Authors: Block Analytics Ltd X Merkle 3s Capital

We have already answered this question three times.

Is there a bubble in AI?

This is the question we've been asked most often over the past two years, and we've written about it more than once. Every time we give a conclusion, we're forced to re-examine it due to new surges and crashes.

This time, we don't intend to give a simple "yes" or "no" answer.

The question itself is flawed. AI isn't an asset; it's an entire industry chain—from wafer fabs to power plants, from trillion-dollar giants to newly funded startups. Asking "Is there a bubble in AI?" is as crude as asking "Is there a bubble in real estate?": can a prime location in a first-tier city and a ghost town in a third-tier county have the same answer?

Applying the same question to all levels will inevitably result in a wrong answer.

The correct question is: At what level is the AI ​​bubble?

Bubbles never ask "whether they exist," they only ask "where they are and how thick they are."

If you break this down, you'll see a picture that contradicts your intuition: the layer where everyone is focused on the worries is precisely the safest; while the areas where things are really bubbling up are rarely discussed seriously.

The Ghost of 2000: What's Different This Time?

When discussing the AI ​​bubble, the year 2000 is unavoidable. But most people only remember that "the dot-com bubble burst," without remembering how it burst .

The script from back then: first, establish a stock price, then find revenue.

The collapse of 2000 unfolded like this: Telecom companies borrowed massive amounts of money, frantically laying fiber optic cables, like building an eight-lane highway for a ghost town. The road was finished, but where were the cars? None. 85% to 95% of the fiber optic cables laid that year were "dark" —lying underground, never transmitting a single bit. Assets were on the books, revenue was zero, and the debt was real. Then, bang.

Fiber optics is just the story of the infrastructure layer. The application layer is even more absurd.

The most famous pet supplies e-commerce company at the time had only a few million dollars in revenue the year it went public, while its marketing expenses were several times its revenue—it spent money on advertising during the Super Bowl, losing money on every sale, and the more it sold, the faster it lost money. About nine months after its IPO, it went into liquidation and went bankrupt. This was not an isolated case; it was the standard profile of the application layer at that time: zero profit, relying on financing to stay afloat, and using "eyeballs" and "clicks" instead of revenue to value itself .

Even more surreal is that some scholars once calculated that a company's stock price could rise significantly on average simply by changing its name and adding ".com" to the end, without changing any of its business operations.

The market is paying for suffixes, not for business.

Let's look back at the "shovel sellers" of yesteryear. Cisco was Nvidia in 2000—internet traffic all passed through its routers, a seemingly impeccable logic. But at the peak of the bubble, Cisco's price-to-earnings ratio soared to triple digits . What does that mean? It meant the market demanded that it maintain its then-current profit levels for over a century, or grow tenfold or more within a few years, for the deal to break even. Later, the internet truly changed the world, and traffic exploded—it took Cisco's stock price over two decades to return to its 2000 peak.

Remember this case; it's the most important footnote in the entire text:

The biggest tragedy back then wasn't buying a fake company, but buying a real company for a hundred times the price.

The current script: generate revenue first, then increase stock price.

Now let's switch the focus to 2026.

Not a single GPU is idle . Every chip produced is immediately racked up the moment it rolls off the production line, running tokens at full capacity to earn real money. It's not just high utilization, it's 100%; customers are lining up with money but still can't buy one.

What about the application layer? Let's compare it to leading large-scale companies. One leading player's annualized revenue was less than $100 million 18 months ago, but now it's $45-47 billion , and it has already achieved quarterly profitability. Management originally planned for 10-fold growth, but it actually achieved 80-fold growth.

Compare the leading companies from the two eras side by side:

• Back then: Revenue of several million, losses of tens of millions, and bankruptcy after nine months on the market.

• Now: My income has increased several hundred times in 18 months, and I've already started making money.

Back then, companies relied on "stories" to raise money from the capital market; now, leading companies collect money from customers through contracts. This isn't a difference in degree, it's a difference in business models.

The valuation logic for "shovel sellers" has also changed. Today, Nvidia's price-to-earnings ratio is around 30—only a fraction of Cisco's peak. And what supports this valuation is not imagination of the future, but the backlog of orders that have already been signed and written into the production schedule.

Back then, the priority was stock price, then revenue – a process that often led to failure. Now, the priority is revenue, then stock price increases – a process that's easier to catch up with. The order is different, the outcome is different.

The buyers have changed. In 2000, the companies laying fiber optic cables were debt-ridden telecom companies; today, the companies buying computing power are Microsoft, Google, Meta, and Amazon—the four companies with the thickest cash flow on Earth, spending their own earned money.

In 2000, borrowed money was used to buy assets that no one used; in 2026, earned money will be used to buy assets that are not enough—these are two different species!

