AI has already ceased to be a beautiful presentation about the future. Now these are factories, power plants, hundreds of billions in capital expenditures and an uncomfortable question for investors: when will all this start to pay off?
Wall Street has an amazing talent for turning a good technology into a very expensive investment story. Railroads, the internet, electric vehicles - each era has its own symbol of progress and its own line of people wanting to buy it at any price. Today, artificial intelligence is at the center of the line. And I have no doubt that AI will radically change entire industries. What I doubt is something else: that any stock with the letters AI in the investment presentation automatically deserves a cosmic multiplier.
In February Reuters reported on plans by Microsoft, Amazon, Alphabet and Meta to spend more than 600 billion dollars on artificial intelligence in 2026. This is an investment program of industrial scale. It is not only about model developers: the costs go to chip manufacturers, data center builders, energy companies, transformer suppliers, network equipment and cooling systems. The paradox is that the stock exchange is already pricing in many future results, while the physical infrastructure has yet to be built.
My main thesis: AI is a real technological revolution, but the valuations of its participants can outpace economic returns by years. The market loves the future so much that it periodically forgets to ask it for financial statements.
Four different businesses under the AI label
I propose dividing the industry into at least four links. The first is computing hardware: NVIDIA, AMD, memory and network component manufacturers, including Broadcom. Their business benefits from accelerator purchases and growing computing requirements. Here demand is already translating into revenue. But its pace depends on the budgets of a relatively small group of the largest buyers.
The second link is hyperscalers: Microsoft, Alphabet, Amazon, Meta and other owners of cloud and digital infrastructure. They are today paying for most of the construction boom. For them AI is simultaneously a source of new products and a huge expense item. When an investor sees growth in capital expenditures, it cannot automatically be counted as growth in future profits. Capital expenditures must first be recouped.
Third - energy and infrastructure: GE Vernova, Eaton, Vertiv, gas turbine manufacturers, network systems and cooling. These are not trendy chatbots, but without energy the smartest data center turns into a very expensive server warehouse. In the US, connection times to power grids and the availability of generation are becoming a separate constraint on projects.
Fourth - applied software, where most of the promised productivity should be realized. Buyers must not just test an AI assistant, but demonstrate measurable cost savings, revenue growth or efficiency gains. The problem is that demonstrating an impressive product and signing a solvent contract are two different things.
Who is collecting money today, and who is still only writing checks
In July Reuters, using consensus estimates from LSEG, highlighted the pressure of AI infrastructure spending on the free cash flow of the largest technology companies. Analysis of expectations for 2027 showed: the increase in capital expenditures could significantly exceed the additional operating cash flow. This is not proof that the investments will fail. It is a reminder that building the future requires a lot of money in the present.
The question for the investor is: how many dollars of cash flow will appear after the company spends a dollar on computing infrastructure? And when exactly? In one year, three or seven? At a high cost of capital, the difference in timing matters enormously.
For semiconductor manufacturers the task is different. If clients in 2027 begin to increase spending more slowly, the drop in order growth may matter more than absolute revenue growth. The market prices expectations about expectations. Sometimes a company earns record profit and its shares fall because investors expected an even larger record.
A telling episode is the canceled IPO of Australian Firmus Technologies in October. Reuters Breakingviews linked the withdrawal to investor doubts about the project valuation and infrastructure execution risks. Interest in AI has not gone away, but willingness to pay any price no longer looks unconditional.
Physics versus presentation
On the stock exchange you can recalculate a company's value in a minute. It does not work that way with a power system. Reuters described constraints on the American AI boom: a queue of projects waiting to connect, a shortage of gas turbines and the need to build own generation. This is more important than another speech about superintelligence.
There is also a macroeconomic side. On October 1 Federal Reserve Vice Chair Philip Jefferson named the large-scale construction of AI infrastructure one of the factors of economic activity and pressure on certain prices. AI may raise productivity in the future, but today construction competes for labor, electricity, equipment and capital.
The result is a curious structure: a technology that promises lower costs can, at the deployment stage, raise some of them. And if the Fed keeps rates higher than expected, discounting distant AI profits becomes less generous.
The hidden side of the AI race: debt burden, capacity leasing and the fight for cost
There is another question the market discusses far less often than the number of video cards: who ultimately finances the infrastructure and on whose balance sheet does the risk remain? A data center has a long operating life, but individual computing systems can become technologically obsolete much faster. A business built on expensive equipment must achieve sufficient utilization in a relatively short time to cover initial costs and future equipment replacement.
