Where the AI Bubble Really Is: Which Layer of Players Are Naked

Author: Block Analytics Ltd X Merkle 3s Capital

We've Answered This Question Three Times Already

Is there an AI bubble?

This has been one of the most frequently asked questions over the past two years, and we've written about it more than once. Each time we gave a conclusion, each time we had to re-examine it after new surges and crashes.

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

Because the question itself is wrong. AI is not an asset; it's an entire industrial chain—from wafer fabs to power plants, from trillion-dollar market cap giants to startups that just secured funding. Asking "Is there an AI bubble?" is as crude as asking "Is there a real estate bubble?": Can the answer be the same for prime locations in first-tier cities and ghost towns in eighteenth-tier counties?

Applying one question to all levels is bound to yield a wrong answer.

The correct question is: Where is the bubble in AI?

Bubbles never ask "if," only "where and how thick."

When you break down this question, you see a counterintuitive picture: the layer everyone is worried about is precisely the safest; the place where bubbles are truly forming is rarely seriously discussed.

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

Talking about the AI bubble inevitably leads to the year 2000. But most people only remember "the dot-com bubble burst," not how it burst.

The Script Back Then: Stock Price First, Revenue Later

The script for the 2000 collapse went like this: Telecom companies borrowed astronomical amounts of money, frantically laying fiber optic cables, like building an eight-lane highway for an empty city. The road was built, but where were the cars? Nowhere. Back then, 85% to 95% of the fiber laid was "dark"—lying underground, having never transmitted a single bit. The assets were on the books, revenue was zero, and the debt was real. Then, bang.

Fiber was just the infrastructure layer story. The application layer was even more absurd.

One of the most famous pet supply e-commerce companies back then had annual revenue of only a few million dollars the year it went public, while its marketing expenses were several times its revenue—it splurged on Super Bowl ads, losing money on every sale it made, losing faster the more it sold. It liquidated and went bankrupt about nine months after its IPO. This wasn't an isolated case; it was the standard profile of the application layer at the time: zero profits, reliant on financing to stay afloat, using "eyeballs" and "clicks" instead of revenue for valuation.

Even more surreal, scholars at the time found that: a company could simply change its name, adding ".com" at the end, change nothing about its business, and its stock price would, on average, jump significantly.

The market was paying for a suffix, not for a business.

Now look at the "pickaxe sellers" back then. Cisco was the Nvidia of 2000—all internet traffic had to pass through its routers, an impeccable logic. But at the peak of the bubble, Cisco's P/E ratio soared to triple digits. What does that mean? It meant the market required it to maintain its then-profit scale for over a hundred more years, or grow tenfold in a few years, for the investment to break even. Later, the internet truly changed the world, and traffic did explode—Cisco's stock price took over twenty years to return to its 2000 high.

Remember this case; it's the most important footnote of this article:

The greatest tragedy back then wasn't buying fake companies, but buying real companies at a hundred times the price.

The Script Now: Revenue First, Stock Price Later

Now cut to 2026.

Not a single GPU is dark. Every chip produced is plugged into a rack the moment it rolls off the line, running at full load generating tokens, exchanging for real cash. It's not high utilization; it's 100%, it's customers queuing with cash and still unable to buy.

The application layer? Take a leading large model company for comparison. One leading player's annualized revenue was less than $1 billion 18 months ago; now it's $45-47 billion, and it's already achieving quarterly profitability. Management originally planned for 10x growth, but actually achieved 80x.

Put the leading companies from both eras side by side:

• Then: Revenue of a few million, losses of tens of millions, bankrupt within nine months of IPO.

• Now: Revenue multiplying hundreds of times in 18 months, already turning a profit.

Companies back then relied on "stories" to ask capital markets for money; leading companies today rely on contracts to collect money from customers. This isn't a difference in degree; it's a difference in business model.

The "pickaxe sellers" have also changed their valuation logic. Nvidia's P/E ratio today is around the low thirties—only a fraction of Cisco's peak back then. And what supports this valuation isn't imagination about the future, but order backlogs already signed and written into production schedules.

Then, it was stock price first, then find revenue, find it to death; now, it's revenue first, then stock price rises, and it catches up. The sequence is different; the outcome is different.

The buyers have changed too. In 2000, it was debt-laden telecom companies laying fiber; today, it's Microsoft, Google, Meta, Amazon buying compute—the four companies on the planet with the thickest cash flows, spending money they earned themselves.

