The AI Economy: Infrastructure Owners Win

The first great fortune of the artificial intelligence era will reward those who own the servers more than those who build the most brilliant chatbot. The new AI economy is redrawing the map of capital.

I have been studying capital flows for years. The pattern here is ancient; the technological disguise is recent. My previous analyses remain on the blog.

The Precedent: 1849, Those Who Sold the Shovels

In 1849, in California, thousands of men searched for gold. Few found it.

Samuel Brannan read the mechanism before anyone else. He bought every shovel, every pickaxe and every sieve available in San Francisco, then resold them to prospectors at multiplied prices. He became California's first millionaire.

The mechanism was elementary: sell the tools. Whoever controlled the supply captured the flow, regardless of the outcome of any individual search. Gold enriched a few lucky ones; the shovels enriched those who owned them. The lesson outlives the metal.

This is the point the public forgets every time. Durable wealth follows ownership of the scarce asset, more than skill in using it. Brannan himself found zero grams of gold, and his fortune grew regardless. Every prospector who went down to the river passed through his warehouse first. The flow was guaranteed by structure rather than by luck.

The Current Pattern: 600 Billion in 2026

In 2026 the same mechanism is active again. The context is different; the structure remains identical.

Goldman Sachs economists estimate that investment in artificial intelligence will reach approximately 600 billion dollars in 2026, according to the analysis reported by Elysian, distributed across semiconductors, data centers, cloud, energy and increasingly sophisticated models.

The numbers are impressive. So is the opportunity. Beneath the enthusiasm lies a more substantial question: who actually gets rich from this boom?

The answer lies in the distinction between use and ownership. The shovels of AI have a precise name: fabs, data centers, networks, power plants, intellectual property. Each of these items requires capital that few possess. Each captures a share of the flow regardless of which application wins.

The Mechanism: Capital Versus Labor

For most people, AI remains a tool. It writes emails, analyzes data, generates images, automates administrative work. The economic benefit can be real.

There is a second, more decisive layer: ownership. Those who own the companies, the infrastructure and the intellectual property collect the payments of everyone else.

International Monetary Fund research has long documented the shift in income share from labor toward capital. AI accelerates this structural dynamic. Capital income grows faster than labor income.

The consequence for the reader is direct. Those who sell their time alone, even with the tool's boost, remain on the slow side of the curve. Those who own the asset remain on the fast side. The tool improves individual productivity; ownership captures aggregate value.

The real divide, therefore, concerns how much capital each person can invest, which assets they can reach and what share of income comes from ownership. This asymmetry defines the wealth gap of the next decade.

The Divergence the Market Is Ignoring

Since 2023, two curves have been moving in opposite directions. Access to AI is democratizing; ownership of AI assets is concentrating.

A professional subscribes to a platform for the cost of a dinner. An investor owns shares in the companies providing the chips, the electricity and the computing capacity behind that platform. Institutional investors also reach private companies, venture funds and infrastructure operations beyond the reach of ordinary families.

The divergence between access and ownership always resolves. The question is how. In documented precedents, the resolution rewards those who control the scarce resource.

The market has already priced in the growth of applications. The market has yet to price in the scarcity of physical infrastructure.

The Regime, Beyond the Cycle

Let me clarify the nature of the phenomenon. This goes beyond the cycle. It is a regime change.

The interest rate cycle of 2020-2026 has closed the thirty-year bond rally. The regime that follows differs from everything current models have been calibrated for.

In a world of more expensive capital, rent shifts toward assets with physical barriers to entry. Fabs, energy and the grid become the new long-duration holdings: scarce, defensive, remunerated by structural demand.

Most asset managers treat AI as a growth theme. I read it as a theme of rent and scarcity. The two readings diverge in portfolio allocation. The first chases application multiples. The second buys the physical bottleneck that every application must pass through.

My Position

I put forward a clear thesis: over the current decade, capital invested in the physical infrastructure of AI will beat capital invested in software applications.

The reason is structural. Models replicate easily; chip factories take years and tens of billions of dollars to build. Those who control TSMC and Samsung govern the outcome of the technology competition, more than those who train the best model.

US sanctions on GPUs toward Huawei confirm this reading. Real power resides in access to the fabs, a geographically concentrated and politically contested resource. Three historical precedents of industrial scarcity are enough to call it a pattern.

What would change my view? A technological leap that makes model training independent of frontier semiconductors. Until that point, physical scarcity commands.

Three Implications for Capital

For those allocating capital, three operational consequences emerge clearly.

Family offices and sovereign funds (36-month horizon): shift weight toward owners of physical infrastructure, energy producers and data center operators. Scarcity rent rewards ownership more than use.

CEOs and boards (18-month horizon): geopolitical risk along the semiconductor chain remains absent from many strategic plans. A disruption in Taiwan would hit revenues ahead of the VAR models.

CFOs and investor relations (18-month horizon): the narrative that “AI reduces costs”, brought to investors, risks aging fast. The real question becomes who owns the asset that generates the rent.

The Forecast

I formulate a verifiable forecast, with an explicit horizon and indicator.

By December 31, 2027, the Philadelphia Semiconductor Index (SOX) will outperform an equal-weighted basket of AI software application companies, measured in total return from January 1, 2026.

Confidence: 68%. Horizon: December 31, 2027. Verification: the compared total return of the two baskets. An underperformance of the SOX would archive the thesis.

What to watch

Three leading indicators will confirm or refute the thesis over the coming quarters.

• Capital expenditure by the large cloud providers, quarter over quarter.
• Delivery times and prices of high-end GPUs.
• Industrial electricity rates in areas with a high density of data centers.

These signals precede the repricing of infrastructure assets by several months. Watch physical scarcity. Ignore the noise of the demos.