Market Letter XIII: The multi-trillion dollar fantasy

“It is not reality that shapes human behavior, but the image of reality.”
— George Soros

Years ago, while studying, I wrote an essay on valuation bubbles in equity markets. At the time, my thinking leaned heavily on so-called noise trader theory. Its logic is relatively straightforward: irrational investors push stock prices toward unsustainable levels, after which more rational investors begin joining the trade simply because underperforming the market starts to hurt more than the risks associated with participating.

George Soros brought an additional layer of nuance into this discussion through his theory of reflexivity. The idea that markets do not merely reflect fundamentals but actively shape them is one of the most fascinating ways to understand how bubbles emerge. Once a company’s share price rises high enough, access to financing improves, talent becomes easier to attract, visibility increases, and often so does customer demand. High valuations are therefore not merely a consequence of a compelling narrative, they gradually become part of the narrative itself.

This is one of the reasons why bubbles can reinforce themselves and persist far longer than fundamental analysis would otherwise suggest.

Recently, I came across an idea that, in my view, captures this phenomenon even more elegantly. At the end of the day, a surprisingly large portion of valuation is simply built on imagination.

Anyone who has ever opened a discounted cash flow model understands how this works. Particularly in technology companies, assumptions can very quickly spiral out of control. If future cash flows are pushed far enough into the future, even a minor change in growth assumptions or cost of capital can work wonders for present value calculations. Suddenly, a company generating very little profit begins to look, at least on paper, capable of conquering the world.

This is precisely why imagination plays such a massive role in valuation bubbles. In order for a company to justify something like a 200x P/E multiple, investors need to hold a fairly strong conviction that the future will look dramatically different from the present.

A recent example of this is SpaceX. The company combines all the classic ingredients: a massive new addressable market, multiple simultaneous technological disruptions, and a founder around whom an almost cult-like following has emerged. This obviously does not mean SpaceX cannot eventually become the most valuable company in the world. That is entirely possible. But at this very moment, a substantial portion of its valuation rests simply on what people imagine the future will look like.

Perhaps the most fascinating aspect of bubbles, however, is not that people make mistakes. That happens all the time. What is more interesting is how markets gradually begin constructing themselves around those mistakes in ways that temporarily make them rational.

In the technology sector, this dynamic is visible in its purest form. If markets collectively begin believing that a particular company will dominate the future, that company can attract so much capital, talent, and visibility that it may genuinely become the eventual winner. At that point, imagination is no longer detached from reality, it becomes a mechanism that partially builds the future itself.

In today’s market environment, this phenomenon has only accelerated. Information moves in seconds, narratives are born on social media in real time, and algorithms together with momentum strategies amplify moves at unprecedented speed. In the past, investment themes developed over months or even years. Today, an entire market narrative can emerge practically over the course of a single weekend.

In an environment like this, one of the most underappreciated competitive advantages is patience.
Most market participants live quarter to quarter, constantly reacting to news flow while optimizing for extremely short time horizons. A long-term investor can exploit situations where markets either overprice hype or, alternatively, overreact to temporary setbacks.

In the 1930s, B.F. Skinner conducted his famous experiments where rats were given a lever. When pressing the lever always produced food, interest faded quickly. When food never arrived, interest naturally disappeared as well. But when the reward came only randomly, the rats pressed the lever almost endlessly.

Later, Wolfram Schultz demonstrated in similar research that dopamine response is not primarily a function of pleasure itself, but rather of the prediction error between expectations and actual outcomes. Variable rewards create a stronger dopamine response than guaranteed rewards.

This is precisely why gambling works so effectively. The connection to investing is fairly obvious.

Over short periods, the stock market often resembles gambling far more than investing. Particularly during bubbles, markets become little more than a collective game of musical chairs where investors chase rapid gains in pursuit of dopamine hits. Perhaps it is still better than feeding money into slot machines, but the underlying mentality is often not very far removed.

The AI Narrative

Over the past several years, artificial intelligence has gradually become the dominant narrative in equity markets. The story is naturally extremely compelling, as it combines many of the classic ingredients that historically have often led to bubbles.

First, we are dealing with a new technology whose potential impact on the global economy is genuinely enormous. Second, the scale of investment is extraordinary. Capital is not only flowing from the private sector, but governments and public institutions are also participating, as AI is increasingly being viewed as strategically critical infrastructure.

