Jensen Huang's Compute Temple: Who Is Burning Incense to GPUs in the AI Bubble?

The market thinks it is investing in artificial intelligence, but it is actually paying incense money to a compute temple. The temple was inconspicuous at first. It was just a server room, a row of servers, several stacks of hot chips, and a few engineers watching curves on their screens. Then models began speaking like humans, code began completing itself, and images started growing out of prompts. Suddenly, the capital market seemed to hear a bell ringing in the night.

After the bell rang, components were no longer just components.

GPUs were originally computing tools; later, they became relics of the technology market. Data centers were originally cost centers; later, they were renamed AI factories. In 2024, NVIDIA founder and CEO Jensen Huang said at Dell Technologies World that the previous industrial revolution manufactured software, earlier ones manufactured electricity, and humanity was now beginning to manufacture intelligence. He also said that data centers would convert data and electricity into tokens.

There was something theological about this statement. It gave the server room a monk's robe and found an afterlife for depreciating assets.

The market quickly learned this language. Server rooms became temples, electricity became incense, chips became ritual implements, and capital expenditure became merit. Investors talked about productivity, but what they were really buying was a system of devotion. The closer a company was to the GPU, the closer it was to the altar.

I. GPU Turns from Component into Sacred Object

The most fascinating thing about the AI bubble is that, at first, it looks very unlike a bubble.

It has real products, real users, and real revenue, and it has made many tasks faster. After ChatGPT appeared, white-collar workers collectively felt for the first time that they had a silent colleague beside them. It did not drink coffee, take sick leave, or complain about procedures in meetings. It occasionally talked nonsense, but at three in the morning it could still deliver a passable first draft.

The shock was real. And precisely because it was real, the story that followed was easy to push out of control.

IMF Managing Director Kristalina Georgieva wrote in January 2024 that AI would affect nearly 40% of jobs worldwide, with roughly 60% affected in advanced economies. She cautioned that AI could raise productivity, but could also replace jobs and worsen inequality. Her tone lacked the excitement of a technology launch. It sounded more like a medical report from a macroeconomic institution: a new role had indeed appeared beside the production function.

The market had no patience to wait for the examination results to unfold.

Investors did not want to calculate, company by company, exactly how much labor AI had saved. Nor did they want to wait for CFOs to record efficiency gains in the income statement. The market needed something that could be quoted every day, report results every quarter, and give fund managers a clear talking point at the morning meeting. So the GPU was lifted onto the altar.

NVIDIA delivered a set of fiscal 2025 results that read almost like a scripture. Full-year revenue reached $130.5 billion, up 114% year over year; fourth-quarter data-center revenue reached $35.6 billion, up 93% year over year; and full-year GAAP gross margin reached 75%. In its February 2025 earnings release, Jensen Huang said Blackwell demand was "amazing," that reasoning AI had added another scaling law, that training made models smarter, and that long thinking made answers smarter.

Those numbers briefly quieted the market. People stopped urgently asking how much each AI application was actually worth. Cloud vendors were still buying chips, so faith came with a receipt. Data centers were still being built, so the temple still had an expansion program. NVIDIA's gross margin was still standing there, allowing investors to imagine future cash flows in respectable clothing.

This is the most ruthless feature of the GPU: it turns a vague future into current-period revenue. Others talk about visions; it ships products. Others talk about changing the world ten years from now; it recognizes revenue this quarter.

Believers can dream, but the temple should at least have a point-of-sale machine.

II. NVIDIA Is the Temple's Cash Register

Every technology bubble has someone who can turn imagination into financial statements.

In the railway bubble, tracks gave the future a physical shape. In the internet bubble, fiber-optic cables and routers brought "traffic" down to earth. In the AI bubble, NVIDIA does the job even more directly. It turns the almost theological word intelligence into equipment that can be delivered, priced, scheduled, and installed in a data center.

Jensen Huang understands this very well. He repeatedly uses the word factory. During GTC 2024, he said in an interview that data centers had previously been viewed as costs and capital expenditure, whereas factories make money. He placed data, electricity, and tokens inside one production process, effectively adding an illusion of return to capital expenditure.

The phrase is like a building permit for the entire temple.

