Vertex Macro | Financial Trader Cloud · The Fourth Wall of the Market
Who Sold Shovels in the AI Bubble, and Who Is Using Shovels to Dig Their Own Grave
The Fourth Wall of the Market
The AI industry chain looks like the future, but once you take it apart, it looks more like a construction site divided among layers of subcontractors. Investors sit in the client's meeting room, cloud vendors hold the general-contracting agreement, Nvidia sells materials nearby, large-model companies bring construction crews onto the site, and application companies paint the walls in SaaS colors. Enterprise clients are eventually invited to the inspection, where a quotation sits on the table beneath an attractive heading: efficiency revolution, intelligent transformation, cost optimization. Retail investors stand outside the hoarding, staring at the sea in the artist's rendering and imagining they have bought a sea-view apartment facing the future.
Then the wind blows, and everyone realizes the sea is still inside the PowerPoint deck.
The most absurd part of this AI boom is that it does not depend on outright fabrication. The models really have improved, chips really have been in short supply, cloud vendors really have been spending heavily, and enterprises really have been running pilots. Real things give the bubble a convincing exterior, so prices gain the nerve to climb into the sky. The entire chain is doing the same thing: pushing costs onto the next layer while selling stories to the layer above. Everyone says they are selling shovels because it sounds safe, intelligent, and close to cash flow. But if an entire street is made up of shovel shops, there had better be a real mine at the end of it.
In 2024, Nvidia founder and CEO Jensen Huang said at Dell Technologies World that humanity was beginning to “manufacture intelligence,” and that data centers would convert data and electricity into tokens. He called data centers AI factories. The phrase was clever, like a construction permit that repackaged server racks, electricity, cooling, land, and depreciation as a factory that could print money.
The market loves this kind of renaming. Data centers move from cost centers to factories, chips from hardware to scarce capacity, and electricity bills from operating expenses to fuel for intelligence production. Change the language and valuations get new clothes. Once financial markets are prepared to pay a high price, what they often need most is not evidence, but a presentable story.
I. The Gold Rush Begins; Shovel Shops Open First
Every speculative mania has a phrase that sounds seasoned. In this AI cycle, the market loves this one: in a gold rush, the people selling shovels make the most money.
The saying was not originally wrong. Very few people actually find gold; tool sellers get paid first. Nvidia was the first to taste that sweetness. In fiscal 2025, Nvidia's 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,” and that reasoning AI had introduced another scaling law.
Translated into trading-desk language, it was simple: the shovels were still in short supply.
The danger grew from there. Chip companies said they sold shovels, and the market nodded. Cloud companies said they sold shovels too, and the market began to get excited. Data centers, power companies, enterprise software vendors, consulting firms, and outsourcers crowded onto the shelves one after another, with valuation models packed together like stalls in a night market. Everyone wanted to stand upstream; no one wanted to admit they were merely selling lunch boxes beside the construction site.
IMF Managing Director Kristalina Georgieva wrote in January 2024 that AI would affect nearly 40% of jobs globally, with about 60% affected in advanced economies. She cautioned that AI could raise productivity, but could also replace jobs and worsen inequality. The statement gave the market a macroeconomic pass. As long as a company's name touched AI, it was as if the productivity revolution had rained on it, leaving its valuation with a damp, glossy sheen.
But productivity is not perfume. You cannot spray a little on every company and turn the whole street into Paris.
That was the source of the market's vague unease. Everyone dimly knew the development was important, and everyone knew prices were moving too quickly, but it was hard to say exactly where the bubble was growing. It did not necessarily grow in a single stock. It grew in a collective impulse: everyone wanted to describe themselves as a shovel seller.
When everyone in an industry chain stands upstream, the bubble is probably standing upstream too.
II. Nvidia Collected Entrance Fees First; Everyone Else Queued to Enter
Nvidia's position on this construction site was clean. It did not stand on the street corner shouting slogans; it sold cement, steel bars, and cranes directly. Whether the site could ultimately be completed was debatable. Whether the materials had already been paid for was not.
Jensen Huang had long presented the story as an industrial revolution. During GTC 2024, he said in an interview that enterprises had previously viewed data centers as costs and capital expenditure, while factories made money; data and electricity entered AI factories and tokens came out. The language was extremely effective financial cosmetics. Depreciation sounded uncomfortable; capacity sounded comfortable. Electricity bills sounded vulgar; tokens sounded sophisticated. Server racks occupied space; factories created the future.
