From Municipal Debt to Compute Debt: Is Silicon Valley Reliving History in 2026?
Hello everyone, this is the Financial Veteran from FinSages.
Right now, I’m sitting in my study in China, reviewing the events of the past January. Across my three monitors, a dense web of K-line charts and the minutes from two major overseas conferences are weaving together a picture that sends a chill down my spine.
These two signals, one of fire and one of ice, are like giant mirrors reflecting the most absurd and dangerous fissures of our time.
Let’s turn the clock back slightly, to Las Vegas in early January.
That was CES 2026. Through the livestream, I saw that familiar figure in a leather jacket—Jensen Huang. If you looked closely, you’d notice this year’s jacket seemed to be a deep purple with sharper, more defined lines. Perhaps it was a metaphor: he is the emperor of this era.
On that stage, bleached white by countless flashbulbs, Huang held up a chip as thick as a brick—a brand-new computing unit based on the “Rubin Ultra” architecture. His voice, as rousing as ever, boomed out his famous slogan to the thousands of near-fanatical developers and investors: “Buy more, Save more!”
The data flashing on the screen was dizzying: the Rubin Ultra’s FP4 compute performance was ten times that of the previous Blackwell generation, with memory bandwidth breaking a staggering 8TB/s. The applause was thunderous. Nvidia’s stock drew another steep upward line that day, as if gravity itself had been suspended.
Yet, just two weeks later, the scene changed dramatically.
Our attention shifts to Davos, Switzerland. At the snow-covered World Economic Forum, Elon Musk was not physically present; he joined via a holographic projection. In that room filled with global political and business leaders, Musk’s virtual image appeared weary, even tinged with a hint of mockery.
He addressed the elites, who were still discussing ESG and globalization: “Are you still worried about chips? That was last year’s problem. We are hitting a physical wall at light speed. What we lack is no longer silicon, but transformers, substations, and electricity. At the current pace of compute expansion, by 2027, the power demand from the AI industry alone will surpass the total residential electricity consumption of the United States.”
These two scenes from January, when viewed together, form a brilliant piece of black comedy.
On one side, the digital carnival in Las Vegas, where people believed compute could grow exponentially without limit. On the other, the physical world’s warning from Davos, where the hard constraints of energy are tightening a suffocating noose around this very carnival.
As a veteran who has navigated the financial system and witnessed three decades of tempests, this scene of “half fire, half seawater” gives me a powerful, almost terrifying sense of déjà vu.
This feeling is all too familiar. It reminds me of China in the early 2010s.
Back then, I traveled through every county in central and western China. Every local official who received me would point out the car window at the dust-filled sky and the freshly leveled farmland, telling me with unwavering confidence: “Veteran, look, this will be the future CBD, and that will be a hundred-billion-dollar high-tech industrial park.”
In those days, tower cranes were articles of faith; steel and concrete were the future. To realize these grand blueprints, local government financing platforms (known as chengtou) took on massive debt, and bank credit flooded into infrastructure like a deluge. Everyone believed in a single logic: land prices will always rise, tomorrow will be better, and today’s debt will be diluted by future growth.
If you transpose this logic onto Silicon Valley in 2026, you’ll find that the script is identical, down to the punctuation.
Today, Silicon Valley’s GPUs are yesterday’s steel and concrete. Its data centers are yesterday’s industrial parks. And the tech giants, leveraging debt and burning through cash flow to hoard compute power, look exactly like those local financing vehicles that once believed in the gospel of “land finance.”
Mark Twain once said, “History doesn’t repeat itself, but it often rhymes.”
Today, I want to strip away the glossy veneer of technology and, from the simple perspective of a financial balance sheet, take you on a tour of a brewing superstorm called “Compute Debt.” Let’s see if the tech elites across the ocean, revered as gods, are simply re-enacting a history we have already walked—a path full of temptations and traps.
Chapter 1: The Silicon Age’s “Land Finance”
To understand the crisis of 2026, we must first shatter an illusion: in the AI era, compute power (GPUs) is no longer a simple industrial consumable or a production tool. It has been thoroughly financialized.
On the balance sheets of Silicon Valley’s giants, the GPU is not listed as “equipment”; it is the “land” of the digital world. To own it is to own the right to collect rent in the future AI world.
I. The 2026 “Compute Land Rush”: An Uncontrolled Arms Race
Let’s begin with some deep data reconnaissance.
In recent days, the Nasdaq has been turbulent as Microsoft, Alphabet (Google), and Meta released their Q4 2025 earnings reports. Wall Street analysts are still cheering for revenue growth, but I urge you, like me, to fix your gaze on one line in the cash flow statement: “Capital Expenditures” (Capex).
The figures are staggering.
The combined quarterly Capex of these three giants has surpassed $55 billion. What does this number signify? It means that for every $100 they earn from advertising and cloud services, they must immediately reinvest over $40 back into the ground.