2000 was borrowed money buying unused assets; 2026 is earned money buying insufficient assets—these are two different species!

But, There's a Crack in the Wall

We must hit the brakes here.

This "free cash flow" story is already starting to fray at the edges. The combined capital expenditure of the four major cloud providers this year is about $725 billion, a staggering 77% increase year-over-year. What scale is this? Roughly equivalent to the GDP of a medium-sized developed country for an entire year, poured into data centers.

Even more glaring is Amazon: free cash flow plummeted from $26 billion to $1.2 billion, almost to zero, while long-term debt is creeping up. In other words, the money the giants themselves earn is almost insufficient to fuel the burn; they're starting to borrow.

This isn't a signal of a bubble bursting—the giants' balance sheets are still among the most solid in human commercial history. But it's the first crack in the wall: the hard logic of "cash flow buyers," the hardest logic of this cycle, is sliding from "fully valid" to "broadly valid."

Worth watching every quarter.

To wrap up the 2000 review: the biggest misdirection left by that bubble is making everyone remember "the story was fake," but forgetting that what truly killed the market was uncontrolled supply: no matter how true the story is, as long as everyone on the supply side can infinitely leverage and expand capacity, oversupply is a matter of time, and collapse is a matter of mathematics. Conversely, judging whether this round will repeat the mistake doesn't lie in how compelling the demand-side story is, but in whether anyone on the supply side can hit the brakes.

This leads to the next question: This round, whose foot is on the brake?

First, the Map, Then Layer-by-Layer Demining: The Five-Layer Pyramid of AI Compute

Before naming names, let's map out the entire industry chain. The AI compute industry chain, from bottom to top, can be sliced into five layers:

Let's say it again with a table:

This diagram reveals a pattern at a glance:

The closer to physics, the less bubble; the closer to story, the more bubble.

At the L0 layer, expanding production takes three to five years, building a fab costs tens of billions of dollars—it's impossible to blow a bubble even if you want to, supply simply doesn't cooperate. The higher you go, the looser the physical constraints, the larger the narrative space: by the long tail of L4, a PowerPoint deck can secure funding, bubbles naturally gather there.

The only exception is the L2 interconnect layer—it's clearly hardware, should be protected by physical constraints, but has become the place with the strongest scent of bubble. Why? We'll dissect it later.

The first step in judging an AI bubble is not to look at market sentiment, but to first see clearly which layer of the pyramid you're standing on.

In this map, the reason L0 dares to directly label "No Bubble" is because it's locked by two physical locks. Let's talk about the locks first, then demine layer by layer.

The First Lock: TSMC

Why do we judge that this round of AI capital expenditure won't spiral out of control? The answer isn't on the demand side, but the supply side.

Bubble bursting has a necessary condition: oversupply. Tulips have to be planted everywhere, fiber has to be laid with no one using it, houses have to be built that can't be sold. Without oversupply, there's no crash. The real culprit of the 2000 disaster wasn't that the internet story was wrong, but that fiber supply was completely out of control—any telecom company could borrow money to dig trenches and lay cables; no one could hit the brakes.

And the supply of AI compute is held in the hands of some of the world's most conservative people.

The "Central Bank" of the AI Era

TSMC's market share in advanced process nodes exceeds 90%, with a lead of about 9 to 15 months over Intel and Samsung, and this gap shows no signs of narrowing at the most advanced 2-nanometer node. This means one thing: global AI chip production isn't determined by the market; it's determined by TSMC.

It's like the central bank of the AI era—the Fed controls how much money is printed, TSMC controls how much compute is printed. The Fed holds meetings, votes, and faces political pressure to raise rates; TSMC controls compute supply simply by not nodding on expansion plans.

And the "governors" of this "central bank" are a group of engineers in their seventies who experienced 2001 and 2008. They see themselves as guardians of the founder's legacy; they've personally seen how semiconductor bubbles inflated and buried the entire industry. In their memory, "the crash after the surge" isn't a textbook case; it's the employees they had to lay off, the production lines they had to shut down with their own eyes.

So when Jensen Huang came knocking, asking to double or even triple capacity—they said no.

Think about how counterintuitive this is: the hottest company on the planet, with unlimited orders and cash, comes to you, begs you to expand production, and you say no. This kind of "no" can only be uttered by one company in the world, and only one company's "no" counts.

