The $3.7 Trillion Demand Backstop

Why the AI Infrastructure Cycle Is Not What the Crowd Thinks It Is

Dear Partners,

There is an old story about a man who discovers oil on his property. Neighbours arrive. Speculators arrive. Someone builds a derrick. Someone else builds a pipeline. Within months, an entire town springs up around the well. Everyone is talking about the oil. But almost nobody is asking the right question: who is buying it, and how much do they need?

The artificial intelligence infrastructure cycle has reached a similar moment. Headlines are dominated by the sheer scale of spending. Hyperscalers are pouring hundreds of billions into data centres, GPUs, and power capacity. The numbers are staggering and they are accelerating. But the conversation is stuck on the supply side. How much are they building? Is it too much? Will there be a bust?

These are natural questions. They are also the wrong ones. The right question is simpler and far more important: what is the demand backstop? What economic force, already in motion, guarantees that this infrastructure will be used?

The answer is not a technology trend. It is a labour market fact. There are approximately $3.7 trillion in annual wages paid to knowledge workers in the United States alone. These are the salaries of managers, analysts, accountants, lawyers, software engineers, marketers, and administrators whose work product is fundamentally digital. Their output lives on screens. Their tools are keyboards. Their deliverables are documents, spreadsheets, code, and decisions.

That $3.7 trillion is the demand backstop. It is the gravitational force that makes this infrastructure cycle structurally different from every previous one. And almost nobody is framing it this way.

The Wage Bill Is the Signal

Let me explain the arithmetic, because it is disarmingly simple.

The Bureau of Labour Statistics publishes detailed occupational data every year. When you isolate the categories of work most exposed to AI, the numbers are large and concrete. Management occupations alone account for roughly $1.55 trillion in annual wages across nearly 11 million workers. Business and financial operations add another $970 billion. Computer and mathematical roles contribute $607 billion. Legal professions add $175 billion. That subtotal is already $3.3 trillion, and it excludes the vast majority of administrative, design, and engineering roles that are also highly exposed.

The $3.7 trillion figure is not a forecast. It is not a projection based on optimistic adoption curves. It is the current annual cost of cognitive labour in the American economy. It exists today, right now, on the payrolls of every corporation in the country.

Now consider what happens when even a small fraction of that wage bill converts into compute demand. At 3 percent, you get $111 billion in annual AI infrastructure spending. At 5 percent, you get $185 billion. These are not aspirational numbers. They are arithmetic.

Think of it the way a farmer thinks about irrigation. You do not ask whether a river will flow. You measure the watershed. The watershed here is $3.7 trillion in wages already being paid to do work that AI can increasingly augment or perform. The river will flow because the water is already there. The only question is how quickly and through which channels.

Why the Bears Are Asking the Wrong Question

The sceptics have a framework, and it is not unreasonable on its surface. They look at the revenue currently generated by AI products and compare it to the capital being deployed. The gap is enormous. One prominent venture firm estimated that end users would need to generate over $600 billion in annual revenue to justify current infrastructure levels. By that measure, you would need thousands of products the size of ChatGPT just to break even.

This analysis is precise and wrong. It makes the classic mistake of measuring a new technology by the revenue model of the old one. It assumes the return on AI infrastructure must come from AI-native subscription revenue, from chatbot fees and API calls. That is like measuring the return on electricity by counting lightbulb sales while ignoring the factories, refrigerators, and cities that electricity would eventually power.

The real return on AI infrastructure will not come primarily from selling AI as a product. It will come from displacing, augmenting, and restructuring $3.7 trillion in existing labour costs. The demand signal is not new revenue. It is the redirection of old spending.

A law firm that pays $500 per hour for associate research does not need a new AI product to generate revenue. It needs an AI tool that does that research in four minutes instead of four hours. The saving is not recorded as AI revenue. It is recorded as higher margins, fewer hires, or greater throughput. The compute cost is buried inside the firm’s technology budget, invisible to anyone counting chatbot subscriptions.

This is why the bear case feels compelling but misses the structural point. It is measuring the wrong river.

The Supply Side Tells You Everything

If there is one qualitative data point that matters more than any financial model, it is this: not a single major hyperscaler reports excess capacity. Not one.

Google’s chief financial officer said publicly that the company exited 2025 with more demand than available capacity. Google Cloud’s backlog surged past $240 billion, a 55 percent increase in a single quarter. Amazon Web Services reported $200 billion in total backlog and added nearly 4 gigawatts of capacity in twelve months. Microsoft disclosed $80 billion in unfulfilled Azure AI orders, constrained not by demand but by the physical limits of power availability. Oracle’s remaining performance obligations hit $455 billion, up 359 percent year over year.

Read those numbers again. These are not speculative. They are contractual commitments, signed and waiting for infrastructure that does not yet exist. The binding constraint is not whether customers want the compute. It is whether the hyperscalers can build fast enough to deliver it.

Think of a port city during a trade boom. Ships are queuing outside the harbour. Warehouses are full. The cranes run day and night and still the backlog grows. You can debate whether the boom will last forever. But you cannot credibly argue that the port is overbuilt while ships are still waiting to dock.