And where does this money go? The vast majority flows into a single pocket: Jensen Huang’s Nvidia.
At CES, besides Huang’s Rubin, I also paid close attention to the moves of AMD’s Lisa Su. “Su-Ma,” as she is known, remains as sharp as ever, unveiling the MI450 chip to compete with Nvidia. Her strategy is to play the “cost-performance” and “inference-side optimization” cards, attempting to punch a hole in Nvidia’s monopolistic fortress.
But what unsettles me is not the intensity of the competition, but its “homogenization.” It’s just like the real estate market of the past; whether it was Vanke, Evergrande, or Country Garden, everyone was ultimately doing the same thing: grabbing land.
Today, Silicon Valley is trapped in a “prisoner’s dilemma”-style arms race.
Microsoft’s Satya Nadella cannot stop, because OpenAI’s model parameters continue to expand exponentially. Mark Zuckerberg cannot stop, because he believes Meta must possess the world’s most powerful open-source compute cluster. Google’s Sundar Pichai dares not stop even for a moment, because the moat around his search business is being filled, bit by bit, by AI.
They are caught in a vicious cycle: to sustain their stock price, they must prove their leadership in AI; to prove leadership, they must own the most H100s, H200s, or Rubins; to buy these chips, they must burn through future cash flow.
Is this not the same high-turnover model of “acquire land – build – sell – acquire more land” from years past?
II. The “Vacancy Rate” Anomaly: Defensive Hoarding and FOMO
In the real estate market, the most terrifying metric isn’t high land prices; it’s the “vacancy rate.”
And in 2026, the AI industry is witnessing a bizarre phenomenon that is rarely mentioned: the “compute vacancy rate.”
According to a friend working at a top cloud provider in Silicon Valley who shared internal data with me (data that, of course, will never appear in public financial statements), at least 30% of the top-tier GPUs in major data centers are currently in a “low-load” or even “idle” state.
This sounds utterly preposterous. The world is clamoring about a chip shortage, so why are the purchased chips sitting idle?
My friend used an analogy that made me understand instantly. He said, “Veteran, it’s just like your ‘ghost cities’ of the past. The buildings are up, but no one lives in them. Why keep building? Because the land prices are rising.”
This is “defensive hoarding.”
In 2026, companies are not buying chips because their current operations genuinely require this much compute power. Today’s AI applications, aside from coding assistance, search, and simple text-to-image generation, have not yet produced a true “killer app” to consume such vast amounts of compute.
They are buying chips purely out of FOMO (Fear of Missing Out).
Every CEO is worried: What if AGI (Artificial General Intelligence) is suddenly achieved tomorrow? What if a competitor develops a model ten times more powerful than GPT-6? If I don’t have enough “compute reserves” on hand, I won’t even have a seat at the table.
Thus, GPUs have become strategic nuclear weapons. Even if they sit collecting dust in a warehouse, even if they consume power while idling, they must be owned. This logic has directly led to an “artificial boom” in compute demand. The market demand we see is not driven by real user needs, but by the insecurity of giants.
This has created a massive misallocation of resources. On one hand, small and medium-sized developers are begging for a single chip. On the other, in the data centers of giants, millions of the most advanced GPUs are humming away, consuming electricity, only to run trivial test data.
III. The Transfer of Debt and the “Curse of Moore’s Law”
If “vacancy” is merely a waste of resources, then the core crisis of “compute debt” lies in a fatal flaw of the asset itself.
This is the point I, as a risk management veteran, most want to emphasize: hoarding GPUs as if they were land is the greatest financial miscalculation of this century.
The reason municipal debt in the past could be sustained for so long, even rolled over by issuing new debt to pay off old, is that land possesses a core attribute: spatial exclusivity and value preservation. A plot of land in downtown Shanghai remains downtown Shanghai ten years later; it may even be more valuable due to surrounding development. Land is an ally of time.
But what is a GPU? It is an electronic product. And electronics are the enemy of time.
Moore’s Law is the sword of Damocles hanging over all “compute landlords.”
The Rubin Ultra that Jensen Huang unveiled at CES, with its ten-fold performance increase over the two-year-old Blackwell, is a gospel for technicians but a death knell for chief financial officers.
It means that the H100 cluster you spent a fortune on, perhaps even issued bonds for, just two years ago, has instantly become an “electronic relic” in 2026—with one-tenth the compute power and twice the energy consumption of the new product.
This isn’t “hoarding land”; this is hoarding a pile of ice cubes melting under the sun.
Let’s do the math.
A tech giant issues $10 billion in bonds to build an H100 data center. At the pace of Moore’s Law, the “economic life” of this data center is only three to four years. After four years, its competitiveness will drop to zero.