By the way, a detail: Jensen Huang and TSMC have collaborated for over thirty years, never having signed a formal purchase contract. All on a handshake. This isn't a management loophole; it's a system built on thirty years of accumulated trust—and also why TSMC dares to say "no" to its biggest customer, and the biggest customer can only accept it.

How Tight is This Lock

On the numbers:

• The most advanced 2nm process capacity for this year-end is already completely sold out, not a wafer left.

• Kaohsiung is simultaneously building five 2nm fabs—the largest-scale parallel construction of advanced process fabs in human history, but building one advanced fab from ground-breaking to mass production takes three to five years, with an upfront investment exceeding twenty billion dollars.

• Even building like crazy, by 2030, the estimated monthly demand for 2nm is 400k-450k wafers, but capacity will only be 300k-350k wafers—a long-term shortage of 100k-150k wafers/month, equivalent to one-quarter to one-third of demand never being met.

There's an even more hidden bottleneck: advanced packaging. Chips are just semi-finished products after fabrication; they need to be "packaged" with memory to be usable—this is the "last mile" for AI chips, and this road is also essentially guarded by TSMC alone, with capacity also chronically insufficient.

If TSMC completely let go, Nvidia could theoretically ship $2 to $3 trillion worth of GPUs a year—this figure is close to ten times the current actual shipment scale. It's TSMC that has locked down this number.

The combined AI ambitions of the entire world have to queue up in front of TSMC's capacity sheet.

This Lock Could Also Be Picked

To be fair, let's also lay out the opposite scenario. This lock isn't a perpetual motion machine; there's a script for it being picked: if someone—whether a Musk-like madman or an Intel desperate for a comeback—bypasses TSMC, builds a super-fab cluster with support from equipment makers, breaks the monopoly on advanced capacity, then the discipline of capacity expansion would collapse.

At that point, every chipmaker would expand capacity like the telecom companies of 2000, and the engine of oversupply would truly ignite.

The good news: The physical cycle of fab construction is there; this script has almost no chance of playing out before 2027. The bad news: Once this script starts filming, there won't be a trailer.

Bubbles need runaway supply. And AI's supply valve is held by a group of old men who've seen two crashes and said no to Jensen Huang!

The Second Lock: Electricity

Even if TSMC decided tomorrow to expand like crazy, the chips produced would need a place to be plugged in.

That's the second lock: electricity and land.

Many think the bottleneck for AI infrastructure is chips, but the real choke point right now is something more mundane—land approval and grid access for data centers.

The absurdity of this lies in the mismatch of time scales. Designing a chip, two years; building a data center, two to three years; but powering a data center with enough electricity—building new power plants, expanding substations, pulling high-voltage transmission lines, completing environmental reviews and approvals—easily takes five years or more. Chips evolve by nanometers; power grids plan by decades.

Chips iterate by the month, power grids by the decade—this is the biggest time lag of the AI era.

So you see a bizarre scene: tech giants holding tens of billions in budgets scour the world for "land with electricity," like prospectors looking for water. Buying land next to nuclear plants, signing twenty-year power purchase agreements, even directly funding the restart of decommissioned nuclear reactors. Money isn't the problem; electricity is.

The power gap isn't expected to ease until 2027-2028—the construction cycle of power plants and grids determines this timeline; no amount of money can compress it much.

The two locks together have this effect: AI compute growth is forcibly "flattened." Demand wants to explode, supply can only crawl. Growth thus becomes slower, but also longer, more stable—this is precisely the treatment that historical technological revolutions like railroads, canals, and the internet never enjoyed. They all saw supply go out of control first, then crash.

Every technological revolution in history died from runaway supply. AI is the first one whose rhythm is forcibly restrained by physical laws—this is its greatest piece of luck.

A Variable from Space

Let's leave a long-term variable here: space data centers.

The logic is sci-fi but solid—in sun-synchronous orbit, solar power is infinite, free; the satellite's shaded side faces deep space at minus two hundred degrees Celsius, cooling costs are near zero. The envisioned form: the satellite's front end is solar panels, the middle is standard server racks, the tail drags a radiator hundreds of meters long, multiple satellites interconnected with lasers, piecing together a virtual data center floating in orbit.

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

Timeline: Proof-of-concept might be seen within two years, potentially starting to shake the investment logic of ground data centers around 2030.

Remember this variable. It can't change anything yet, but it's a sword hanging over the entire L3 infrastructure layer—we'll come back to this immediately below.