Once a data center is called a factory, investors find it easier to tolerate its appetite. Servers consume chips, chips consume electricity, electricity consumes the grid, and the grid consumes land and time. These things originally sounded like costs. After passing through the ritual of the AI factory, they can be renamed future capacity. Language is not innocent here; it is performing a ceremony for accounting categories.

This is also where NVIDIA looks most like a central bank.

Federal Reserve statements change expectations for dollar liquidity. NVIDIA earnings change expectations for AI liquidity. At the September 2024 FOMC press conference, Fed Chair Jerome Powell said the Federal Open Market Committee had decided to cut the policy rate by 50 basis points because it had greater confidence that inflation was declining and the labor market was coming into balance. That meeting repriced the path of funding costs. NVIDIA's Blackwell shipments, supply-chain arrangements, and guidance repriced the market's confidence in the future supply of imagination.

That is why NVIDIA earnings look like the results of a technology company but trade like the nonfarm payrolls of the AI world. Good numbers make the market feel that the god is still speaking. Tight supply makes the incense more expensive. High margins give the temple pricing power. A successful Blackwell ramp gives the next pilgrimage a reason to continue.

A truly powerful idol does not need to perform a miracle every day. It only needs to keep the pilgrims in line, and the offering box will fill by itself.

III. Cloud Vendors Begin Interrogating One Another

A temple needs pilgrims; the AI bubble needs hyperscalers.

Microsoft, Google, Amazon, and Meta all appear to be investing in the future. In practice, they watch one another every day. A company that buys fewer GPUs looks like a country that bought one fewer steam engine on the eve of the Industrial Revolution. A company that slows data-center construction must explain to the market whether it has begun to feel afraid.

Microsoft Chairman and CEO Satya Nadella wrote in his 2024 annual letter to shareholders that fiscal 2024 was the second year of the AI platform shift, and that the company had moved from talking about AI to helping customers turn AI into real outcomes. In Microsoft's July 2024 earnings materials, the company also quoted Nadella as saying that it wanted to meet customers' immediate critical needs while ensuring that it led the AI era.

That language is polished. The version heard on the trading desk is much shorter: the money still has to be spent.

Meta founder and CEO Mark Zuckerberg said on the third-quarter 2024 earnings call that the company was building AI infrastructure faster than it had imagined at the beginning of the year, and that he was pleased with the speed of execution. Meta subsequently raised the low end of its 2024 capital-expenditure guidance from $37 billion to $38 billion, with actual full-year capital expenditure reaching $39.23 billion.

Amazon CEO Andy Jassy said on the third-quarter 2024 earnings call that generative AI was a once-in-a-lifetime opportunity. Amazon expected capital expenditure of approximately $75 billion in 2024, with 2025 even higher. He brought out AWS's early capital-investment cycle to reassure the market. The meaning was clear: bury the money in the ground first and talk about free cash flow a few years later.

Alphabet CEO Sundar Pichai said in the February 2025 earnings release that the company expected to invest approximately $75 billion in capital expenditure in 2025 to accelerate AI progress. Alphabet CFO Anat Ashkenazi added that most of the spending would go toward technical infrastructure, mainly servers, data centers, and networks.

Taken together, these statements give shape to the market's hard-to-describe unease.

Cloud vendors are buying more than equipment. They are also buying proof that they cannot afford to be absent. AI has become the attendance sheet of the capital market. ROI can arrive late, but capex cannot be missing. Management can say that it will maintain investment discipline, but it cannot allow a competitor to build a data-center wall first.

The most exhausting part of this arms race is that nobody wants to be the first to put down the incense. The first company to stop will be suspected by the market of having lost divine favor.

IV. Large-Model Companies Recite Sutras, but Temple Rent Is Expensive

The busiest people in the temple often have no pricing power.

Large-model companies are responsible for making believers think the god can really speak. They demonstrate reasoning, multimodal, agent, and programming capabilities. At a product launch, a model resembles a priest: calm voice, elegant answers, and occasionally even a poem. Once it leaves the launch stage and enters an enterprise workflow, it begins to look less sacred.

It needs supervision, data cleaning, compliance review, system integration, hallucination management, and lower inference costs. It remains useful, but increasingly resembles a highly paid intern.