This means the company must, within three to four years, earn back the $10 billion principal plus interest by running AI services on this data center. It needs to generate an annual return of at least 30% for this model to work.
But what is the reality?
The reality is that in the current AI application layer, aside from Microsoft’s Copilot and a few top SaaS products, the revenue growth rate for most AI products is only between 15% and 20%.
Revenue Growth (15%) < Hardware Depreciation Rate (30%).
This is a fatal scissors difference. In finance, it’s known as a precursor to a balance sheet recession.
The giants are transferring hard cash into Nvidia’s pockets in exchange for a pile of rapidly depreciating silicon chips. And to support these purchases, they are continuously issuing more bonds.
It’s like a person taking out a high-interest loan to buy a truckload of fresh seafood, fantasizing that the price of seafood will keep rising. The reality is, if the seafood isn’t sold in three days, it rots.
This is not just a risk for individual companies; it is a systemic risk. The bonds of these tech giants are held by pension funds and insurance companies worldwide. Once the “compute bubble” bursts, once Wall Street realizes these massive data centers are non-performing assets unable to generate cash flow, the ensuing collapse will be no less severe than the 2008 subprime mortgage crisis.
Moreover, as Musk stated, the physical world is beginning to say “no” to the greed of the digital world.
When electricity becomes the new “credit line” and transformers the new “land permits,” how long can this game of hot potato with compute power last?
Between the clamor of Las Vegas and the austerity of Davos, I feel as if I am watching a great pendulum swing from the extreme of irrational exuberance, accelerating towards the abyss of a cyclical reckoning.
Chapter 2: The Scythe of Depreciation and the Curse of Moore’s Law
Having described the phenomenon, we must now delve into the underlying ledger to dissect the brutal mathematical logic behind this “compute carnival.” As a veteran who has dealt with risk for thirty years, I have always adhered to one principle: all business models must ultimately answer to the rate of return on assets. If an asset’s depreciation outpaces its ability to generate cash flow, then no matter how magnificent its technological cloak, it is, in financial essence, a black hole that devours wealth.
Today, in 2026, the seemingly invincible giants of Silicon Valley are collectively facing an unprecedented challenge. I call it “the backstab of Moore’s Law.”
To understand compute debt, we must first clearly see its fundamental genetic divergence from the land debt of the past. This is a fatal point that many investors, and even some top Silicon Valley venture capitalists, have chosen to ignore.
Think back to our real estate logic over the past three decades. Why could land carry such astronomical debt? Because the underlying collateral was land and property.
What is land? In financial terms, land is an ally of time.
When I was in credit approval at the bank, if a company used a plot of land in the city center as collateral, I would dare to approve a twenty-year loan. I knew that this land possessed “spatial exclusivity.” As long as the city continued to develop, the value of that land would most likely spiral upwards. Even if the business failed, leaving the land to bake under the sun for five or ten years might actually increase its value. Land can hedge against inflation and self-heal the balance sheet through rent or asset revaluation.
Therefore, while land debt is heavy, it has “roots.”
However, the giants of Silicon Valley in 2026 are attempting to play this game with an entirely different kind of asset. They are no longer hoarding land, but GPUs (Graphics Processing Units).
What is a GPU? Physically, it is a sophisticated semiconductor. Financially, it is the most brutal of “consumables.” It is the enemy of time.
This brings us to the curse that hangs over all “compute landlords”—Moore’s Law.
The “Rubin Ultra” architecture that Jensen Huang unveiled at this year’s CES is undoubtedly an engineering marvel. Its single-chip FP4 compute power is a full ten times that of the Blackwell architecture from two years ago, its memory bandwidth has quadrupled, and its energy efficiency has improved by 40%.
Technicians cheered, but in my eyes, this was nothing short of a financial slaughter of existing assets.
Let’s conduct a deep, even brutal, financial projection.
Imagine you are an ambitious AI cloud service provider in 2024. To keep up, you raise $1 billion through debt and purchase over 30,000 of the then state-of-the-art H100 GPUs at $30,000 apiece, building a massive intelligent computing center.
In your presentation to investors, you described this as a long-term, infrastructure-like investment, promising to slowly recoup costs and reap handsome profits over the next decade through compute rental services.
However, the calendar has now turned to February 2026.
The moment Rubin Ultra was announced, your billion-dollar H100 cluster was instantly relegated to second-class status in the compute market. Customers are extremely pragmatic. When they discover that training models on a Rubin chip is ten times faster than on an H100, and the electricity cost per unit of compute is lower, who would be willing to rent your slow, power-hungry old machines?
Unless, of course, you offer insane discounts.
According to data I’ve obtained from second-hand hardware dealers in Silicon Valley, the buyback price for a used H100 has plummeted from its 2024 peak of $35,000 to less than $4,000 today.
What does this mean? It means that in just two years, the value of your assets has evaporated by nearly 90%.