However, there was a crack in the wall.

At this point, we must put on the brakes.

This "free cash flow" story is beginning to wear down. The four major cloud vendors' combined capital expenditures this year totaled approximately $725 billion, a staggering 77% increase year-on-year. What kind of scale is that? It's roughly equivalent to the entire annual GDP of a moderately developed country, poured into data centers.

Even more striking is Amazon: its free cash flow plummeted from $26 billion to $1.2 billion, almost to zero, while its long-term debt is climbing. In other words, the giants' own earnings are quickly becoming insufficient, and they've started borrowing.

This is not a sign of a bursting bubble—the balance sheets of these giants remain among the most robust in human business history. But it is the first crack in the wall: the most solid logic of this round, "cash flow buyers," is sliding from "fully valid" to "largely valid."

It's worth keeping an eye on every quarter.

Let's conclude our review of 2000. The biggest misconception that bubble left for posterity is that everyone remembered "the story was fake," forgetting that what truly killed the market was uncontrolled supply : no matter how real the story seemed, as long as everyone on the supply side could infinitely leverage and expand production capacity, overcapacity was only a matter of time, and a collapse was simply a matter of mathematics. Conversely, judging whether this round will repeat the same mistakes depends not on how appealing the demand-side story is, but on whether anyone on the supply side can apply the brakes.

This leads to the next question: In this round, who is putting the brakes on?

First, release the map; then, clear the mines layer by layer: The five-layer pyramid of AI computing power

Before listing them one by one, let's first outline the entire industry chain. The AI ​​computing power industry chain can be divided into five layers from bottom to top:

Let's repeat it using a table:

There's a pattern in this image that's immediately obvious:

The closer you are to physics, the less hype there is; the closer you are to a story, the more hype there is.

At Level 0, expanding production takes three to five years, and building a factory requires investing hundreds of billions of dollars, making it impossible to create a bubble—supply simply doesn't cooperate. The higher you go, the looser the physical constraints become, and the greater the narrative space: at the long tail of Level 4, a single PowerPoint presentation can secure funding, and bubbles naturally accumulate there.

The only exception is the L2 interconnect layer—it's hardware, and should theoretically be protected by physical constraints, yet it's become the area with the strongest bubble. Why? We'll break it down in detail later.

The first step in judging an AI bubble is not to look at market sentiment, but to figure out where you stand in the pyramid.

The reason why layer L0 can be directly marked "no foam" on this map is because it is locked by two physical locks. Let's talk about the locks first, and then clear the mines layer by layer.

The first lock: TSMC

Why do we believe this round of AI capital expenditure won't get out of control? The answer lies not on the demand side, but on the supply side.

There's a necessary condition for a bubble to burst: oversupply . Tulips need to be planted everywhere, fiber optic cables need to be laid out so much that no one uses them, and houses need to be built so many that they can't be sold. Without oversupply, there's no collapse. The real culprit behind the disaster of 2000 wasn't that the internet story was wrong, but that the supply of fiber optic cables was completely out of control—any telecom company could borrow money to dig trenches and lay cables, and no one could apply the brakes.

The supply of AI computing power is in the hands of a group of the most conservative people in the world.

The "Central Bank" of the AI ​​Era

TSMC holds over 90% market share in advanced process technologies, maintaining a lead of approximately 9 to 15 months over Intel and Samsung, and this gap shows no signs of narrowing, even in the most advanced 2nm process. This implies one thing: global AI chip production is not determined by the market, but by TSMC.

It's like a central bank in the AI ​​era—the Federal Reserve controls how much money is printed, while TSMC controls how much computing power is printed. The Federal Reserve needs to hold meetings, vote, and face political pressure to raise interest rates; TSMC controls the supply of computing power by simply not approving its expansion plans.

The heads of this "central bank" are a group of engineers in their seventies who experienced the crises of 2001 and 2008. They see themselves as guardians of the founder's legacy, having witnessed firsthand how the semiconductor bubble inflated and how it buried the entire industry. In their memory, the "crash after the boom" is not a textbook example, but rather employees they personally laid off and production lines they watched shut down.

So when Huang Renxun approached them, demanding that production capacity double or even triple—they refused.

Think about how counterintuitive this is: the hottest company on Earth comes knocking on your door with endless orders and cash, begging you to expand production, and you say no. Only one company in the world can say this "no," and only one company has the final say .

Incidentally, here's a detail: Jensen Huang and TSMC have collaborated for over thirty years, and they've never signed a formal procurement contract . It's all based on handshakes. This isn't a management loophole; it's a system built on thirty years of trust—which is why TSMC dares to say "no" to its biggest customer, and the biggest customer has no choice but to accept.