Nvidia's earnings therefore became a policy meeting for AI trading. At the September 2024 FOMC press conference, Fed Chair Jerome Powell said the FOMC had decided to cut the policy rate by 50 basis points, citing greater confidence that inflation was declining and the labor market was coming into balance. That meeting repriced the cost of dollar funding. Nvidia's Blackwell shipments, supply-chain progress, and gross margins repriced the supply of AI imagination.
That is why professionals watched Nvidia the way they watched nonfarm payrolls. Retail investors watched the stock price; fund managers watched data-center revenue; strategists watched the capex cycle; traders watched to see whether the story could spill over into power, optical modules, servers, cloud services, and software.
There was no inspirational quotation here. Upstream profits did not guarantee downstream comfort. Shovel sellers turned miners' anxiety into revenue, which was good business and the cleanest way to collect money in the early bubble. The trouble came later: the shovels were being sold before the gold mine had even been dug, while valuations already assumed the whole village had struck gold.
The market can be an impatient client. The foundation has not even been poured, and it is already calculating the skyscraper's future rent.
III. Cloud Vendors Bought Shovels, Then Rented Them to Others
Cloud vendors occupied the most awkward position on the construction site.
They were like general contractors: first buying materials from Nvidia, then renting computing power to model companies, enterprise clients, and application developers. From the outside, people saw them holding vast numbers of GPUs and assumed they stood upstream in the gold rush. Finance departments opened the books and saw land, electricity, servers, networks, depreciation, and payback periods growing longer and longer.
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 Microsoft had moved from talking about AI to helping customers turn AI into real outcomes. The company's July 2024 earnings materials also quoted Nadella saying Microsoft wanted to meet customers' immediate critical needs while ensuring that it led the AI era.
The words sounded measured, but trading desks heard another sentence: capital expenditure still could not stop.
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 about $75 billion in 2024, with 2025 even higher. He invoked AWS's early capital-investment cycle to reassure the market, saying the company could turn investment into operating profit, free cash flow, and return on invested capital.
That passage sounded like a collective prayer from the cloud vendors. Burn the money first, build the plant first, carry the depreciation first; customer usage would catch up later. AWS's history had shown that this path could work, so the market was willing to buy another ticket. This time, however, the ticket was more expensive and the audience was more crowded.
When discussing AI infrastructure investment in 2026, Jassy laid out the cash cycle even more plainly. Data-center land, power, buildings, hardware, chips, and network equipment all required upfront investment—some six months ahead, some two years ahead. Data-center assets could last more than thirty years, while hardware and networks lasted about six. Once revenue caught up with capex, profit margins and cash flow would begin to look good.
That was the cloud vendors' real position. They bought shovels, carried the depreciation, and sold future usage rights. AI demand certainly existed, but could price, time, and utilization all hold together? If even one of the three failed, the general contractor's smile would begin to thin.
Cloud vendors looked like the people controlling the construction site, but they were also like subcontractors being pressed for progress by the client. The client was called the capital market, the blueprint was called the AI era, and the payment schedule was called “we'll talk about it later.”
IV. Shovels Started Costing More Than Gold
In the early stage of a gold rush, somewhat expensive shovels are understandable. Everyone fears missing the mine; toolmakers raise prices, and people still buy.
The trouble begins when shovels become so expensive that it takes several gold mountains to break even. AI already has that flavor. Chips are expensive, data centers are expensive, electricity is expensive, model training is expensive, and inference is not cheap either. Enterprise adoption also requires system integration, data cleaning, process changes, compliance clearance, and permission setup. The final bill will not be sent to the future; it will be sent to the current quarter's cash flow.
Alphabet CEO Sundar Pichai said in the February 2025 earnings release that the company expected to invest about $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.
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 capex guidance from $37 billion to $38 billion, with actual full-year capital expenditure reaching $39.23 billion.
These numbers were not visions. They were footprints left on cash-flow statements.
The market initially viewed AI capex as an offensive move, then slowly began to detect the flavor of a forced position. Management did not dare stop; stopping looked like an admission that they had no seat on the next platform. The giants stared at one another's data centers like large holders comparing positions in a stock without particularly deep liquidity. The first to withdraw could easily be branded by the market as having lost faith.
The toxic point was simple: shovels were already valued like gold mines, while the gold was still stuck in customer pilots and product road maps.
If enterprise clients ultimately became willing to pay real money for large volumes of AI services, revenue could absorb the investment. If enterprises merely wanted to try the products at a discount for a few months, then cut the price dramatically at renewal—so dramatically that even SaaS procurement managers felt embarrassed—everyone except the shovel sellers would have to recalculate their destiny.