Where the Bubble Really Is: Demining the Pyramid Layer by Layer

The two locks are explained. Back to the five-layer map, from bottom to top, layer by layer.

L0 + Application Layer Head: Large Cap—Expensive, But Not a Bubble

Microsoft, Google, Meta, Amazon, Nvidia. The capital expenditure in this layer corresponds to real contracts, real revenue, full-load utilization.

Two numbers are enough.

First: AWS's signed but not yet executed backlog reached $360-370 billion in Q1, up over 90% year-over-year—and this doesn't include the additional $100 billion commitment from a certain leading AI lab later. What does this mean? It means even if AWS signed no new customers from today, the work already signed is enough to keep it busy for several years. These aren't expectations; these are signed contracts.

Second: The leading large model company mentioned earlier—18 months, revenue from under $1 billion to over $45 billion, already profitable quarterly. This growth rate has no second sample in human commercial history.

There's another calculation few people make: the economics of inference. Training a frontier model is pure investment, burning money without blinking; but once the model is trained, every time it's called, every token generated, is revenue. According to current industry estimates, the inference revenue opportunity over a model's entire lifecycle is roughly 5 to 10 times its pre-training investment. That means today's astronomical capital expenditure isn't buying a one-time "model" product; it's buying a "compute toll booth" for many years to come.

The toll booth model has a characteristic: the upfront investment is terrifying, the later cash flow drowns you. Highways, power grids, telecom networks are all like this—provided there are really cars running on them. And we've already confirmed: not a single GPU is dark, every lane is full.

Expensive? Yes. Is it a bubble? The definition of a bubble is price detaching from fundamentals, and the fundamentals in this layer are chasing the price at a speed of 80x every 18 months.

Back then, valuations stood still waiting for revenue, waited until bankruptcy; now, revenue is chasing valuations, and it's catching up.

Summarize the buyers in this layer in one sentence: they aren't betting on a story by buying compute; they are left with no choice in the face of already secured orders. If they don't expand, they can't deliver the contracts they've signed—this is capital expenditure pushed by demand, not pulled by illusion.

L1 Memory Layer: A Long-Short Battleground

One layer up, memory chips. This is currently the most fiercely contested battlefield between bulls and bears.

First, explain why this layer is important. If GPUs are the chefs, memory (especially High Bandwidth Memory - HBM) is the prep station—no matter how fast the chef's knife skills are, if ingredients aren't handed over, it's useless. And AI inference is precisely a task that voraciously consumes "prep speed": the larger the model, the longer the conversation, the demand for memory bandwidth grows even faster than the demand for compute.

Current situation: Memory prices have risen 60-70% in a year, Micron's profit margins have soared from a historical average of 16% to 70%.

How scary is this number in historical context? Over the past twenty-five years, the memory industry has been notoriously cyclical—price rises, crazy expansion, oversupply, price crash, collective losses, rinse and repeat. Profit margins at the 70% level in this industry have, every time they appeared, been followed by a funeral. According to the old script, one should sell and run now.

But the bull logic is: This time the demand isn't inventory restocking; it's structural. AI inference's demand for HBM will continue to increase, and memory manufacturers, having been taught by the cycle for twenty-five years, are expanding extremely cautiously this time—no one wants to be the one crashing prices.

There's a structural change worth mentioning separately: After twenty-five years of bloody consolidation, the global high-end memory market is down to three players. In the 1990s, this industry had over twenty manufacturers; when price wars started, no one could stop. Today, three oligarchs watch each other's expansion plans across the Pacific, none wanting to make the first move. An oligopoly naturally comes with capacity discipline—this is the hardest structural reason why "expansion won't go out of control this time," more reliable than any management statement.

Moreover, HBM is quietly "crowding out" capacity for regular memory: on the same production line, wafers allocated to HBM yield far fewer finished units than regular memory. The stronger the HBM demand, the tighter the supply of regular memory, pushing up prices for the entire industry—that's why even the price of the regular RAM sticks in your computer is rising.

An even more important number: Currently, only about 0.1% of the global population truly uses AI correctly. If this number moves towards 5%—that is, from "geek toy" to "daily tool for ordinary office workers"—the demand ceiling for memory is above the clouds.

The bear logic is equally solid: The current price increase is driven by price itself, not shipment volume—hoarding, withholding supply, buying on the rise not the fall, this is a classic signal of supply-demand mismatch, not healthy demand.

A 70% profit margin is either the starting point of a new structural era or the climax of the old script. Bulls are betting on "this time is different"—and these five words happen to be the most expensive five words in investment history.