How tight is this lock?

Digital level:

• The most advanced 2-nanometer process technology has sold out its entire production capacity by the end of this year; not a single unit remains.

• Kaohsiung is simultaneously building five 2-nanometer wafer fabs—the largest-scale parallel construction of advanced process technology fabs in human history. However, it takes three to five years for an advanced wafer fab to go from groundbreaking to mass production, with an initial investment of over US$20 billion.

• Even with such frantic construction, by 2030, the monthly demand for 2nm wafers is projected to be 400,000-450,000 wafers, while production capacity will only be 300,000-350,000 wafers—a long-term shortfall of 100,000-150,000 wafers per month, meaning that one-quarter to one-third of the demand will never be met.

There is another, more hidden bottleneck: advanced packaging. Chips are only semi-finished products once they are manufactured. The computing chip and memory need to be "packaged" together before they can be used—this is the "last mile" of AI chips, and this path is also basically controlled by TSMC alone, whose production capacity is always in short supply.

If TSMC were to fully unleash its potential, Nvidia could theoretically ship $2 to $3 trillion worth of GPUs annually—a figure nearly ten times its current actual shipment volume. It is TSMC that has locked in that number.

All the AI ​​ambitions in the world combined would have to queue up in front of TSMC's production capacity list.

This lock could also be picked.

To be fair, let's also clarify the negative side. This lock isn't a perpetual motion machine; it has a script that can be broken: if someone—whether it's a visionary like Musk or an Intel desperate to turn things around—bypasses TSMC, builds their own super foundry cluster with the support of equipment manufacturers, and breaks the monopoly on advanced production capacity, then the discipline of capacity expansion will collapse.

At that time, every chip manufacturer will frantically expand its production capacity like telecom companies did in 2000, and the engine of oversupply will truly ignite.

The good news is: given the physical timeline for building the factory, this scenario is unlikely to materialize before 2027. The bad news is: once filming begins, there will be no trailers.

Bubbles need an uncontrolled supply. And the supply valve for AI is in the hands of a group of elderly people who have witnessed two crashes and rejected Jensen Huang!

The second lock: Electric

Even if TSMC comes to its senses tomorrow and goes on a massive expansion of production, the chips will still need to be installed somewhere.

This is the second lock: electricity and land.

Many people believe that the bottleneck of AI infrastructure is chips, but the real bottleneck right now is something much more basic— land approval and power grid access for data centers .

The absurdity of this lies in the mismatch of time scales. Designing a chip takes two years; building a data center takes two to three years; but providing sufficient power to a data center—building new power plants, expanding substations, laying high-voltage transmission lines, and completing environmental impact assessments and approvals—often takes at least five years. Chips are developed in nanometer-scale evolution, while power grids are planned in ten-year periods.

Chips are iterated monthly, while power grids are measured in decades—this is the biggest time lag in the AI ​​era.

So you see a bizarre sight: tech giants with budgets of tens of billions of dollars are searching the world for "land with electricity," like gold prospectors looking for water. They're buying land next to nuclear power plants, signing 20-year power purchase agreements, and even directly funding the restart of decommissioned nuclear reactors. Money isn't the issue; electricity is.

The power shortage is not expected to ease until 2027-2028 – the construction cycle of power plants and grids determines this timeline, and no amount of money can significantly reduce it.

The two locks stacked together had the effect of forcibly "flattening" the growth of AI computing power. Demand wanted to explode, but supply could only climb. Growth thus became slower, but also longer and more stable—a treatment that historical technological revolutions like railways, canals, and the internet never enjoyed. Those revolutions all experienced supply spiraling out of control first, followed by collapse.

Historically, every technological revolution has died from uncontrolled supply. AI is the first to have its rhythm forcibly thwarted by the laws of physics—that's its greatest stroke of luck.

A variable from space

Here we leave one long-term variable: space data centers.

The logic is futuristic yet hard-hitting—solar energy is unlimited and free in a sun-synchronous orbit; the satellite faces away from the deep space at temperatures below -200 degrees Celsius, making heat dissipation virtually cost-free. The envisioned form is: a solar panel at the front of the satellite, a standard server rack in the middle, and a radiator hundreds of meters long trailing behind. Multiple satellites are interconnected by lasers, forming a virtual data center floating in orbit.

The two most expensive things for terrestrial data centers—electricity and cooling—are free in space.

Timeline: Proof of concept may be seen within two years, and the investment logic for terrestrial data centers may begin to be shaken around 2030.

Remember this variable. It doesn't change anything right now, but it's a sword hanging over the entire L3 infrastructure layer—we'll use it shortly.

Where is the real bubble: Demolishing mines layer by layer along the pyramid

Now that we've finished explaining the two locks, let's go back to the five-layer map and proceed from bottom to top, layer by layer.