Not all obligations are visible in the headline about capital expenditures. Long-term leases of premises, power purchase agreements, capacity buyback guarantees and project financing can distribute risk between the cloud platform, developer, electricity supplier and lender. The very existence of a long-term obligation does not mean a financial problem; but it cannot be ignored when assessing economic return. That is why I look at leases including future obligations, interest payments and free cash flow after capital expenditures. The stock market may love a growth story, but a lender prefers to receive money on schedule.
A separate line of competition is the in-house accelerators of cloud companies. If hyperscalers expand the use of chips developed internally, third-party equipment suppliers may retain revenue growth but face price pressure or changes in order structure. The buyer's gain in computing does not have to be the gain of every accelerator manufacturer. For NVIDIA, AMD, Broadcom and infrastructure suppliers it matters not only the number of AI projects but also what share of investments goes to a specific architecture, how sustainable the technological advantage is and how much training and inference cost per unit of useful work.
One simple test of valuation rationality
Imagine a hypothetical company investing $100 billion in new computing capacity. If after amortization, electricity, personnel, taxes and infrastructure maintenance costs it generates additional free cash flow of only $5 billion per year, the simple payback period on the initial capital is about twenty years. This is an illustration of the mechanics, not a calculation for a specific company; discounting and the need for repeat investments make the economics even more demanding. If the flow is $20 billion, the picture is fundamentally different.
That is why when comparing issuers I would highlight four metrics: capital expenditures to operating cash flow; additional operating result from new services; amortization and assumed useful life of equipment; cost of capital raised. Two businesses with the same AI revenue can have completely different value to shareholders.
A scenario that reinforces my cautious thesis of overheating is sustained order growth accompanied by declining free cash flow and rising debt obligations. A scenario that would make me revise it is confirmed marginal profit from AI services growing faster than capital expenditures. Convincing an investor that “this time you do not need to count” is usually the most expensive way to explain that counting will still be necessary.
Bubble or healthy growth? The wrong question
Fans of extremes demand a choice: either AI is the new internet, or it is an outright scam. But both positions are too convenient to be professional. The internet turned out to be a revolution; that did not save shareholders of many internet companies from losses in the early 2000s. Successful technology does not guarantee a successful investment at any price.
I see three scenarios. Base: AI monetization improves, spending remains high, and the market gradually differentiates companies by investment efficiency. Positive: corporate productivity and demand for new services accelerate faster than capex; current valuations receive fundamental confirmation. Negative: construction and financing rates continue to rise, but returns lag; investors revise valuations and rotation out of the most expensive stories begins.
It is especially important that a slowdown scenario in AI capex is not necessarily equally bad for everyone. An accelerator manufacturer may lose order momentum while a cloud company improves free cash flow thanks to lower spending. These are opposite effects inside one industry.
What I watch instead of loud promises
First, the dynamics of capital expenditures of Microsoft, Alphabet, Amazon and Meta relative to their operating cash flow. Second, backlog and forecasts for orders of accelerators, memory and networking equipment. Third, monetization: paid corporate deployments, AI service margins, average revenue per customer. Fourth, the cost and timing of access to electricity. Fifth, Treasury yields - they determine how much the market is willing to pay for a distant future.
A sign of strengthening of the bullish scenario is growth of cash flow following invested funds and expansion of the number of companies creating measurable value. A sign of its weakening is ever higher valuations amid deteriorating returns and project delays.
Practical filter: count not the promise but the return
To compare participants in the AI chain it is useful to ask one question: what additional cash flow arises for every dollar invested - and over what time frame? For a hyperscaler it is important to look not only at cloud revenue growth but also at the profitability of AI services after electricity, amortization and infrastructure maintenance costs. For an equipment manufacturer - at the concentration of orders among the largest buyers and whether declared orders turn into deliveries and cash flow. For an applied software developer - at paid deployments, customer retention and confirmed savings for the customer, not at the number of pilot projects.
This check also has a temporal dimension: servers and accelerators become obsolete, and data centers require ongoing expenses. Therefore not only the economics of launching the project matter, but also the period during which it can generate income before the next major upgrade. For each company it is worth comparing cautious, base and optimistic scenarios and checking whether acceptable returns are preserved at lower utilization, delayed monetization or higher financing costs. If the investment thesis works only under a flawless scenario, the price may already leave no margin for error.
Why the strongest technologies often turn out to be bad investments
There is one uncomfortable question rarely asked during technological euphoria: who will capture the economic rent when the technology becomes mass market? Successful diffusion of a product does not guarantee high profits for its first supplier. Computers changed the world, but personal computer manufacturers competed for low margins for many years. The internet radically expanded markets, yet a huge number of early investors did not recover their capital. The automotive industry created the foundation of the modern world, but far from every car company became an attractive long-term investment.