By then, many companies would discover that they were not actually selling shovels. They were renting someone else's shovel and putting their own trademark on it.
V. Model Companies Used Shovels to Dig Dreams
Model companies were like construction crews. They stood in the middle of the site, worked the hardest, and were the easiest to use in a story.
They had indeed built astonishing things. Models could write programs, summarize, draw, translate, chat with people, and sometimes perform like an always-on junior employee on particular tasks. But a construction crew's problems are never only whether it can work. There are also material costs, schedules, rework rates, and final payments.
At the July 2024 ECB Central Banking Forum, Fed Chair Jerome Powell said, while discussing generative AI, that the Fed 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.
That slow central-bank tone was useful for traders. It reminded the market that demos were not productivity, downloads were not profit, and enterprise pilots were not long-term budgets.
Model companies most feared users trying the product and saying, “Pretty good, but not worth that price.” The sentence hurt, but it came close to the underlying rule of the SaaS world. Users found the product useful; CFOs asked how many people it had saved. CEOs called it a strategic investment; boards asked when it would reach the income statement. Employees liked having another chat box; finance departments asked whether next year's renewal could be cut in half.
Model companies burned GPUs, data, engineers, and market patience. Many free users did not automatically become a good business. High API call volumes still had to be tested against unit economics. Before inference costs declined, much of the growth looked lively but felt thin-margin, and when calculated out, resembled a subsidy.
Construction crews most feared a client who demanded more spectacular renderings every day, but began talking about cash-flow discipline when it was time to sign payment authorizations.
At that point, dreams would be forced to put on hard hats and accept cost accounting.
VI. Application Companies Packaged Shovels as SaaS
AI application companies were like renovation crews. They wrapped model capabilities in attractive interfaces, applied them to industry scenarios, and gave them names that sounded as if they belonged to the future. Demos were usually beautiful because demos did not need to handle a company's legacy systems, dirty data, permission problems, or legal clauses.
Once the products actually entered the building, the world was not so clean.
Where was the enterprise data? Who could see it? Could the model remember it? Who was responsible for an incorrect output? Would employees bypass the process? Would a supplier lock in its pricing? These questions were not suitable for a product launch, but they would sit in the procurement conference room.
In its 2024 annual report, Microsoft Chairman and CEO Satya Nadella emphasized that the company was turning AI into real outcomes for customers, and mentioned enterprise customers using Microsoft 365 Copilot to improve creativity and productivity. This kind of statement mattered to enterprise software because the long-term retention of Copilot premium would determine how quickly AI could turn from a feature into revenue.
The cruelest part of enterprise software was that customers kept the books.
A feature that merely made presentations prettier would soon have its price pushed down. A tool that merely saved employees half an hour would prompt finance to ask whether that half hour had become revenue or meant hiring one fewer person. Many AI applications looked like companies, but when unpacked were APIs with skins; unpacked again, they were prompt wrappers; and at the end, they were business plans that needed retention rates to prove themselves.
Real moats were usually not in the demo. They were in data, workflows, distribution, compliance, industry know-how, and customer trust. These things grew slowly, were not as attractive as a model launch, and did not fit neatly into a three-minute video. But they determined whether a renovation crew could survive the first wave of hype.
AI application companies most feared platforms suddenly building their features in-house. Yesterday, they were a startup story; today, they were a menu option in Office, Google Workspace, or a cloud platform. The market first gave them TAM, then peer comparisons, and finally might leave them only a plugin valuation.
Plugins could make money too. They simply could not easily pretend to be infrastructure.
VII. Enterprise Clients Were the Owners Finally Asked to Pay
The entire AI capital chain eventually had to arrive in one room. Sitting in that room was the enterprise client's CFO.
This CFO was not very romantic. He did not care much whether the model was approaching AGI, nor did he care much about a particular benchmark score. He had budget tables, procurement processes, internal-control requirements, and pressure from the board. The questions he asked were mundane and fatal: How many people were saved? How much revenue was added? How much did error rates fall? Why were renewal prices so high?
At that moment, magic would be repriced.
IMF Managing Director Kristalina Georgieva said in 2024 that AI would affect high-skilled jobs, and could allow people who knew how to use it to increase their productivity and income while leaving those who could not use it behind. In enterprise procurement, that statement translated into a colder sentence: enterprises would pay for tools that genuinely changed workflows, while remaining politely skeptical of products that merely added another interface layer.