L0 + Application Layer Header: Large Cap – Expensive, but not a bubble.

Microsoft, Google, Meta, Amazon, Nvidia. This layer of capital expenditure corresponds to real contracts, real revenue, and full utilization.

Two numbers are enough.

First: AWS's signed but unfulfilled order backlog reached $360-370 billion in the first quarter, a year-on-year increase of over 90%—and this doesn't even include the additional $100 billion commitment from a leading AI lab. What does this mean? It means that even if AWS didn't sign a single new customer from today onwards, the work it already had would be enough to keep it busy for several years. These aren't projections; they are signed contracts.

The second example is the leading large-scale model company mentioned earlier—in 18 months, its revenue grew from less than 100 million to over 45 billion, and it became profitable every quarter. This growth rate is unparalleled in the history of human commerce.

There's another factor that few people consider: the economics of inference. Training a cutting-edge model is a pure investment, burning through cash without batting an eye; but once the model is trained, every time it's used and every token generated is revenue. According to current industry estimates, the inference revenue potential of a model over its entire lifecycle is approximately 5 to 10 times its pre-training investment . In other words, today's astronomical capital expenditures aren't buying a one-time product like the "model," but rather a "computing power tollbooth" for many years to come.

The tollbooth model has one key characteristic: the initial investment is terrifying, and the subsequent cash flow is overwhelming. This applies to highways, power grids, and telecommunications networks—provided there are actually cars on the road. And we've already confirmed: not a single GPU is idle, and every lane is full.

Is it expensive? Yes. Is it a bubble? A bubble is defined as a price deviating from fundamentals, while the fundamentals are catching up with the price at a rate of 80 times every 18 months.

Back then, valuations stood still while waiting for revenue, eventually leading to bankruptcy; now, revenue is catching up with valuations, and it's doing so.

To sum up the buyers at this level in one sentence: they're not betting on a story when they buy computing power; they have no other choice but to accept orders they've already secured. Without expanding production, they can't deliver on the signed contracts—this is capital expenditure driven by demand, not capital expenditure pulled by illusions.

L1 memory layer: a battleground between bulls and bears

Moving up one level, we come to memory chips. This is currently the most contentious battleground between bulls and bears.

Let me first explain why this layer is important. If the GPU is the chef, then memory (especially high-bandwidth memory HBM) is the prep station—no matter how fast the chef's knife skills are, it's all for nothing if the ingredients can't be served quickly enough. And AI inference is precisely a job that crazily relies on "prep speed": the larger the model and the longer the dialogue, the faster the demand for memory bandwidth increases compared to the demand for computing power.

The current situation: memory prices have increased by 60-70% in a year, and Micron's profit margin has soared from the historical average of 16% to 70%.

This number is alarming when viewed in historical context: Over the past 25 years, the memory industry has been notorious for its "pig cycle"—price surges, rampant capacity expansion, oversupply, price collapses, and collective losses, repeating endlessly. Every time a profit margin in this industry reaches 70%, it's followed by a financial disaster . Following the old script, it's time to liquidate and run.

However, the bulls argue that this demand isn't for restocking, but rather structural. AI inference will drive continued growth in demand for HBM, while memory manufacturers, having learned their lesson from cyclical changes over the past 25 years, are extremely cautious about expanding production this time—nobody wants to be the one to crash prices.

There's a structural change worth mentioning separately: after 25 years of bloody reshuffling, only three players remain in the global high-end memory market. In the 1990s, the industry had over 20 manufacturers, and price wars were rampant; today, these three oligopolies watch each other's expansion plans across the Pacific, none wanting to make the first move. This oligopolistic structure inherently brings with it production capacity discipline—the strongest structural reason that "this expansion won't get out of control," more reliable than any management statement.

Moreover, HBM is quietly "squeezing out" the production capacity of regular memory: on the same production line, the wafers cut for HBM produce far fewer units than those for regular memory. The stronger the demand for HBM, the tighter the supply of regular memory becomes, driving up prices across the entire industry—which is why even the price of regular memory sticks in your computer is rising.

An even more important statistic: Currently, only about 0.1% of the global population is truly using AI effectively . If this number were to reach 5%—that is, to transform from a "geek toy" into a "daily tool for ordinary white-collar workers"—the ceiling for memory demand would be sky-high.

The bears' logic is equally sound: the current price increase is driven by prices themselves, not by the volume of goods sold—hoarding, withholding sales, and buying when prices are rising rather than falling are typical signs of a supply-demand mismatch, not healthy demand.

A 70% profit margin could mark either the beginning of a new era or the climax of an old drama. Bulls are betting on "this time it's different"—and those five words happen to be the most expensive five words in investment history.