At the July 2024 ECB Central Banking Forum, Fed Chair Jerome Powell said that the Federal Reserve was spending considerable time studying AI's effects on productivity, inflation, and the labor market. He also said it was still too early to determine whether AI would eliminate jobs, augment jobs, or create new ones.

The central bank's tone is usually slow, and that slowness has value. The market likes to discount every demo immediately. The central bank at least reminds us of one thing: productivity does not automatically flow out of a keynote.

This is the dilemma facing model companies. Capability improves quickly, but commercial adoption moves slowly. Enterprise procurement, information security, internal procedures, legal responsibility, and budget cycles are all checkpoints that a miracle must pass before entering reality. Employees feel that AI saves time; management asks whether it can reduce headcount. Product managers say engagement is high; finance asks how much gross margin inference costs consume. The CFO does not oppose the god. The CFO only wants to know the god's monthly ARPU.

NVIDIA can collect money upstream because everyone fears a GPU shortage. Large-model companies face a different judgment: they must prove that they have not burned GPUs merely to produce beautiful answers. A monk may recite sutras beautifully, but the temple rent still has to be paid each month. If pilgrims only want to attend free morning prayers, the offering box will eventually ring with emptiness.

V. Electricity and Data Centers Are the Temple's Foundation

Every religion eventually collides with the physical world.

No matter how intangible a token is, it requires real electricity. No matter how light the cloud appears, it rests on land, power grids, cooling systems, water resources, transformers, and construction crews. The market began by discussing model parameters, then discovered that the real bottleneck might be power connections and data-center delivery.

When discussing AI infrastructure investment in 2026, Amazon CEO Andy Jassy described AWS's cash cycle plainly. Data-center land, power, buildings, hardware, chips, and network equipment all require upfront investment: some six months in advance, some two years in advance. Data-center assets can last more than thirty years, while networks and hardware last approximately six years. Only after revenue catches up with capital expenditure do operating profit, free cash flow, and return on invested capital begin to look attractive.

That passage is an accounting footnote to the entire AI bubble.

The market sees miracles; companies first receive depreciation schedules. The market talks about productivity; companies first sign power contracts. The market talks about intelligence; companies first secure power capacity. These things have no mystique, but they determine whether the temple can open its doors.

In a March 2025 speech at the University of Leicester, Bank of England Governor Andrew Bailey said AI could become a general-purpose technology like electricity in the early twentieth century, potentially raising growth rates and per-capita income over the long term. He also emphasized the need to invest in human skills so that the economy could use AI effectively.

Bailey's electricity analogy is clever. In the previous era, electricity was the infrastructure of technological revolution. In the AI era, electricity has once again become the bottleneck of a model revolution. History has bitten its own tail, while the capital market sits beside the snake building a DCF model.

The foundation will not be built faster simply because the narrative is attractive. Data centers must be constructed, the grid must be connected, chips must be delivered, and cooling systems must operate. Once certain expenditures have been committed, they cannot be exited with one click like a stock position. AI capex may look like an option on the future, but much of it consists of hard fixed assets.

Once the temple is built, the pilgrims had better actually come.

VI. The Bubble Is Real, but the Price Is Moving Too Fast

Calling AI a fraud is easy and cheap.

AI plainly contains real substance. It has already changed software development, content production, customer service, search, data analysis, and the use of enterprise software. The painful question is how much of tomorrow today's price has already consumed.

In financial history, many bubbles grew around real things. Railways were real, but railway stocks could still die. Electricity was real, but electrification stocks could still collapse. The internet was real, but the valuations of 2000 could still send a generation of fund managers away to write self-criticisms.

Speaking about financial stability at the October 2024 Bloomberg Regulatory Forum, Bank of England Governor Andrew Bailey specifically cited the ideas of Hyman Minsky and Charles Kindleberger. He reminded the market that after financial crises recede, people often begin to believe that a new era has arrived, and Cassandra-like warnings are ignored.

When the words new era appear, the risk department should pull its chair closer. A new era may genuinely arrive, but it is often sold many times in advance. Capital markets are very good at placing technological progress into valuation models and then opening champagne today with cash flows that belong far in the future.