What makes the AI narrative somewhat different, however, is that it possesses characteristics not typically seen in every bubble. Most importantly, demand is currently so strong that it exceeds supply across almost the entire value chain. As a result, many companies positioned around AI-related investment are currently generating exceptionally strong earnings, which is why valuation multiples in many cases do not even appear particularly aggressive.

This has given rise to the current bottleneck investing theme. Markets are attempting to identify the critical choke points through which the entire investment wave must pass. Up until now, the greatest attention has focused on semiconductor companies, whose share price performance has been almost explosive due to demand for computing power and, particularly, scarcity in memory supply. Semiconductors, however, are far from the only bottleneck.

The discussion now extends to energy, electrical grids, cooling systems, rare earth materials, and practically every piece of physical infrastructure required to support AI. The narrative has gradually become surprisingly hard-asset intensive.

From an investor’s perspective, the challenge lies in the fact that the entire theme is inherently opaque.
Technology has always been a sector where imagination is given an unusually large amount of room to operate. Partly because very few investors truly understand what is happening behind the scenes. When understanding is limited, narrative begins to fill the vacuum left behind by analysis.

When it comes to AI itself, I do not personally see the problem in the technology. In my view, the technology has already reached a stage where its utility is undeniable, and its impact on the global economy will be enormous both today and in the future. This is a genuine industrial revolution. The potential problem lies in infrastructure.

At the moment, the hottest area of the market is not AI itself, but the physical capacity on top of which the entire ecosystem is being built. In practical terms, this means data centers.

Where Does The Money Come From?

Put simply, data center investments are currently being financed by three groups. First are the American hyperscalers: Microsoft, Amazon, Alphabet, and Meta, whose collective investment budgets are already moving toward the trillion-dollar range.

Second come external strategic financiers such as Oracle, SoftBank, and sovereign wealth funds.
The third group consists of infrastructure-focused capital allocators such as Blackstone, Brookfield, KKR, and Apollo, all of whom are financing the buildout of physical capacity.

Where is the money going?

Table 1: Estimated Cost Structure of a 1 GW Data Center (Source: Nvidia, SemiAnalysis, Semiconductor Industry Association, Microsoft, Alphabet, Amazon, Meta, Vertiv, Schneider Electric, Eaton Corporation, IBM)

When examining the cost structure of data centers, the first observation is fairly obvious.

Semiconductors account for a completely dominant share of total investment, something that has also been clearly reflected in both stock price performance and earnings expectations over the past two years.

As a result, enormous amounts of capital are currently flowing directly into the semiconductor sector.

The investment wave is, of course, benefiting the entire value chain from networking infrastructure and power distribution to cooling systems, but at this stage semiconductors remain by far the single largest beneficiary.

At the same time, this particular segment also appears to be the clearest candidate for a potential bubble.

Figure 1: Development of Semiconductor Index P/E Multiples and Earnings Expectations
(Source: Bloomberg)

If we look at current earnings growth expectations, it is difficult to categorize the situation as a traditional valuation bubble. These companies are generating genuinely strong earnings, and demand at the moment remains exceptionally robust.

What investors should perhaps be more concerned about is not a valuation bubble, but rather an earnings bubble. At this point, we return once again to investors’ tendency to extrapolate present conditions far too optimistically into the distant future.

In practice, the entire earnings growth story built around AI infrastructure rests almost entirely on a single assumption: that the current pace of investment will continue for years to come. And the scale involved is enormous.

Figure 2: Projected AI Investment Through 2030 Compared to the GDP of Various Countries
(Source: Business Insider, Goldman Sachs)

Goldman Sachs estimates that the AI investment wave could reach approximately $7.6 trillion over the next five years. IBM CEO Arvind Krishna, meanwhile, estimates that building a single one-gigawatt data center costs roughly $80 billion, of which as much as $60 billion is allocated to semiconductors. If, for example, 100 gigawatts of new capacity are built globally, we are already talking about roughly $8 trillion of investment.

Krishna summarized the problem rather well in my view: “There is no plausible ROI if everyone keeps scaling this way.” And this, ultimately, lies at the heart of the issue.