We won't draw a conclusion on this layer. It's a gambling table, not a bubble; both sides have real chips.

L2 Interconnect Layer: Optical Modules—The Bubble Scent Starts Here

Finally, we reach where we really want to underline. Also the only "hardware exception" on that map.

First, explain what an optical module is in thirty seconds. An AI data center has tens of thousands of GPUs; they don't work in isolation but must constantly exchange data, collaboratively computing the same model—the "conversation volume" between chips is so large that copper wires simply can't handle it, necessitating the conversion of electrical signals to optical signals transmitted via fiber. The little box responsible for "electrical-to-optical, optical-to-electrical" conversion is the optical module.

GPUs are muscles, optical modules are blood vessels. The larger the cluster scale, the demand for inter-chip interconnection increases by the square—so the hotter AI is, the crazier optical modules become. This industry logic is real; the entire optical module market is expected to grow close to 60% this year, capacity truly "sold out until 2028."

The logic is real. But let's see what the stock prices have done, company by company.

First: Lumentum—The Darling of the Last Bubble, the Frontrunner of This One

This company makes lasers and optical components, essentially the core "light source" in optical modules and optical communication systems. Its pedigree is very telling: its predecessor was one of the most famous star stocks of the 2000 optical communication bubble—that company's market cap once soared to over $100 billion, dropped 99% after the bubble burst, becoming the standard illustration of "infrastructure bubble" in textbooks. Lumentum is the business spun out from that company.

For twenty years in between, it lived a dull life: supplying lasers for iPhone Face ID, components for telecom networks, a typical "good but boring" hardware company.

Then AI came. Data centers need massive quantities of high-speed lasers, a new technology path that "integrates optical circuits directly into switches" pushed it to center stage, and even Nvidia invested $2 billion in it. Result: Stock price up over 10x in the past 12 months.

Is the business improving? It really is. Orders backlogged to 2028, that's real. But please put two numbers together: its revenue growth expectation is tens of percent per year for the next few years, while its stock price rose over a thousand percent in one year. The market's pricing for it is already dozens of times its annual revenue—while the normal level for a mature hardware company is three to five times.

The epicenter of the last bubble burst was optics; the place with the strongest bubble scent this round is still optics. History doesn't repeat, but it sure rhymes.

Second: AAOI—Someone Who Fell Once Before, Standing on the Same Cliff Again

This company makes optical transceiver module units, mainly sold to cloud providers' data centers. Its history is also telling: in the last data center construction wave (around 2017), it was also a big winner—until its largest customer suddenly cut orders and switched to other suppliers; its stock price dropped 90% in the following two years, struggling on the edge of losses for a full seven or eight years afterwards.

Then AI came, demand for new-generation high-speed optical modules exploded, old customers returned. Result: Stock price up over 4x year-to-date.

Note the difference between this company and Lumentum: Lumentum is at least an industry leader, with technological moats, backed by Nvidia; AAOI is a second-tier manufacturer that hasn't made money most of the past decade, has extremely high customer concentration, and was already taught a lesson by order cuts in the last cycle. Its surge is almost purely buoyed by the sector's tide.

And the tide has already started to wobble. Last month, this sector saw single-day double-digit declines more than once—AAOI down over 10% in a day, leaders also down 7%-10%. No substantive negative news, just high-level筹码 starting to loosen.

There's another layer of risk rarely discussed: the technological path itself.

The industry is currently pushing an architectural revolution: integrating optical components from "independent small boxes plugged into switches" directly into the chip package—known as co-packaged optics in the industry. Once this direction becomes mainstream, it means two things: First, the "optical module" as an independent product form will gradually be absorbed, with dominance shifting from module makers to chip giants; Second, value in the chain will concentrate towards the "core light source," margins in the assembly segment will be squeezed dry.

Translation: This technological shift is more opportunity than risk for a company like Lumentum holding the lasers—light sources are always needed, and become more valuable; but for module makers like AAOI that excel at assembly, it's a second sword hanging overhead. Ironically, the market's pricing enthusiasm for both types of companies now is almost indistinguishable—when the tide is big, no one checks who's wearing swim trunks.

Within the same sector, some sell irreplaceable light sources, others sell boxes that could be bypassed by architectural revolution at any moment—and the stock price gains show no difference. This itself is a characteristic of a bubble.