On a financial statement, this is called an “impairment loss.” In reality, it is the bloody destruction of wealth.
Land debt can be extended; you can use time to create space by issuing new debt to pay off old. But compute debt is a race against time. You must earn back the principal and interest within three to four years—before the chips are completely obsolete.
This is a race that is destined to be lost.
Why? This brings us to the second fatal mismatch: the “scissors difference” between revenue and costs.
Let’s look at the income statement.
The current AI industry exhibits a bizarre “inverted” structure: the one selling shovels (Nvidia) takes 90% of the industry’s profits, while those digging for gold (the application layer) are losing money just to stay in the game.
While SaaS revenue is growing—companies like OpenAI, Anthropic, or Microsoft’s Copilot have indeed increased their subscriber counts in 2026—if we strip away the related-party transactions designed to inflate revenue, the real growth rate from external customers is generally between 15% and 25%.
And what is the rate of hardware cost depreciation? Due to the acceleration of Moore’s Law, standard accounting practices are already shortening the depreciation period for AI servers from five years to three. This means the cost pressure from depreciation alone amounts to a staggering 33% of the asset’s total value each year.
Revenue grows by 20%, while assets depreciate by 33%.
This is elementary school math, but it’s a life-or-death equation that gives CFOs sleepless nights.
Worse still, AI business has a characteristic that distinguishes it from traditional software: increasing marginal costs.
In the traditional internet era, if I developed an app, the additional cost of serving 100 million users versus 10,000 was almost negligible. But in the AI era, every user query, every line of code generated, consumes tangible compute power, which means consuming electricity and GPU lifespan.
This means that as your user base expands, your costs rise linearly, or even exponentially. It is incredibly difficult to reduce marginal costs to zero through economies of scale.
So, we see a vicious cycle: the higher the company’s revenue, the deeper its losses. It’s like a giant trapped in quicksand; the more it struggles, the faster it sinks.
To maintain their stock prices and their image as “AI leaders,” the giants can only continue to increase their investment. The old chips haven’t paid for themselves, but the new ones (Rubin) are already out. To avoid being eliminated by the market, they have to grit their teeth, issue more debt, and buy the new chips.
This has the classic characteristics of a Ponzi scheme: using the next round of capital expenditure to cover the losses of the previous one. As long as the growth story is still being told, as long as new financing can be secured, the game can go on.
But all financial games eventually hit a physical wall.
If Moore’s Law is the “soft blade” from within, then energy is the “hard constraint” from without.
Musk’s warning at Davos was no exaggeration. In 2026, another flashpoint for compute debt is not in the chips, but on the electricity meter.
During my research into the data center corridor in Northern Virginia, I witnessed a shocking scene. It is the “crossroads of the global internet,” home to the world’s densest concentration of data centers.
But I saw several newly built, massive intelligent computing centers with their gates locked, patrolled only by security guards. They were filled with expensive Blackwell chips, but the machines were silent, their indicator lights dark.
I asked my local guide, “Why aren’t they turned on?”
The guide pointed to a distant power pylon. “No power. Dominion Energy announced that the substation expansion has been delayed until 2028 due to environmental approvals and equipment shortages. Not a single new kilowatt can be supplied before then.”
This is what’s known as “stranded assets.”
In finance, this is the most tragic state. You’ve borrowed money, bought equipment, and the equipment is there, but a crucial external condition is missing, preventing the asset from generating any cash flow.
For land, if there’s no water or power, the land is still there; you can just wait a few years. But for a GPU, this kind of “waiting” is fatal.
The machines are off, but the depreciation clock has not stopped. With every passing second, that pile of chips gets one step closer to becoming e-waste.
Furthermore, electricity is not only scarce; it is expensive.
In 2026, due to AI’s terrifying consumption of power, global industrial electricity prices have risen by over 20% on average. This has directly shattered the cost baselines of many AI startups.
I’ve done the calculations: in a standard H100 cluster data center, the total lifecycle electricity cost is already approaching half of the hardware procurement cost. If electricity prices rise another 30%, then 90% of the AI model training tasks on the market today would be unprofitable on an economic basis.
It’s no wonder, then, that Microsoft and Google are frantically buying nuclear power plants and even investing in controlled nuclear fusion. This is no longer about environmentalism; it’s about survival.
Electricity has become the “credit line” of 2026.
In a traditional debt crisis, a central bank can provide liquidity by printing money and lowering interest rates to alleviate debt pressure. But in a “compute debt” crisis, the Federal Reserve can print dollars, but it cannot print kilowatt-hours (kWh).
The hard constraints of the physical world offer the most ruthless mockery of the digital world’s infinite expansion.
When we piece all this together, a breathtaking panoramic view emerges:
- On the asset side: rapid depreciation driven by Moore’s Law, with asset values melting like icebergs.