The current buildout of AI infrastructure resembles an arms race far more than a traditional investment cycle. This is not really a game of returns, but rather a competitive race in which every major technology company is trying to secure its position before the market’s eventual winners have been determined. In these kinds of races, capital destruction is almost inevitable.

What makes the situation even more problematic is that a large portion of these investments is directed toward physical infrastructure. Unlike software, this is hardware, where scalability is inherently limited and technological life cycles can be extremely short. Today’s GPU clusters may, in practice, become obsolete within just a few years.

At the same time, major American technology companies are currently allocating the majority of their operating cash flow toward these investments. Ironically, many of the world’s strongest cash flow businesses now find themselves in a position where they are effectively forced to direct nearly all free capital back into infrastructure buildout.

Put bluntly, enormous amounts of capital are currently being deployed extremely aggressively into assets whose economic lifespan may ultimately prove far shorter than expected.

The optimist’s counterargument is, of course, entirely logical. If companies such as Microsoft, Alphabet, or Amazon genuinely intend to integrate AI across their entire product portfolios, they need to own sufficient capacity to make that possible. Without infrastructure, the revenue potential underpinning the entire investment thesis simply does not materialize. The argument is perfectly valid, but the mathematics quickly become challenging.

Take Microsoft as an example. With roughly $120 billion in annual AI-related investment spending, the company effectively needs to generate tens of billions of additional operating profit each year simply to maintain reasonable returns on capital. In practice, this means AI must translate into price increases and entirely new revenue streams across the company’s entire product portfolio. And even then, a persistent structural risk remains in the background.

If AI gradually democratizes over time in the same way many previous technologies have, today’s most aggressive infrastructure investments may ultimately prove poorly rewarded. Open-source language models, continuously declining compute costs, and the broader commoditization of technology could eventually lead to a world where AI increasingly becomes more of a public utility than a genuine competitive advantage.

History offers many good examples of this dynamic. America’s railway boom created enormous societal value while simultaneously destroying substantial amounts of investor capital.

If this realization gradually begins to manifest itself in the market, the investment pace may no longer remain at one trillion dollars annually. And at that point, we may not be looking at the bursting of a valuation bubble. We may instead be looking at the bursting of an earnings bubble.

Finally, there is the broader question of whether the physical world can even scale alongside this pace of investment. A data center can purchase semiconductors relatively quickly, but building a 500 MW power plant is hardly an overnight process. If the energy does not exist, what exactly does one do with everything else?

There is, in other words, a long list of bottlenecks. They will eventually be solved one way or another, because capitalism tends to be remarkably effective at solving these kinds of problems. The real question is whether investors will remain patient enough once valuations are already elevated. They may not.

The Trillion-Dollar Question

Investors should approach the infrastructure theme with caution. The rally may very well continue for a long time, as we are likely still in the early stages of this investment cycle.

At the same time, sentiment remains extremely fragile. Any negative development related to slowing investment momentum, rising interest rates, or a general deterioration in risk appetite could quickly alter market dynamics.

We have ourselves participated in this theme in various ways, and up until now the largest beneficiaries have naturally been found within the semiconductor sector. Recently, however, we have begun gradually taking some profits off the table. Not because our conviction in AI itself has weakened, but because market behavior is slowly beginning to resemble a very familiar pattern once again.

Liquidity is being pulled from every corner of the market toward a single narrative. Investor attention narrows, the forest begins disappearing behind the trees, and analysis gradually turns into little more than price chasing. Once enough participants begin buying simply because others are buying, investing slowly begins transforming back into what it so often becomes during the later stages of bubbles: Gambling.

I remain of the view that the ultimate beneficiaries of AI will be found everywhere, and that the benefits will not be concentrated solely at the infrastructure layer. Perhaps quite the opposite.

And perhaps at this point, we return to where this entire piece began.

Markets are ultimately, above all else, collective psychological systems. Sometimes fundamentals determine price. Sometimes price begins determining fundamentals. And sometimes the collective imagination of investors constructs an entire reality that temporarily makes the impossible appear rational.

Artificial intelligence will almost certainly change the world permanently. That does not, however, mean that all the capital currently being deployed around it will generate attractive returns for its owners.

As Soros once said, the most money is made during bubbles.

As long as you are not the last one heading for the exit.

Niko Fagernäs

Toimitusjohtaja

+358 44 9818 644 niko.fagernas@proprius.fi

View profile