The Capex Acceleration Is Unprecedented

The combined capital expenditure of the five largest hyperscalers, Amazon, Alphabet, Microsoft, Meta, and Oracle, has tripled in two years. In 2023, the big four spent approximately $155 billion. In 2024, that figure rose to $256 billion. For 2025, the combined total reached approximately $443 billion. Projections for 2026 range from $602 billion to $690 billion.

To put this in context, Goldman Sachs estimates that cumulative hyperscaler capital expenditure from 2025 through 2027 will exceed $1.15 trillion. That is more than double the $477 billion spent in the prior three years. And analyst estimates have proven too low for two consecutive years, with actual spending exceeding initial consensus by more than 30 percentage points each time.

Yet even at these levels, AI-related capital expenditure represents only about 0.8 percent of GDP. The telecom build-out of the late 1990s peaked at 1.0 to 1.2 percent of GDP. Reaching that level would require approximately $700 billion in a single year. We are not there yet.

I find this comparison instructive. The 1990s telecom build-out is widely regarded as having created excess capacity. But the infrastructure it left behind, the fibre-optic backbone of the internet, powered the next two decades of digital commerce. Even the “overbuild” turned out to be an underbuild when measured against actual long-term demand. The AI infrastructure cycle has the same structural character, except the demand backstop is more visible and more immediate than it was for broadband in 1998.

Power Is the Real Bottleneck

There is a natural governor on this cycle that prevents the kind of unconstrained overbuild that defined past technology booms. That governor is electricity.

The International Energy Agency estimates that global data centre electricity consumption reached 415 terawatt hours in 2024, roughly 1.5 percent of all electricity generated worldwide. By 2030, that figure is projected to reach 945 terawatt hours, equivalent to the total electricity consumption of Japan.

In the United States, data centre grid demand is expected to more than double from 62 gigawatts in 2025 to 134 gigawatts by 2030. The response from hyperscalers has been extraordinary. Microsoft signed a 20-year agreement to restart a unit at Three Mile Island. Amazon secured a 17-year power purchase agreement for nearly 2 gigawatts from the Susquehanna nuclear plant. Meta committed to a 20-year agreement for over a gigawatt of nuclear capacity.

When companies sign 20-year nuclear power agreements, they are not speculating. They are making irreversible commitments that only make sense if they believe demand will persist for decades. A farmer who builds an irrigation canal is not gambling on next season’s rain. He is investing in the permanent fertility of his land.

The Conversion Is Already Beginning

The theoretical case for AI exposure of knowledge work is well established. A study from OpenAI and the University of Pennsylvania found that roughly 19 percent of American workers have more than half their tasks exposed to AI. Applied to the total US wage bill, the highly exposed segment equates to approximately $3.4 to $4.0 trillion, consistent with the $3.7 trillion figure.

But theory is not what interests me. What interests me is what is already happening on the ground.

Anthropic published research in March 2026 that draws a crucial distinction between theoretical exposure and observed exposure. In computer and mathematical occupations, theoretical AI exposure is 94 percent. Observed exposure, meaning what is actually being automated today, is 33 percent. That gap between 94 and 33 is not a failure of the technology. It is the conversion opportunity. It is the distance between where we are and where the economics are pulling us.

The Dallas Federal Reserve added another layer in February 2026. Employment in the most AI-exposed sectors has already declined by 1 percent since late 2022. That decline falls disproportionately on workers under 25. It is not mass layoffs. It is the quiet cessation of hiring at the entry level, exactly the roles most susceptible to AI augmentation. For workers over 30, employment in these same sectors is growing. AI is not eliminating experienced judgment. It is substituting for the routine cognitive work that used to require junior staff.

This pattern is important because it reveals how the conversion actually works. It is not dramatic. It is not front-page news. It is a gradual reallocation of spending, one hiring decision at a time, one software subscription at a time, one workflow automation at a time. The $3.7 trillion does not convert overnight. It converts the way compound interest builds, slowly and then remarkably.

Where the Infrastructure Pure-Plays Sit

For those of you who follow our portfolio, you know we have held positions in companies that sit directly in the path of this demand. I want to discuss three infrastructure businesses and explain why I believe they occupy a position similar to owning toll roads at the beginning of a continental trade route.

IREN (formerly Iris Energy). We have discussed IREN in previous letters, and the thesis has only strengthened. The company tripled its operational data centre capacity to 810 megawatts during 2025 and grew revenue 168 percent year over year to $501 million. Total grid-connected power now stands at approximately 2.9 gigawatts, of which less than 20 percent is currently utilised. That is not a weakness. It is a reservoir. IREN has pivoted aggressively from bitcoin mining toward AI compute, secured status as an NVIDIA preferred partner, and is building a 75-megawatt liquid-cooled facility designed for next-generation Blackwell GPUs at 200 kilowatts per rack. The company is targeting $500 million in annualised AI cloud revenue by the first quarter of 2026. When you own the land and the power in a supply-constrained market, you do not need to predict the future with precision. You simply need to be standing in the right place when the demand arrives.