Let's sum up the accounts for this layer: Demand growth is close to 60%, stock prices are up fourfold to tenfold. What is the gap in between? It's the market discounting 2028 revenue into the 2026 stock price.

Correct narrative, plus excessive pricing—this is the standard form of a bubble. Not fake, but priced so high it leaves no room for any future mistakes.

Why is this the layer where bubbles form? It becomes clear when you return to the pattern of that map: optical modules are the link in the entire hardware chain with the lowest physical barrier to entry. Building a fab requires tens of billions of dollars and five years; expanding optical module production lines only requires a few billion dollars and a few quarters—it's the only hardware segment where supply can "cooperate" with speculation. When the supply side isn't locked, bubbles find a crevice to grow.

TSMC's lock doesn't protect optical modules—because optical module capacity is precisely the one link in the entire chain that doesn't need TSMC's nod.

Repeated single-day double-digit declines show smart money is already queuing at the exit.

L3 Infrastructure Layer: GPU Cloud Subletters—Alive, But Relying on Others' Bottlenecks

In the past two years, a batch of new cloud providers specializing in GPU leasing have emerged: buying cards themselves, building their own facilities, then renting compute to companies short on cards, called NeoCloud in the industry—we prefer to call them "GPU subletters."

They're doing well, and indeed have some skills: these folks squeeze hardware like F1 drivers race cars, achieving actual GPU utilization rates 2-3 times that of traditional second-tier suppliers. The same batch of cards, they can squeeze out more revenue.

The survival logic also holds: The four major cloud providers' own capacity is far from enough; the overflow demand has to go somewhere. As long as the big premise of "compute shortage" exists, subletters have business.

But note the essence of this business: They are beneficiaries of the bottleneck, not holders of the moat.

Make their situation clear: Every dollar they earn essentially comes from the time lag of "big players not expanding fast enough." But—the electricity bottleneck is expected to ease by 2027-2028; big players' own data centers are being built at the fastest pace in human history; the earlier planted seed, space data centers, if realized in the 2030s, would pull the rug out from under the scarcity logic of ground compute.

Time gaps do close. Subletters don't hold the property deed; they only hold a lease with an unknown expiration date.

Moreover, this business has another structural weakness: high concentration of customers and lifelines. Their cards come from the same chip giant, major customers are often just a couple of AI companies, and for some players, the largest shareholder and the largest supplier are the same name. Upstream holds your supply, downstream holds your revenue, you in the middle earn money by "brokering the time gap"—this kind of business can be very profitable but doesn't deserve a "platform" valuation.

Making money from other people's bottlenecks means preparing for the day the bottleneck disappears.

This layer isn't a scam; today's cash flow is real. But the high valuation the market currently gives them prices in a permanent state of a temporary condition—this is a valuation error, heading towards bubble territory.

L4 Application Layer Long Tail + VC Ecosystem: Where Bubble Signals Are Strongest

Finally, climb to the top of the pyramid. This layer needs to be split in half to view.

The head half—the few large model companies with real revenue—we've already discussed; revenue catches up with valuation, won't elaborate.

The real problem lies in the long tail, and the VC ecosystem funding that long tail. The most glaring numbers are here:

In Q1 this year, AI companies took the overwhelming majority of global venture capital—out of every $10 of VC money, over $8 flowed into AI.

In 1999, at the peak of the dot-com bubble, what was this proportion? About one-third to 40%.

That is to say, today's VC concentration of bets on a single theme is twice that of the peak of the biggest bubble in human history.

And the structure is extremely top-heavy: Just four top deals swallowed 65% of the global quarterly VC total. One quarter of the world's venture capital, two-thirds went into four companies' accounts.

This creates a transmission chain: Headline star companies support sky-high valuations with real revenue—that's fine; but thousands of long-tail startups with no revenue are borrowing the valuation logic of the leaders to price themselves—"That company grew 80x in 18 months, why can't I?"—that's the big problem. The 1999 game of "add .com and it rises" has a new version today: "add an AI Agent and it doubles."

What's more troublesome is that the death of these long-tail companies is already predictable. They won't die from product failure—the product might even be good. They will die from valuation inversion: The money raised in the last round at bubble prices burns out, the next round of investors is only willing to invest at realistic prices, and raising at realistic prices means the previous round investors suffer huge losses, the founding team's equity is wiped out—so negotiations break down, the company stalls between the "dignity of valuation" and "survival," until the cash on hand runs out. Most of those 1999 companies died exactly this way: not killed by the market, but choked by their own previous round's valuation.