When discussing AI in 2024, Fed Chair Powell said central banks were also studying its effects on productivity, inflation, and growth, and that it was still too early to reach a conclusion. The sentence belonged beside every AI commercialization model. Reaching a conclusion too early was the mistake valuations loved most.
Enterprise clients were not ultimately buying a feeling about the future. They were buying the possibility of putting results into the income statement. If AI could halve customer-service headcount, significantly speed software development, raise sales-conversion rates, and reduce fraud losses, prices had support. If it merely gave employees another chat box, made reports sound more polished, or made meeting minutes look more like meeting minutes, renewal season would be ugly.
The sentence the market least wanted to hear was right there: the greatest fear for AI commercialization was that users would try it, think it was pretty good, and then ask for 70% off.
That reaction was more damaging than not using the product. Non-use could be blamed on insufficient market education. Using it and then haggling meant the finance department had already examined the magic.
VIII. From Gold Rush to Liquidating the Construction Site
Bubbles do not necessarily end in crashes. Sometimes they are simply taken over, slowly, by accounting departments.
At the gold rush's most beautiful moment, everyone talks about gold mines. When the wind dies down a little, electricity bills, lease contracts, depreciation curves, utilization rates, retention rates, gross margins, and returns on capital begin appearing on the table. These words have no oracular quality, but they come closest to the market's final judgment.
When discussing financial stability at the October 2024 Bloomberg Regulatory Forum, Bank of England Governor Andrew Bailey cited Hyman Minsky and Charles Kindleberger. He reminded markets that, after financial crises fade, people often believe a “new era” has arrived, and Cassandra-like warnings are ignored.
By the end of 2025, the Bank of England's financial stability assessment had begun discussing AI-related valuation risks directly. Reports cited the Bank of England as noting that enthusiasm for AI investment had pushed some stock-market valuations close to dot-com-bubble levels. Bailey also reminded investors that even if AI ultimately succeeded, not every company currently valued highly would become a winner.
The statement was cold and professional: successful technology, failed stocks. People who had actually spent time in markets knew that these two things often happened at the same time.
Once the construction site began to be liquidated, the industry chain would stratify. Upstream companies such as Nvidia might still earn profits, but their valuations would be repeatedly questioned by gross margins, competition, and the pace of demand. Cloud vendors would need to prove that capex could become stable cloud revenue. Model companies would need to prove that their APIs were not merely coupons for burning money. Application companies would need to prove customer stickiness rather than just a pretty shell. Enterprise software companies would need to prove that premium seats could be retained. Data centers and power companies would need to prove that demand came from real usage rather than a collective illusion in the capital market.
The market would not ultimately ask how much you resembled the future. It would ask only whether you could collect the money.
Once that question was asked, many stories would immediately begin to look old.
Conclusion: Shovels Are Real; Graves Might Be Real Too
This article does not need to offer simple answers.
AI is real technology, Nvidia has made money selling shovels, cloud vendors have strategic reasons for increasing capex, model companies are worth investing in, and application companies may create new workflows. The problem lies elsewhere: too many participants in a single industry chain are presenting themselves as shovel sellers.
When everyone stands upstream, the bubble stands upstream too.
The AI industry chain is like a construction site divided among layers of subcontractors. The client wants the future, the general contractor wants scale, the material seller wants gross margin, the construction crew wants valuation, the renovation crew wants renewals, and the owner wants ROI. Everyone has a reason, so the construction site grows larger and larger. Collective madness is certainly frightening; collective rationality can be more troublesome. It piles the bills onto everyone until each person feels blameless—until no one is willing to admit first that the building may not sell.
Mature traders do not need to shout buy or sell orders. They want to know where the money comes from, which layer it flows to, who recognizes revenue first, who carries the depreciation, who has pricing power, and who will eventually have to explain to the CFO why this future factory has not started producing cash.
The gold rush will not disappear in such a dramatic fashion. It usually becomes quiet first. The renderings remain on the hoarding, while people inside begin checking invoices. Someone in the distance is still shouting about gold mines; nearby, someone is already calculating the electricity bill.
The shovels are real.
The gold may not be enough for everyone to share.
People speculating on shovels may finally discover that they did not buy mining rights, but a half-finished presale building beside the construction site.
> The Fourth Wall of the Market | The market thinks it is watching the central bank, but in reality, the central bank is also always watching the market. This place writes about macro-finance, central bank narratives, asset bubbles, and trader psychology.