CoreWeave (CRWV). CoreWeave went public in March 2025 and has rapidly become the most visible pure-play AI infrastructure company in the market. Revenue reached $5.13 billion in 2025, up 168 percent year over year, with 2026 guidance of $12 to $13 billion. The revenue backlog stands at $66.8 billion, roughly four times where it began the year. Major contracts include OpenAI, Meta, and Microsoft. The company operates 850 megawatts of capacity with 3.1 gigawatts contracted, and was the first cloud provider to offer NVIDIA’s GB200 NVL72 systems. CoreWeave is not selling a dream. It is selling compute to the largest and most sophisticated AI buyers in the world, under long-term contracts, at industrial scale.

Cipher Digital (CIFR). Cipher is the cleanest example of a business transforming itself from bitcoin mining into AI infrastructure. The company formally exited mining in February 2026 and has committed entirely to high-performance computing. Total contracted HPC revenue stands at $9.3 billion with projected average annualised net operating income of $669 million over ten years. The anchor contract is a 15-year lease with Amazon Web Services worth approximately $5.5 billion covering 300 megawatts. An additional $3.8 billion contract with a major cloud provider covers another 300 megawatts at a facility in Texas. The total development pipeline spans 3.4 gigawatts across eight sites. In the space of a year, Cipher has transformed from a commodity mining operation into a contracted infrastructure platform with decade-long revenue visibility. The stock has reflected this, rising over 300 percent in six months.

The Philosophy Behind the Position

I want to step back from the data for a moment and explain why this thesis appeals to me philosophically, not just analytically.

Throughout my career, I have been drawn to situations where the underlying asset is real and the market is distracted by the wrong narrative. In the early days of Soar Aviation, no one believed you could crowdfund an aircraft. The narrative was about the novelty of the funding mechanism. But what mattered was the physical asset and the structural demand for pilot training. The narrative was noise. The asset was signal.

The AI infrastructure cycle presents the same dynamic at a vastly larger scale. The narrative is about whether AI companies can generate enough subscription revenue to justify the build-out. But that is the wrong frame. The signal is the $3.7 trillion wage bill that already exists, the physical infrastructure being built to serve it, and the contractual backlogs that prove demand is not speculative.

Our investment principles have not changed. We still look for what is asset-backed, what is cheap relative to the underlying value, what pays us to wait, what the crowd has neglected, and what is run by people who have their own capital at risk. The AI infrastructure companies we own meet these criteria in the same way that IREN met them when we first invested at $5 per share. The numbers are larger now, but the logic is identical.

There is a passage in Munger’s work that I return to often. He argues that the key to successful investing is not superior intelligence. It is a superior framework, applied with consistency, over a long period of time. The framework we use is designed to identify situations where the physical world is moving in one direction while the financial narrative is stuck debating something else entirely. The $3.7 trillion demand backstop is exactly that kind of situation.

What I Am Watching

I am not naive about the risks. Every infrastructure cycle carries the possibility of excess. The question is always whether the demand is real enough and durable enough to absorb the supply. Here, the evidence is overwhelmingly in favour.

But I am watching three things carefully. First, the pace of enterprise adoption. It is one thing for 78 percent of companies to say they use AI in some function. It is another for that usage to translate into sustained compute spending. The gap between experimentation and integration is where hype cycles die. Second, the power constraints. If permitting and grid expansion lag too far behind, even the strongest demand cannot be served, and the infrastructure companies we own will see their timelines elongated. Third, the health of the hyperscaler balance sheets. These companies are spending 90 percent of operating cash flow on capital expenditure. That is sustainable as long as revenue growth continues. If it stalls, the capex commitments become a burden rather than an asset.

None of these risks are disqualifying. They are the kinds of uncertainties that create opportunity for investors who are willing to do the work and hold the positions through volatility. The farmer does not abandon his fields because of a weather forecast. He checks his irrigation, inspects his fences, and trusts the soil.

Final Thoughts

The $3.7 trillion demand backstop is not a prediction. It is a measurement of the world as it exists today. Knowledge workers are expensive. Their work is increasingly digital. AI can perform meaningful portions of that work at a fraction of the cost. The infrastructure required to deliver that AI is being built at an unprecedented pace, and it is still not being built fast enough.

I do not know which AI model will be dominant in five years. I do not know which application layer company will capture the most value. Those are important questions, but they are not our questions. Our question is simpler: who owns the physical infrastructure that every model, every application, and every enterprise will need? That is where we want to be.

It is the same logic that has guided us from the beginning. Own what is real. Buy it when others are distracted. Be patient enough to let the world catch up to the arithmetic. And never confuse the noise of the crowd with the signal in the soil.

Thank you for your continued trust and partnership. I will keep investing with the same discipline, the same principles, and the same quiet conviction that the best outcomes come not from chasing what is exciting, but from owning what is essential.

With warm regards and steadfast dedication,

Neel Khokhani