narratives, and some have even faced regulatory inquiries or filings for investigation. Most of these companies end up with consecutive limit downs.

What is really worth paying attention to is another category of AI narrative companies.

These enterprises are not completely built on thin air, as they do have related businesses, revenue or products, but their capital market narratives are full of huge uncertainties. A large number of traditional software and internet enterprises, from BlueFocus and Unisound to Huatu Education and Bilibili, have fallen into similar awkward situations.

In the early stage, they successfully drove their stock prices to double with certain AI products or revenue; however, since the beginning of this year, the ones with milder drawdowns have fallen 30% from their highs, and the ones with deeper drawdowns have even plummeted 90%.

The drawdowns of these companies happen because although their AI narratives are not necessarily proven false, they are extremely difficult to verify.

Take BlueFocus as an example. Its AI narrative is to transform from a traditional marketing company into an AI MarTech company, restructure its business model with AI, and has independently developed 136 marketing Agents for this purpose. This set of narratives once made BlueFocus regarded as a core target of the AI + marketing track, pushing its stock price up 1.6 times in a short period, with its P/E ratio once exceeding 200 times.

However, after the sharp rally, its maximum drawdown from the stock price high this year has exceeded 50%.

The reason is not hard to understand: the premise for this tech valuation narrative to hold is that AI can optimize BlueFocus' profit model or bring new revenue, but BlueFocus has not yet proved this delivery path.

Although it publicly claims that AI-driven revenue has reached 3.725 billion yuan, the so-called AI-driven revenue refers to the revenue of advertising placement projects where AI is deeply involved in execution, which is not equivalent to the revenue from selling AI technology externally. BlueFocus' profitability has not improved significantly either, with its net profit margin standing at only 0.32% in 2025.

The drawdowns of these companies also indicate that the first wave of the AI bubble has started to burst.

/ 02 / The bubble bursts at Forward EPS, not Forward P/E

Bubbles in the technology sector are often the rewards and tributes the society gives when facing disruptive advanced productive forces. That's why the capital market has been extremely generous in the past few years, willing to pay for the future imagination space.

However, when a large number of companies have overdrawn the profit valuations for the next dozens or even hundreds of years in advance, coupled with the fact that free cash flow of tech giants has turned negative, investors have finally become cautious from the previous frenzied FOMO state. The market has to recalculate the real value of an enterprise, which leads to the emergence of a new pricing logic: gradually shifting from "whether a company is involved in AI" to "how much actual profit the AI business has generated".

The quantitative manifestation of the new pricing logic is that the bubble bursts at Forward EPS, not Forward P/E. Specifically, stock price = valuation × earnings. In the current market, although valuations have seen corrections, what has been drastically squeezed out is the bubble of overly optimistic future earnings expectations that were driven purely by stories.

The proof is that companies like Micron and SK Hynix, which have clear paths to fulfill their order and profit targets, have staged a strong rebound after valuation corrections, while many companies driven purely by narratives with uncertain earnings are seeing their market capitalizations collapse.

Take Unisound as an example. In the past, the majority of its revenue came from AI solutions in the smart life sector, with clients including home appliance giants such as Midea and GREE, delivering products ranging from voice modules, edge-side chips to overall solutions.

After the AI boom, its narrative was upgraded from voice products to vertical industry large models + AI agents, trying to gradually transform its business model from one-off project delivery to high-margin revenue sources such as subscription and MaaS cloud invocation.

This narrative was very successful in the early stage of its listing, with its highest market capitalization hitting 62 billion Hong Kong dollars and its highest PS (TTM) exceeding 60 times. But now its market capitalization has fallen by 90%. There are two core reasons:

First, the narrative has no performance support. Although its financial report shows "large model-related revenue" of 610 million yuan, the market questions that

this is not purely new MaaS/subscription cloud invocation revenue, but mostly one-off project-based delivery, with a low proportion of standardized subscriptions. In terms of profitability, its long-term negative net profit margin also fails to prove that AI has significantly improved its business model.

Second, its original voice business is facing the impact of AI democratization. In the large model era, out-of-the-box ASR/TTS APIs have greatly lowered the threshold of voice technology, which may divert Unisound's customers and compress the profit margin of its products. The company's gross profit margin also declined in 2025.

It is worth alerting that the crisis of unfulfillable profits is by no means limited to the "storytellers" at the application layer. A large number of domestic AI hardware "concept stocks" that are benchmarked against overseas counterparts are also hanging with the Sword of Damocles.

The revenue of many AI hardware companies in the US stock market is relatively solid, as they receive the capital expenditure from Google, Amazon and Microsoft, with a very smooth transmission chain. Some domestic companies also operate under this logic, such as optical module manufacturers. But for many domestic AI concept stocks that have risen riding the AI trend, it is highly questionable whether their performance can really be delivered to such a high degree.

/ 03 / Gold shines after all the sand is washed away

Just as a bull market does not advance in one single push, a bubble does not deflate in one single drop. The current drawdown among AI concept companies that "have narratives but no delivered performance" may only be one phase in the long industrial cycle.

Even the most hardcore AI "picks and shovels" players at present have very stretched valuations, as their current prices are based on the most optimistic industry expectations and have overdrawn profits that may not even be achieved.

For example, memory stock represented by Micron, under the logic that AI-driven incremental demand has transformed memory from a strong cyclical commodity to a mandatory AI infrastructure that greatly improves earnings stability, has switched its valuation framework from PB to PE.

Micron currently has a 22x TTM P/E ratio, which does not look overly high. But the denominator of this P/E is the super-cycle earnings where the contract price of DDR4 has risen 10 times in 15 months, and the gross profit margin has jumped from 36% to 75%. Multiplying earnings at the cyclical peak by a seemingly "reasonable" multiple itself carries huge risks.

This means that even if there is no bubble in the AI industry, as long as the growth rate of memory chip demand slows down marginally, leading to the inability to continue raising product prices, it will deal a heavy blow to the current valuation model.

So how can we judge the bubble inflection point of such companies? A key observation perspective lies in the flow of debt.

In the current AI industry, more and more debt is being transferred to data center developers, private infrastructure funds, power and energy project financing, GPU leasing and cloud infrastructure companies.

Therefore, when the future hard tech AI bubble bursts, the first to collapse may not be stocks, but the bonds of industrial chain enterprises that provide infrastructure for CSPs and the AI ecosystem and rely on long-term contracts and debt financing. If the yield of high-yield bonds starts to surge, or large tech companies fail in bond issuance, that may be an important inflection point signal.

Of course, the process may be tortuous, but the path is clear. Just like after 2000, the human society entered the digital era where the internet is indispensable for all industries. Today, we are also irreversibly moving towards an intelligent era where all industries are dominated and empowered by AI.

Short-term bubble bursting and valuation correction are regular growing pains in the growth cycle of the technology industry, which cannot stop the industrial wave of AI. The bursting of the bubble is exactly the best time to screen out enterprises with real core advantages that are expected to grow into great companies.

After all, gold shines after all the sand is washed away. And in this process, we not only need the courage to embrace the future, but also calm judgment.

This article is from the WeChat official account "Read Finance", author: Read Finance, published with authorization from 36Kr.