We won't draw conclusions on this point. It's a gambling table, not a bubble; there are real chips on both sides.

L2 Interconnect Layer: Optical Modules—The Scent of the Bubble Begins Here

We've finally arrived at the part we really want to emphasize. It's also the only "hardware exception" on that map.

Let me explain what an optical module is in 30 seconds. An AI data center contains tens of thousands of GPUs. They don't work independently; instead, they constantly exchange data and collaborate to compute the same model. The amount of "communication" between the chips is so great that copper wires simply can't handle it. Electrical signals must be converted into optical signals and transmitted via optical fibers. The small box responsible for "electrical-to-optical and optical-to-electrical" conversion is the optical module.

GPUs are the muscles, and optical modules are the blood vessels . The larger the cluster, the greater the demand for interconnects between chips—hence the booming AI market and the frenzied growth of optical modules. This industry logic is true; the entire optical module market is expected to grow by nearly 60% this year, and production capacity is indeed "sold out until 2028."

The logic is sound. But let's examine what each stock's price action has actually done.

The first one: Lumentum – the darling of the last bubble, and the leader of this bubble.

This company manufactures lasers and optical components—in short, the core "light source" in optical modules and optical communication systems. Its history is quite intriguing: its predecessor was one of the most famous star stocks during the 2000 optical communication bubble—that company's market capitalization once soared to over $100 billion, only to plummet by 99% after the bubble burst, becoming a textbook example of an "infrastructure bubble." Lumentum is a business spun off from that company.

For the next two decades, it was a rather uneventful company: supplying lasers for iPhone's facial recognition and components for telecommunications networks; a typical "good but boring" hardware company.

Then AI arrived. Data centers require massive amounts of high-speed lasers, and the new generation of technology that "integrates the optical path directly into the switch" has pushed it to the center stage again, with even Nvidia investing a substantial $2 billion in it. As a result, its stock price has increased more than tenfold in the past 12 months .

Is the business getting better? Yes, it really is. Orders are booked until 2028, that's a fact. But put these two numbers together: its revenue growth is projected to be several tens of percent annually for the next few years, while its stock price has risen by more than a thousand percent in a year. The market is already pricing it at several tens of times its annual revenue—while the normal level for a mature hardware company is three to five times.

The very center of the last bubble burst was light, and the strongest smell of this bubble is still light. History doesn't repeat itself, but it really does rhyme.

The second company: AAOI – someone who has fallen once, is standing on the same cliff again.

This company manufactures optical transceiver modules, primarily selling them to cloud vendors' data centers. Its history is equally intriguing: during the last wave of data center construction (around 2017), it was a top-performing stock—until its largest customer suddenly canceled orders and switched to other suppliers, causing its stock price to plummet by 90% in the following two years, and then struggled on the verge of losses for seven or eight years.

Then AI arrived, and the demand for next-generation high-speed optical modules exploded, bringing back old customers. As a result, the stock price increased more than fourfold within the year.

Note the difference between this company and Lumentum: Lumentum is at least an industry leader with a technological moat and the backing of Nvidia; AAOI is a second-tier vendor that has been unprofitable for most of the past decade, has extremely high customer concentration, and has already learned its lesson from order cancellations in the last round. Its surge is almost purely due to the buoyancy of the sector's upward trend.

The tide has already begun to sway . Last month, this sector experienced single-day double-digit drops more than once—AAOI fell by more than 10% in a single day, and the leading stocks followed suit, falling by 7%-10%. There was no substantial negative news; it was simply that high-priced shares were starting to loosen.

There is another risk that is rarely discussed: the technology route itself.

The industry is currently undergoing an architectural revolution: integrating optical components directly into chip packages instead of "independent little boxes plugged into switches"—a process known in the industry as co-packaged optics. Once this direction becomes mainstream, it will mean two things: First, "optical modules" will be gradually absorbed as an independent product form, and the dominant position will shift from module manufacturers to chip giants; second, the value in the chain will concentrate on the "core light source," and the profits in the assembly stage will be squeezed out.

To put it another way: For companies like Lumentum, which hold the key to lasers, this technological revolution presents more opportunities than risks—light sources will always be needed and are becoming increasingly valuable. But for module manufacturers like AAOI, which excel at assembly, it's a second sword hanging over their heads. Ironically, the market is currently pricing both types of companies almost equally—when the tide is high, nobody cares who's wearing swim trunks.

Within the same sector, some are selling irreplaceable light sources, while others are selling boxes that could be bypassed by architectural revolutions at any time—yet the stock price increases show no difference. This in itself is a characteristic of a bubble.

Let's summarize this : demand increased by nearly 60%, and stock prices rose four to ten times. What's the difference? It's that the market has discounted 2028 revenue into the 2026 stock price.