There's another amplifier: The cost structure of these long-tail companies this round is more fragile than in 1999. Internet startups back then burned marketing expenses; they could cut ads and survive. Today's AI startups burn compute bills—if the model isn't called, the product stops, this expense cannot be cut. Revenue is a story, costs are rigid—this combination dies faster when capital recedes than the last round.

Note, this doesn't contradict "Large Cap has no bubble"—

Heads have real revenue backing, tails only have stories backing. Bubbles are never in the biggest companies; bubbles are in the small companies using the biggest companies' valuation logic to price themselves.

Remember what the real lesson of 1999 was? Not that "the internet was fake"—the internet was real, e-commerce was real, the biggest e-commerce company survived and dominated the world. The lesson was:

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

The Bears Aren't All Wrong Either: Two Lines of Attack Worth Thinking About Before Bed

If you think we're blindly bullish after reading this far, please read on. The bear camp has real substance, and this time, the real substance is sharper than most bulls are willing to admit.

Bears have two main lines of attack. On the surface, they are two topics; dig deeper, you'll find they are actually two sides of the same problem.

Attack Line One: The Depreciation War—How Many Years Can Your GPU Really Last?

First, explain "depreciation" with a relatable example.

Suppose you drive for a ride-sharing service, buying a car for $300,000. If you account for this car being scrapped in 3 years, the annual cost is $100,000; if accounted for over 6 years, the annual cost is only $50,000. Note: You haven't earned a penny more, the car is still the same car, simply by changing an accounting assumption, your paper profit increases by $50,000 every year.

Now replace the car with a GPU, replace $300,000 with hundreds of billions of dollars.

Tech giants are collectively doing the same thing: lengthening the depreciation period for GPUs. Originally generally accounted for over 3-4 years, now extending to 5, 6 years. Every year extended, the current period's profit looks better by a chunk. Bears estimate, with these changes, the entire industry might under-depreciate by over $100 billion in the next three years; some giants' current profits might be overstated by over 20% because of this.

What does 20% mean? It means one-fifth of the profit you see in financial reports might just be a "gift of accounting assumption," not earned by the business itself.

The bulls' rebuttal also makes sense: Depreciation periods aren't changed arbitrarily. In inference scenarios, older GPUs are perfectly capable—training frontier models needs the latest cards, but using three-year-old cards for daily inference still runs at full load, still makes money. By this logic, GPUs lasting 10, 15 years isn't an exaggeration; depreciating them over 3 years in the past was actually underestimating.

Who's right? The honest answer is: It depends on Nvidia itself. The more dramatic the performance leap of the next two generations of products, the faster older cards depreciate, the more correct the bears; the gentler the leap, the longer the lifespan of old cards, the more correct the bulls. Every time Nvidia releases a new generation, it's voting on its customers' balance sheets.

This is the most ironic scene in AI's financial issues: The more successful Nvidia's products are, the more questionable its customers' financial reports become.

Attack Line Two: GPU Credit—Moving Debt Where It Can't Be Seen

The second attack line is newer, and more hidden. Not many discuss it in the market, but we think it's an order of magnitude more serious than the depreciation issue.

GPUs have already begun circulating through complex off-balance-sheet structures. Unpacking it, this structure works like this:

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

• The shell borrows money: The SPV borrows money from private credit funds to buy tens of thousands of GPUs.

• Lease to card users: The SPV leases the GPUs long-term to AI companies, collects rent, uses rent to repay the loan.

• The card seller joins: The most brilliant step—the chipmaker itself also invests money into the SPV, becoming an anchor investor.

Everyone gets what they want: AI companies get to use cards but don't take on debt; giants and AI companies don't see this debt on their balance sheets; chipmakers lock in sales volume and conveniently earn investment returns; private credit funds get high-interest assets.

A win-win-win-win. Just one small problem: The debt hasn't disappeared; it's just invisible to everyone.

This structure should remind you of something. Actually, it rhymes with two pieces of history simultaneously.

The first is 2000. Few remember that in the telecom bubble, there was a role called "vendor financing": equipment giants themselves lent money to customers, letting customers buy their own equipment. On paper, sales were booming, growth curves perfect, but in reality, it was left hand to right hand—customers used your money to buy your goods. When the bubble burst, these equipment makers held not profits, but a pile of unrecoverable debt, dying more miserably than anyone. The structure today of "chipmakers investing in SPVs, SPVs using that money to buy chips" is a blood brother to the vendor financing of yesteryear.