I view AI through the same lens. Suppose the cost of training and operating models declines substantially over time. For society this is excellent: more companies will be able to adopt the technology, productivity will rise, access to tools will expand. But what will happen to the margins of the developers of the most expensive models? If basic capabilities become a commodity, prices may fall faster than the costs of individual players. The winner will not necessarily be the one who first demonstrated the most impressive model. The winner will be the one who managed to turn the technology into a sustainable advantage in distribution, data, customer base or operational efficiency.
Nor should we forget network effects. A company that controls a corporate ecosystem, authentication system, document workflow and distribution channel sometimes has greater ability to monetize AI than an independent laboratory with a more advanced product. This explains why I watch Microsoft, Alphabet and Amazon so closely. Their advantage lies not only in access to computing power but also in the enormous number of already existing corporate relationships. However, the existing business also creates cannibalization risk: an expensive AI feature may replace revenue from an old product instead of creating an entirely new category of demand.
Therefore my thesis is not that AI companies are too expensive without exception. My thesis is technological victory and shareholder victory are not the same. Technology can become free or extremely cheap for the user precisely because suppliers are engaged in a fierce price war. For the consumer this is a benefit; for the holder of overvalued shares it is not always.
Energy, water, land: the investment story extends beyond software
In conversations about artificial intelligence the market sometimes behaves as though computing exists literally in the cloud: without a physical address, temperature or electricity bill. In practice every large deployment requires land, grid access, capacity reservations, substations, water or alternative cooling systems, cables and maintenance. These are not abstract expenses but concrete projects with permits, construction delays and local politics.
Growing demand for electricity can support companies supplying electrical equipment, transformers, distribution systems and cooling solutions. But here too there is a trap: if a company sells a critical component, the market often prices it in advance as a future monopolist. Then it turns out that expanding production brings in competitors, increases supply and customers bargain over price. I would watch especially closely the ability of infrastructure suppliers to maintain margins after capacity expansion.
There is also a conflict between short-term energy security and decarbonization. Hyperscalers may announce a bet on clean energy, but real connection timelines and the availability of continuous capacity can make gas generation or long-term contracts with nuclear operators commercially attractive. The result is new competition for fuel, equipment and the grid. For the economy this means that the AI boom affects capital goods inflation, utility tariffs and regional infrastructure, even if profits from new digital services have not yet fully materialized.
This effect is two-way. If construction accelerates, generation expands and the cost of computing falls, end users and developers gain additional room for growth. But if projects are delayed by shortages of energy capacity, the investor is effectively financing expensive assets that are not yet operating. Such risk cannot be eliminated by advertising about the revolution.
How I would argue with a convinced advocate of endless AI growth
A supporter of the AI rally would rightly say: technological transformation can create new markets, existing profit forecasts may systematically underestimate demand, and the largest companies have balance sheets that most dot-com bubble participants lacked. These are serious arguments. A high stock valuation by itself does not yet mean a bubble - sometimes it reflects high profitability, a strong competitive position and a long growth period.
My answer is that the best arguments for the technology do not cancel the need to count the entry price. If high returns assume maintaining exceptional margins for ten years, one must show why competitors will not be able to erode them. If the model assumes continuous growth in clients' computing budgets, one must explain through what final savings or new revenue these clients will pay for them. And if a company is today burning enormous resources for infrastructure, one must show that its cost of capital will not swallow the expected benefit.
I would not build the analysis on a forecast of one specific day when “the bubble bursts.” The market can remain overheated for a surprisingly long time, especially when the cash flows of the leaders are real. It is more productive to identify signs of deterioration in the investment thesis. For example, a doubling of capital expenditures with an almost unchanged free cash flow forecast is a reason to recheck returns. Growth in leverage accompanied by declining infrastructure utilization is an even more significant signal. Conversely, rising margins and confirmed willingness of enterprises to pay for AI tools strengthen the foundation.
There is a difference between an analyst and a prophet. A prophet announces in advance the date of the end of a technological boom; an analyst understands what would make him change his mind. I prefer the second approach, even if it looks worse on a cover.
Conclusion
Investing in AI simply because the future lies with it is roughly like buying any building near a railway station simply because trains are not going anywhere. The question is always price, utilization and return on capital.
Artificial intelligence may well become the main industrial transformation of the decade. But that is precisely why it requires not religious enthusiasm but financial discipline. My priority is to distinguish companies that sell equipment and services with proven economics from companies that are still selling expectations of a future economy. If an investor can draw this line, the AI bubble ceases to be a slogan for him and becomes a manageable set of scenarios.
The article was prepared by senior financial analyst of Tradepas Boris Sergeyevich Katz