A sound narrative coupled with excessive pricing—that's the standard form of a bubble. It's not fake; it's so expensive that it leaves no room for future mistakes.

Why did the bubble specifically emerge at this level? The answer lies in understanding the pattern on that map: optical modules represent the link with the lowest physical barrier to entry in the entire hardware chain. Building a wafer fab requires hundreds of billions of dollars and five years, while expanding an optical module production line only takes a few hundred million dollars and a few quarters—it's the only piece of hardware whose supply can "cooperate" with speculation. When the supply side can't be controlled, the bubble has the opportunity to grow.

TSMC's lock-in cannot protect optical modules—because the production capacity of optical modules is precisely the only link in the entire chain that does not require TSMC's approval.

The repeated occurrence of double-digit daily plunges indicates that smart money is already lining up at the door.

L3 Infrastructure Layer: GPU Cloud Sub-Landlord – Surviving, but Relying on Others' Bottlenecks

In the past two years, a number of new cloud providers specializing in GPU leasing have emerged: they buy their own GPUs, build their own data centers, and then lease the computing power to companies that lack GPUs. In the industry, they are called NeoCloud—we prefer to call them "GPU sub-landlords."

They're thriving, and they're truly skilled: these people squeeze the most out of hardware like F1 drivers do with their cars, achieving GPU utilization rates that are 2-3 times higher than traditional second-tier suppliers. They can extract significantly more revenue from the same batch of cards.

The survival logic also holds true: the four major cloud providers' own capacity is simply insufficient, and the overflow demand must be met by someone. As long as the major premise of "computing power shortage" exists, sub-landlords will have business .

But please note the nature of this business: they are beneficiaries of bottlenecks, not holders of moats .

Think about their situation clearly: every dollar they earn essentially comes from the time lag when "major companies haven't kept up with their capacity expansion." However—the power bottleneck is expected to ease in 2027-2028; major companies' self-built data centers are being completed at the fastest pace in human history; and the foreshadowed space data centers, if implemented in the 2030s, will fundamentally undermine the scarcity of ground-based computing power.

The time difference will eventually be closed. The sub-landlord doesn't have a property certificate, only a lease with an unknown expiration date.

Moreover, this business has a structural weakness: its customers and lifeline are highly concentrated . Their cards come from the same chip giant, and their major customers are often just two or three AI companies. In some cases, the largest shareholder and the largest supplier are even the same company. The upstream controls your supply, the downstream controls your revenue, and you in the middle earn money from the "matching time difference"—this kind of business can be very profitable, but it doesn't justify the valuation of a "platform".

If you're making money off someone else's bottlenecks, you have to be prepared for the day those bottlenecks disappear.

This layer isn't a scam; the cash flow today is real. But the market is currently giving them high valuations, pricing in the permanence of a temporary state —this is a valuation error and is heading towards a bubble.

L4 application layer long tail + VC ecosystem: the place with the strongest bubble signals

Finally, climb to the top of the pyramid. This level needs to be viewed in two halves.

The top half—a few large model companies with real revenue—have already been mentioned; their revenue keeps up with their valuation, so I won't elaborate further.

The real problem lies in the long tail, and the VC ecosystem that fuels it. The most glaring issue is precisely this:

In the first quarter of this year, AI companies took the vast majority of global venture capital investment—more than 8 out of every 10 dollars of VC funding went to AI.

In 1999, at the height of the dot-com bubble, what was this percentage? Approximately one-third to four-tenths.

In other words, today's concentration of VC bets on a single theme is twice that of the peak of the biggest bubble in human history.

Moreover, the structure is extremely top-heavy: just four major deals alone consumed 65% of the total global VC funding for the quarter. Two-thirds of the world's venture capital investment in a single quarter went into the accounts of these four companies.

This creates a chain of transmission: top-tier star companies use real revenue to support sky-high valuations—that's fine; but thousands of long-tail startups with no revenue are using the valuation logic of the top companies to price themselves—"That company's stock increased 80 times in 18 months, why can't mine?"—that's the big problem. The game of "adding a .com and it goes up" in 1999 has today's version of "adding an AI Agent and it doubles."

What's more troublesome is that the way these long-tail companies die is already predictable. They won't die from product failure—the products might even be good. They'll die from valuation inversion : the money raised in the previous round at bubble prices runs out, and the next round of investors are only willing to pay at the current market value. Raising at the current market value means huge losses for the previous round of investors and the founding team losing all their shares—so negotiations break down, and the company is stuck between "valuation dignity" and "survival" until the money in its account is gone. Most of the companies from 1999 died in this way: not killed by the market, but choked to death by their own previous valuations.