- On the liability side: rigid interest payments and massive capital expenditures, like an ever-tightening noose.
- On the revenue side: meager growth that cannot keep up with depreciation, like a funnel that can never be filled.
- On the environmental side: a physical shutdown caused by power depletion, like an oxygen tube suddenly being cut.
This is the true picture of Silicon Valley in 2026.
On the surface, it is a revolution of “new quality productive forces,” a grand journey towards AGI. But in the eyes of a financial veteran, it is clearly a superstorm brewing on the balance sheet.
It is so much like the fiber-optic bubble of 2000. Back then, companies like Global Crossing also took on massive debt to lay undersea fiber, convinced that internet traffic would double every three months. The result? The fiber was laid, the traffic didn’t come, and the companies went bankrupt.
Of course, after the fiber bubble burst, it left behind extremely cheap infrastructure that laid the groundwork for the subsequent prosperity of the internet.
Today’s “compute debt” may share the same fate. When the bubble bursts, when Nvidia’s stock price reverts to the mean, when those expensive GPUs are sold off at scrap metal prices, the true AI era may finally arrive.
But before that, in this most frenzied and dazzling year of the bubble, 2026, countless fortunes are destined to be turned to ash under the scythe of depreciation.
The Diamond Sutra says: “All conditioned phenomena are like a dream, an illusion, a bubble, a shadow, like dew or a flash of lightning; thus should you view them.”
This verse from over two thousand years ago, used to describe today’s compute assets—as ephemeral as dew, as fleeting as lightning—is astonishingly precise and poignant.
Faced with this situation, do the CEOs at the center of the storm, the helmsmen of trillion-dollar capitals, truly not know?
Of course, they know.
But why can’t they stop? Why do they keep their foot on the accelerator, knowing there is a cliff ahead?
To understand this, we must move from cold financial figures to the depths of human nature. In the next chapter, we will push open the door of the “prism of human nature” to see what kind of “prisoner’s dilemma” is being played out in that decision-making room accessible to only a few.
Chapter 3: The Prisoner’s Dilemma Through the Prism of Human Nature
By this point in the story, I believe many of you, especially those with a background in finance or industry, have a burning question in your minds.
You will ask: “Veteran, if the math is so clear, if the scythe of Moore’s Law is so sharp, if the energy ceiling is pressing down, and even Wall Street analysts are privately muttering about the rate of return, then why?”
Why do Satya Nadella, Mark Zuckerberg, and Sundar Pichai—the smartest minds in the world, the helmsmen of trillion-dollar empires—continue to act like frenzied gamblers, pouring hundreds of billions of hard cash into this clearly overheated furnace?
Can they not see the steep depreciation curve? Do they not know about the resistance from the physical world?
The answer may send a chill down your spine. Of course, they know. They know better than anyone. On their desks, the risk assessment reports from their CFOs are likely thicker than the GPUs themselves.
But they cannot stop.
This is no longer just a matter of economics; it has evolved into a deep game of survival, power, and human nature. To understand this layer, we must put down our calculators, pick up a scalpel, and cut through the glamorous suits of these Silicon Valley giants to examine their anxiously beating hearts.
This is the third layer of logic I want to discuss today: the prisoner’s dilemma through the prism of human nature.
In the field of financial risk management, we have a classic theory known as agency risk. Simply put, the interests of the person playing the cards (the CEO) are not always perfectly aligned with the interests of the person funding the game (the shareholders).
Imagine you are the CEO of a major tech company. It is 2026, and the whole world is proclaiming that AGI, or Artificial General Intelligence, is just around the corner. This is a civilization-level paradigm shift, comparable to the steam engine, electricity, or the internet.
At this moment, you have two buttons in front of you.
The red button is: “Call the bet.” This means you have to burn ten billion dollars this quarter to buy GPUs, which will make your financial report look ugly and turn your company’s free cash flow negative. If you lose the bet—that is, AGI doesn’t arrive as quickly, or the GPUs you bought become obsolete—you’ve wasted a few tens of billions of dollars at most. For a trillion-dollar company, this is painful but not fatal. You can explain it as necessary strategic trial and error, a ticket to the future.
The blue button is: “Wait and see.” This means, for the sake of financial prudence, you decide to pause the arms race and wait for the technological path to become clearer. Your financial report will be beautiful, your profit margins high. However, hidden within this option is a massive, devastating risk. What if your rival, the company next door, presses the red button and they win the bet? They achieve AGI first.
What is the outcome then? The outcome is that you will be completely eliminated. The search empire, the social empire, the office software empire you’ve built over the past twenty years could be dimensionally crushed and crumble overnight.
For a professional manager, wasting a few tens of billions of shareholder money is an “operational loss,” which might earn you a reprimand. But missing an entire era due to conservatism, leading to the company’s demise, makes you a “historic criminal,” the end of your career.