By the end of 2025, the Bank of England's financial stability assessment had begun discussing AI-related valuation risks directly. Reports citing the Bank of England said that enthusiasm for AI investment had pushed some stock-market valuations close to dot-com-bubble levels. Bailey also warned that even if AI ultimately succeeded, companies with high valuations today would not necessarily all become winners.

That statement is cold enough: successful technology, failed stocks. The market dislikes it because it cuts off the lazy route from "the story is right" to "the stock must be right."

The real danger of a bubble is that the market begins treating the future as money already collected. AI may genuinely change the world, but where that change will lead, whose hands the profits will reach, and whether returns on capital can cover depreciation and electricity costs are questions that will not disappear simply because Jensen Huang walks onto a stage in a leather jacket.

The god may exist. The temple may even make money. But the offering box is already valued like St. Peter's Basilica. That is the problem.

VII. The Puppeteer Behind the Scenes Is Called "Cannot Be Absent"

In this play, Jensen Huang is the priest, NVIDIA is the cash register, cloud vendors are the pilgrims, large-model companies are the monks, and the capital market is the congregation.

The real character hiding behind the scenes is called "cannot be absent."

It sits on the shoulder of every CEO. When Microsoft increases AI investment, it whispers that no one can afford to miss the platform shift. When Google expands its data centers, it warns that the search gateway may be rewritten. When Amazon raises AWS capex, it says that the next wave of cloud demand must not fall into someone else's hands. When Meta buys GPUs to train Llama, it warns that the social empire must not be left behind by the next-generation interface.

Zuckerberg said in 2024 that Meta was building AI infrastructure faster than expected at the beginning of the year. Jassy said generative AI was a once-in-a-lifetime opportunity. Pichai said Alphabet would invest approximately $75 billion to accelerate AI progress. Nadella said fiscal 2024 was the second year of the AI platform shift. Linked together, these public statements resemble a collective confession from the large technology companies.

They may not be insane. More troubling is that they are all rational.

Viewed separately, every company has reasons to invest. Taken together, the industry may be heading toward a mismatch among capacity, depreciation, pricing, and return cycles. That is where the market's vague unease comes from. It is not simply that AI seems too expensive, nor that AI seems useless. The market senses that these companies are threatening one another with capital expenditure and then packaging that threat as a vision.

Traders who shout buy or sell orders like solving the world with one sentence. The real market problem is less obedient. The technology can work while the valuation becomes excessive. Demand can exist while supply overshoots. Upstream companies can make money while downstream companies remain uncomfortable. Enterprise adoption can rise while unit economics still face judgment.

Mature traders do not need inspirational slogans.

They only want to know where the incense money flows, who recognizes revenue first, who carries the depreciation, who has pricing power, and who will ultimately be required by the CFO to prove that the miracle can be renewed.

Conclusion: The Market Is Buying the Right to Renew the Miracle

The sophistication of the AI bubble is that it does not need to be entirely false.

It only needs to collect too much money too early for things that are real. It only needs to make the market mistake technological shock for ownership of profits. It only needs to make every CEO believe that buying one fewer batch of GPUs today will put the company into tomorrow's failure case. It only needs to make investors see data centers as temples, electricity as incense, and depreciation as the cost of pilgrimage.

Jensen Huang's power lies in more than selling chips.

He gave the market a complete language. The data center became an AI factory, the GPU became an intelligence engine, and the token became a new commodity. Once the language succeeded, capital expenditure could put on ceremonial robes. NVIDIA's fiscal 2025 revenue and gross margin stamped that language with financial legitimacy.

The temple must eventually face the ledger.

Cloud vendors' capex must be depreciated. Model companies' inference costs must fall. Enterprise customers' willingness to pay must be demonstrated. AI applications must retain their users. Power and data-center investments must be supported by real demand. If they are not, the market may not immediately reject AI. It will first reject the excessively excited price that was attached to it.

AI will change the world. That may be true.

The price has already consumed part of that world in advance. That sentence belongs beside the trading desk's screen.

The idol may be real, and the incense may be real too.

The temple is not yet fully built, but the offering box has already been valued as if it were the gateway to heaven.