There's another amplifier: the cost structure of this round of long-tail companies is more fragile than in 1999. Back then, internet startups burned through marketing expenses; cutting advertising allowed them to survive. Today, AI startups burn through computing power bills—if the model isn't used, the product stalls, and there's no way to cut that money. Revenue is the story, costs are rigid; this combination will lead to a faster death when capital recedes than in the last round.

Note that this does not contradict the statement that "Large cap has no foam"—

The top performers are backed by real revenue, while the long tail is only supported by stories. Bubbles are never in the biggest companies; bubbles are in the small companies that use the valuation logic of the biggest companies to price themselves.

Do you remember the real lesson of 1999? It wasn't "the internet is fake"—the internet was real, e-commerce was real, and the biggest e-commerce company survived and even dominated the world. The lesson was:

In a real technological revolution, you can still lose all your money—if you buy the wrong floor.

The bears aren't entirely wrong: Two attack lines worth considering before bed.

If you think we're just mindless bulls, please read on. There's something real in the bear camp, and this time it's sharper than most bulls are willing to admit.

The bears have two main lines of attack. On the surface, they are two separate topics, but if you dig deeper, you'll find that they are actually two sides of the same issue.

Attack Line 1: The War of Depreciation – How many years can your GPU actually last?

Let's start by using a relatable example to explain "depreciation".

Let's say you drive for a ride-hailing service and spent 300,000 yuan on your car. If you calculate the cost based on a 3-year lifespan, the annual cost is 100,000 yuan; if you calculate it based on a 6-year lifespan, the annual cost is only 50,000 yuan. Note: You haven't made any extra money, and the car is still the same car. You've only changed the accounting assumptions, and your book profit has increased by 50,000 yuan every year out of thin air .

Now let's replace the car with a GPU, and turn 300,000 into hundreds of billions of dollars.

Tech giants are collectively doing the same thing: extending the depreciation period for GPUs. Previously, it was generally 3-4 years, but now they're extending it to 5 or 6 years. Each year of extension makes current profits look much better. Short sellers estimate that this change could reduce depreciation by hundreds of billions of dollars over the next three years, potentially overestimating the current profits of some giants by more than 20%.

What does 20% mean? It means that one-fifth of the profit you see in the financial statements may just be a "gift of accounting assumptions" rather than the profit earned by the business itself.

The bulls' rebuttal also has merit: the depreciation period isn't something that can be changed arbitrarily. In inference scenarios, older GPUs are perfectly capable—training cutting-edge models requires the latest cards, but running everyday inference with a three-year-old card still runs at full capacity and still generates revenue. Following this logic, it's not an exaggeration to say that GPUs can be used for 10 or 15 years; the previous 3-year depreciation period was actually an underestimation.

Who's right? The honest answer is: it depends on Nvidia itself . The more dramatic the performance leap between the next two generations of products, the faster the older cards depreciate, and the more right the bears are; the slower the leap, the longer the lifespan of the older cards, and the more right the bulls are. With each new generation of products released, Nvidia is essentially voting a vote on its customers' balance sheets.

This is the most ironic scene in the world of AI finance: the more successful Nvidia's products are, the more suspicious its customers' financial statements become.

Attack Line Two: GPU Credit – Moving Debt to Unseen Places

The second line of attack has been updated and is more covert. It's not widely discussed, but we believe it's an order of magnitude more serious than the depreciation issue.

GPUs are already using complex off-table structures for data flow. Breaking it down, this structure works like this:

• Set up a shell company : establish a special purpose vehicle (SPV) – a shell company with no business other than "holding GPUs".

• Shell companies borrow money : Shell companies borrow money from private credit funds to buy tens of thousands of GPUs.

• Renting to GPU users : Shell companies lease GPUs to AI companies on a long-term basis, collect rent, and use the rent to pay off loans.

• The most ingenious part is this: the chip manufacturers themselves also invest in the shell company, becoming anchor investors.

Everyone got what they wanted: AI companies got the credit card without incurring debt; the debt was not visible on the balance sheets of the giants and AI companies; chip manufacturers locked in sales and also made investment returns; and private credit funds acquired high-interest assets.

A win-win situation for all four parties. There's just one small problem: the debt hasn't disappeared, it's just that nobody can see where it is .

This structure should remind you of something. In fact, it rhymes with two historical periods at the same time.

The first paragraph is from 2000. Few people remember that a key player in the telecom bubble was "manufacturer financing": equipment giants lent money to their customers, who then bought their equipment. On paper, sales were booming and growth curves were perfect; in reality, it was just money laundering—customers were using your money to buy your goods. When the bubble burst, these equipment manufacturers weren't left with profits, but with a mountain of uncollectible debt, suffering the most severe losses. Today's structure—"chip manufacturers investing money in shell companies, and the shell companies using that money to buy chips"—is essentially a direct sibling of that manufacturer financing from back then.