Faced with these two options, any rational CEO would unhesitatingly choose the red button.
This is why we see them buying, buying, and buying, even when GPUs are idle, even when electricity prices are soaring, even when applications have not yet landed. They are not buying compute power; they are buying “insurance against obsolescence.”
What does this remind me of? It is so much like the local government debt expansion in China back in the day.
At that time, I researched many counties and cities. Did those local officials not know the risks of building a massive new district in the middle of nowhere? Did they not know that the roads and buildings constructed with debt might not turn a profit for a decade?
Of course, they knew. But this was not just about economics; it was about politics.
Under the performance evaluation system of that era, GDP growth was a hard target, and attracting investment was a military order. If your neighboring county took on debt to build infrastructure, making their roads wide and their industrial parks beautiful, and attracted all the major enterprises, while you remained idle to save money, then a few years later, your neighbor would rise, you would be marginalized, and your career would be over.
So, it was a tournament from which no one could withdraw. In Silicon Valley in 2026, the stock price is political achievement, and the parameter scale of AI models is GDP.
Driven by this mechanism, all participants are trapped in a classic prisoner’s dilemma. Everyone knows that a collective pause would be best for the industry, avoiding vicious competition and wasted resources. But no one dares to be the first to stop, because whoever stops, dies.
Thus, we witness an extremely distorted allocation of resources.
On one hand, the contracting forces of the physical world are sounding alarms: power grids are overloaded, chips are depreciating, and applications are scarce. On the other hand, the expansionary forces of the capital world are relentlessly upping the ante. Because tech giants with the highest credit ratings can raise money from the market at extremely low costs, they use this cheap capital to fill the gaps created by the physical world’s limitations, attempting to bend the rules of physics with money.
This forced distortion has given rise to the most bizarre phenomenon of 2026: the bubble of fake demand.
If in the first two chapters we discussed oversupply, now I must pull back the curtain on the demand side.
Have you ever wondered what this astronomical amount of compute power is actually running?
In the official promotional videos, it’s running protein folding to cure cancer, complex models to solve climate change, and simulation training for autonomous driving.
But in the real data centers, in the backend logs I’ve seen, a significant portion of the compute power is trapped in a vicious cycle of “entropy increase.”
This is the strange loop of “AI training AI.”
Because high-quality human data—the books, articles, and code actually written by humans—has already been scraped in previous years, models now, in pursuit of greater capabilities, have begun to use “synthetic data.” That is, they use content generated by the previous generation of AI to train the next generation.
It’s like an Ouroboros, a snake eating its own tail.
A new type of “digital sweatshop” has emerged in Silicon Valley. Tens of thousands of H100s run day and night, generating vast amounts of text, images, and videos. This content is not consumed by humans, nor does it generate real commercial value. Instead, it is packaged into new datasets and fed to the next version of the model.
In thermodynamics, this is called entropy increase. In economics, it is internal friction. In finance, it is idling.
We are investing extremely expensive energy and hardware to produce a pile of information “waste.” This growth is illusory and unsustainable. It is like some places in the past that, in order to boost GDP figures, dug up roads and repaved them, only to dig them up again. The GDP figures went up, but the city’s real wealth did not increase.
However, this false prosperity is precisely what the capital markets crave.
Wall Street no longer cares about your active user numbers or your retention rates. They only care about one metric: the Scaling Law.
As long as you can tell them that as compute power increases, the model’s intelligence continues to grow—even if this growth cannot yet be monetized, even if it is achieved by “eating its own tail”—they are willing to continue inflating your stock price.
This is an extremely dangerous collusion.
To prove to Wall Street that they have a future, tech giants must maintain high capital expenditures, creating the illusion of a compute shortage. To maintain the bull market bubble, Wall Street must ignore basic valuation logic and sing praises for this money-burning game. And Nvidia, as the biggest beneficiary, continuously launches more powerful and more expensive shovels to maintain the intensity of this gold rush.
In this closed loop, everyone is winning. Who is losing?
Only the underlying physical world is losing, because energy is being wasted. Only the ultimate payer is losing, because the dividends of this technological cycle are being spent in advance.
I often say that people in risk management tend to have a “crow’s eye.” We are always staring at the shadows behind the prosperity.
In this early spring of 2026, looking at these still-frenzied K-line charts, my mind always flashes back to the scene just before the dot-com bubble burst in 2000.
Back then, Cisco’s market capitalization once surpassed Microsoft’s, making it the world’s number one company. Everyone believed that internet traffic would double every three months, so the demand for routers and switches was infinite. Cisco was the “shovel seller” of that era.
The result? The fiber was laid, the routers were installed, but the applications didn’t follow. People discovered they didn’t need that much bandwidth to send emails. Demand plummeted, Cisco’s stock price fell by 90%, and countless fiber-optic companies went bankrupt.