The second is 2008. The last time the entire financial system was enamored with "packaging, layering, moving risks to places regulators and investors couldn't see clearly" was the mortgage securitization before that crisis. What was packaged back then was houses; what's being packaged now is GPUs.

When an industry starts giving its own customers money to buy its own products, every growth number you see needs a question mark.

Depreciation is an accounting problem; accounting problems never prick bubbles; leverage is a financial problem; every bubble in history was pricked by a financial problem.

The Two Lines Are Actually One Line

Now connect the two attack lines, and you see the true killing power of the bear logic.

The essence of the depreciation debate is: How many years can a GPU last, what is its residual value?

What is the collateral for GPU credit? Still the residual value of GPUs.

That is: The basis for an SPV borrowing billions is the assumption that "these GPUs will remain valuable for many years to come, will continue to generate rent." If Nvidia's next-generation product performance leaps another step, rents for old cards plummet—the first to blow up won't be the giants (they can withstand it), but these SPVs and the private credit funds that lent them money.

Then the question you have to ask becomes: How much has private credit ballooned in recent years? What else is stuffed in there? That's another article.

Currently, the scale of this structure is still small, far from enough to cause systemic trouble—that's the truth. But even the staunchest bulls themselves list "large-scale leveraged GPU collateral financing" as the number one risk signal of this cycle. When both bulls and bears, rarely in agreement, point to the same place saying "look there," that place deserves your serious attention.

The moment GPUs are stuffed into off-balance-sheet SPVs, 2026 smells, for the first time, a hint of 2008. It's just a hint for now—watch the speed at which it thickens.

Conclusion: Expensive, But the Door is Still Locked

Compress the entire article into one diagram, still that pyramid:

No Bubble (L0 + L4 Head): TSMC, Nvidia, the four major cloud providers, leading large model companies. Real contracts, real revenue, full-load utilization, plus the two physical locks of TSMC and power grids. Expensive, but expensive ≠ bubble.

Long-Short Battleground (L1): Memory. 70% profit margins either mark the start of a new structural cycle or the climax of the old script; the gambling table is set.

Bubble-Scented (L2, L3, L4 Long Tail): Optical modules—the only hardware link not protected by TSMC's capacity discipline, pricing 2028 revenue into 2026; GPU subletters—treating temporary bottlenecks as permanent moats; VC ecosystem—single-theme concentration reaching twice the 1999 peak, long-tail startups borrowing leaders' valuation logic to price stories.

Three Potential Minefields That Truly Need Watching:

• Algorithm Efficiency Revolution. If one day, smarter algorithms achieve the same effect with one-tenth the compute, the entire capital expenditure logic of "throwing compute at it" collapses overnight. This is the lowest probability but most lethal one.

• GPU Credit Leverage. Once off-balance-sheet structures, collateral financing, securitization spread, cash flow buyers become leverage buyers, the 2000 script replays with a 2008 engine. This is currently the most real seedling.

• TSMC Abandons Conservatism. Whether its monopoly is picked by competitors, or it changes its own mind to expand like crazy—the moment supply goes out of control, the necessary condition for a bubble truly materializes. This is the one requiring the longest-term tracking.

Before any of these three things happen, AI is a technological revolution forcibly restrained by physical laws: expensive, crowded, locally feverish, but the foundation is solid.

Finally, turn this map into three questions you can carry with you. Next time you see any AI asset, be it a stock or a startup project, first ask:

First question: Which layer of the pyramid is it on? The closer to physics, the safer; the closer to story, the more dangerous. Those who can't articulate which layer they're on, default to the most dangerous layer.

Second question: Is its revenue actually happening, or is it "borrowed" from the valuation of leading companies? The frequency of the phrase "comparable to XX company" is directly proportional to bubble concentration.

Third question: Does it earn money from structure, or from bottlenecks? Money from structure can be earned for many years; money from bottlenecks has an expiration date—and the expiration date is usually much shorter than the time implied by the valuation.

Only when you can answer all three questions, then discuss price.

Bubbles never notify you which layer they'll burst in. But you can at least choose not to stand on the layer that prices itself using other people's stories.

Next time someone asks you "Is AI a bubble?", you can ask back: Which layer are you talking about?

Those seventy-something-old engineers at TSMC might be the only people on this planet who can stop an AI bubble. So far, they're still on duty.