The second point is from 2008. The last time the entire financial system was keen on "packaging, layering, and moving risks to places where regulators and investors couldn't see them clearly" was before that crisis, with mortgage securitization. Back then, houses were being packaged; now, GPUs are being packaged.

When an industry starts paying its own customers to buy its own products, you should question every growth figure you see.

Depreciation is an accounting issue, and accounting issues can never burst a bubble; leverage is a financial issue, and historically every bubble has been burst by financial issues.

The two lines are actually one line.

Now, if you connect the two lines of attack, you will see the true destructive power of the bearish logic.

The essence of the depreciation controversy is: how many years can a GPU be used, and what is its residual value?

What is the collateral for GPU credit? Or is it the residual value of the GPU?

In other words, the shell companies' basis for borrowing billions of dollars is the assumption that "these GPUs will remain valuable and generate rental income for many years to come." If Nvidia's next-generation products take another leap in performance, the rental income of the old cards will plummet—the first to collapse will not be the giants (they can withstand it), but these shell companies and the private credit funds that lent money to them.

Then the question you'll have to ask becomes: How much has private credit expanded in recent years? And what other things have been crammed into it? That's another article for you.

The current structure is still small in scale, far from enough to cause a systemic problem—that's the truth . But even the most ardent bulls themselves list "large-scale leverage in GPU collateralized financing" as the number one risk signal in this cycle. When both bulls and bears are unusually pointing to the same place and saying "look there," that's something worth watching closely.

The moment the company crammed the GPU into the casing, 2026 began to smell like 2008 for the first time. Right now, it's just a hint—watch how fast it gets stronger.

Conclusion: Expensive, but the door is still locked.

Compress the entire text into a single image, which is still that pyramid:

No bubble (L0 + L4 leaders) : TSMC, Nvidia, the four major cloud providers, and leading large-scale companies. Real contracts, real revenue, full utilization, plus the physical locks of TSMC and the power grid. Expensive, but expensive does not equal a bubble.

The battle between bulls and bears (L1) : Memory. A 70% profit margin either marks the beginning of a new structural cycle or the climax of an old scenario; the table is already set.

The following sectors have a bubble-like feel (L2, L3, L4 long tail) : Optical modules – the only link in the entire hardware chain not protected by TSMC's capacity discipline, pricing 2026 revenue with 2028 revenue; GPU sub-landlords – treating temporary bottlenecks as permanent moats; VC ecosystem – the concentration of a single theme has reached twice the peak of 1999, with long-tail startups using the valuation logic of the top to price their stories.

Three potential pitfalls that truly need to be monitored :

• A revolution in algorithm efficiency . If one day, a smarter algorithm achieves the same effect with one-tenth of the computing power, the entire capital expenditure logic of "piling up computing power" will collapse overnight. This is the least probable but most devastating scenario.

• GPU credit leverage . Once off-balance-sheet structures, mortgage financing, and securitization are rolled out, cash flow buyers become leveraged buyers, and the script of 2000 will be repeated with the engine of 2008. This is currently the most realistic indication of this trend.

• TSMC has abandoned its conservative approach . Whether its monopoly is broken by competitors or it changes its mind and expands production wildly—the necessary conditions for a bubble are truly met only when supply gets out of control. This is the one that needs the most long-term monitoring.

Before any of these three things happened, AI was a technological revolution whose rhythm was forcibly restrained by the laws of physics: expensive, crowded, and locally heated, but with a solid foundation.

Finally, turn this map into three portable questions. The next time you see any AI-related target, whether it's a stock or a startup, ask these questions first:

First question: Which level of the pyramid is it on ? The closer to the physical world, the more secure it feels; the closer to the story, the more dangerous it is. If you can't say which level you're on, assume you're on the most dangerous level.

The second question: Is its revenue real, or is it "borrowed" from the valuation of leading companies ? The frequency with which the phrase "benchmarking against a certain company" appears is directly proportional to the concentration of the bubble.

The third question: Does it make money from structural factors or from bottlenecks ? Structural factors can be earned for many years, while bottlenecks have an expiration date—and that expiration date is usually much shorter than the time implied by the valuation.

If you can answer all three questions, then we can discuss the price.

Bubbles never tell you which layer they're about to burst. But at least you can choose not to stand on the layer where you're pricing yourself based on other people's stories.

The next time someone asks you "Is AI a bubble?", you can ask them in return: Which layer are you referring to?

Those 70-something-year-old engineers at TSMC are probably the only people on this planet who can stop the AI ​​bubble. And so far, they're still there.