Is today’s AI re-enacting this scene?
When all the giants have stored up compute power to the extreme, when all the electricity has been drained, if at that time, the legendary AGI still has not appeared, or even if it has, it turns out that it cannot solve the most complex, non-standard problems for humanity—like fixing pipes, caring for the elderly, or handling complex interpersonal disputes—
Then, what will become of these trillions of dollars in compute assets?
They will become massive “negative assets.”
On the financial statements, they will face huge impairment charges, directly piercing the income statement. In the real world, they will become a pile of expensive scrap metal, unable to even earn back their electricity costs.
And the most ironic thing is that this kind of “surplus” has often been the soil for the next round of innovation in history.
The bursting of the fiber-optic bubble left behind extremely cheap bandwidth resources, which led to the later rise of YouTube, Netflix, and the golden decade of the mobile internet.
If the compute bubble of 2026 bursts, perhaps we will usher in an era of “cabbage-priced compute.” When an H100 drops from thirty thousand dollars to three hundred, perhaps the era of true AI popularization will arrive. At that time, AI will no longer be the nuclear weapon in the hands of giants, but the screwdriver in the hands of ordinary developers.
But that is all in the future.
Right now, we are at the most brilliant and most fragile moment of the bubble.
This is not just a game of capital; it is a trial of human nature. In this giant casino, every CEO is a hostage, driven by fear, tempted by greed, and pushed forward by a sense of “having no other choice.”
They cannot stop. Because in this game, to stop is to die.
But as observers, as investors, as witnesses of this era, we must remain clear-headed. We cannot assume that just because they are running, there must be a road ahead.
Sometimes, the cliff is right at the end of the flower-strewn path.
In the next chapter, we will leap out of these suffocating games and, from a higher dimension, contemplate the endgame of this cycle. When the bubble dissipates and compute returns to common sense, how can we, the ordinary people, find our own ark amidst the ruins?
After all, history teaches us: every bursting of a bubble is a violent redistribution of wealth. Some are left swimming naked when the tide goes out, while others have been quietly building ships.
Chapter 4: The Endgame of the Cycle and the Path to Redemption
History is a scriptwriter with a poor imagination. It always takes the same script, recasts it with new actors, and performs it again in a different era.
As we stand in the early spring of 2026, watching this relentless arms race in compute power, we can already faintly see the final page of the script.
Written upon it is not the end of the world, but two words: The Reckoning.
As a veteran accustomed to cycles, I dare to make a prediction here—one that may not be popular, but is likely to be proven true: in the second half of 2026, or by the first half of 2027 at the latest, we will witness Silicon Valley’s “Lehman Moment.”
It will be a superstorm known as the “compute fire sale.”
The trigger may not be a specific technological failure, but a financial report. Perhaps it will be Microsoft, or Meta. After several consecutive quarters of massive investment without realizing the expected revenue growth, the CFO will finally be unable to persuade the vultures of Wall Street with the rhetoric of “strategic losses.”
In that moment, confidence will collapse like a line of dominoes.
The H100 and Rubin chips, coveted just yesterday, will become hot potatoes overnight. Because everyone will suddenly realize that the costs of electricity, maintenance, and depreciation to keep these behemoths running far exceed the value they can generate.
And so, the sell-off will begin.
We will witness a surreal scene: the second-hand market will be flooded with top-tier GPUs in pristine condition, their prices plummeting from tens of thousands of dollars to just a few thousand. Nvidia’s sky-high stock price will undergo the most brutal “reversion to the mean.”
It’s analogous to the housing market in some of our third- and fourth-tier cities. When the reality of population outflow could no longer be concealed, those river-view apartments once sold for 20,000 RMB per square meter might not find a buyer even at 5,000.
It sounds cruel, doesn’t it?
But please note, as a long-term thinker, what I call a “reckoning” is not the death of the industry. On the contrary, in my eyes, the bursting of a bubble is often the beginning of the industry’s true maturation.
This is where the final force in our “Four Forces Model” comes into play: Evolutionary Force.
In this darkest hour, Silicon Valley might do well to learn a bit of wisdom about “debt resolution” from across the Pacific.
Over the past two decades, China experienced the largest infrastructure boom in human history, accumulating massive debt in the process. When the period of high-speed growth ended, what did we do? We did not dynamite the bridges and highways. We began a strategic transformation: from a “build-out” model of massive construction to an “operational” model of revitalizing existing assets.
The same logic will be the only way out for the AI industry.
Before 2026, the main theme of AI was “Training.” This was like building infrastructure; everyone competed on the size of their model’s parameters, on who could dig the deepest foundation. It was a crude, capital-intensive approach.
But after the bubble bursts, the main theme of AI will be forcibly switched to “Inference.” This is like running operations; the competition will be about who can run more cars and transport more goods on the highway that has already been built.
We will see a profound paradigm shift:
First, from “large and comprehensive” to “small and beautiful.”
Companies will no longer be obsessed with an all-powerful GPT-6. Since compute is expensive and power is scarce, they will discover that a small, 7-billion-parameter model, fine-tuned for a specific vertical and running on cheap, local chips, can perfectly solve 90% of concrete problems. Think of the edge-side agents like Clawbot we discussed earlier; they don’t need massive data centers, they live on your phone and laptop.
Second, from “showing off” to “cost reduction.”
During the bubble, the focus was on whether AI could write poetry or paint pictures. In the reckoning phase, people will only care about one thing: Can you reduce the cost of processing financial statements by 50%? Can you lower the complaint rate in our call center by 30%? The standard of value assessment will shift from “shock and awe” to “ROI” (Return on Investment).
Third, the “public-utilitization” of compute infrastructure.
When GPU prices crash, compute will no longer be the private property of tech giants. Like fiber optics of the past, it will become an extremely cheap, basic resource.
This is the most fascinating aspect of the “Evolutionary Force.”
Let’s recall the dot-com bubble of 2000. Back then, companies like Global Crossing took on massive debt to lay enough undersea fiber to circle the globe dozens of times. Then the bubble burst, these companies went bankrupt, and investors lost everything.
But the fiber did not disappear. These expensive assets lying at the bottom of the ocean were auctioned off at rock-bottom prices to later operators. It was precisely because of this almost-free bandwidth that we later had YouTube’s video streaming, Netflix’s media empire, and the golden decade of the mobile internet.
One generation plants the trees; the next enjoys the shade. One generation goes bankrupt; the next picks up the pieces.
The compute bubble of 2026 will ultimately leave behind a rich legacy.
When those millions of H100s are released onto the market at low prices, when the cost of compute becomes dirt-cheap, the true era of “AI popularization” will arrive.
At that time, AI will no longer be the nuclear weapon in the hands of Nadella and Zuckerberg. It will become the screwdriver in the hands of ordinary entrepreneurs, the toy in a college dorm room, and an affordable utility for every small and medium-sized enterprise.
We will no longer debate whether AI will replace humans. We will start discussing how to use AI to design an automatic cat feeder for five cents.
This is the endgame of the cycle. This is the redemption of technology.
Conclusion: Finding Peace and Waiting for the Bloom
The night is deep. I close my laptop and walk to the window. Looking at the scattered lights of the city outside, I reflect on my thirty-year career in finance. I have seen buildings rise and fall, witnessed the madness and demise of the crypto world, and watched the ascent and twilight of internet upstarts.
The roaring compute frenzy of today is but another cycle of human greed and dreams intertwined. It is no nobler than the tulip mania of the past, no more frenetic than the railway fever.
As ordinary people, as specks of dust in this grand era, how should we conduct ourselves?
I have three pieces of advice for you:
First, do not be the one to catch the “last baton” at the peak of the bubble.
No matter how miraculously the media portrays AI, no matter how enviably high Nvidia’s stock price soars, remember this simple common sense: no asset can defy gravity and rise forever. When the grandmothers at the wet market start discussing GPU funds, it’s your time to leave.
Second, pay attention to those who are building ships on the ruins.
When the bubble bursts and the industry is in turmoil, do not panic. Instead, open your eyes wide and look for the teams that are still persistently solving real problems, still polishing their products. Those who are not swimming naked when the tide goes out are the kings of the next cycle.
Third, build your inner “antifragile” embankment.
This era is changing too fast. Yesterday it was the Metaverse, today it is AI, tomorrow it may be quantum computing. If you drift with the current, your anxiety will be endless. The only things that can give you a firm footing are your understanding of underlying principles, your ability to think independently, and a mind that is not swayed by the noise.
I am reminded of Su Dongpo’s famous poem, “Calming the Storm.” After experiencing the great tempests of his life, he wrote that immortal line:
“Looking back at the windswept place I came from, I return, finding neither storm nor shine, only calm.”
On the cusp of this uncertain era of 2026, I hope that you and I can both possess this steadfastness, finding calm amidst storm and shine.
Don’t be dazzled by the neon lights of Las Vegas, nor be terrified by the snowstorms of Davos.
The bubble will eventually dissipate, and compute will return to common sense. What we must do is maintain our own rhythm, quietly sharpening our swords in the trough of the cycle.
Because when the clamor fades, the real future has only just begun.
I am the Financial Veteran from FinSages. In this age of information overload and scarce truth, I am willing to be the fellow traveler who watches the winds and clouds for you, and deciphers the puzzles.
If you hope to find a measure of cognitive certainty in the turmoil to come, if you wish to find your own coordinates at the crossroads of technology and finance, you are welcome to follow our community: FinSages.org.
There are no myths there, only common sense; no anxiety, only wisdom.
Until next time.
