Two Ways to Slice the AI Pie

Programming note: ARPU will return next Friday to take another look at the financing of AI.

The Micro-Macro Split

If you want to understand how institutional investors are trying to process the $2 trillion AI infrastructure boom, a recent debate on X between two very different types of fund managers perfectly captured the tension.

In one corner is Gavin Baker, the Managing Partner at Atreides Management. He evaluates the market with a bottom-up perspective, focusing on company-level economics.

In the other corner is Andy Constan, the CIO of Damped Spring and a former strategist at Bridgewater. He is a global macro trader—someone who looks at the economy from the top down.

Both men are looking at the exact same data. Both agree that AI demand is real, GPUs are scarce, and the required capital expenditure over the next few years is going to be measured in the trillions of dollars. But because they inhabit completely different financial universes, they have arrived at completely opposite conclusions about whether this build-out is actually sustainable.

The debate essentially boils down to one question: Are the tech giants going to pay for this infrastructure with their own cash, or are they going to trigger a credit squeeze trying to finance it?

The Magic of Cash Flow

To a tech investor like Baker, the AI build-out is a classic corporate finance problem that is about to solve itself.

Right now, the credit markets are getting somewhat nervous about Big Tech. Spreads are widening, and bond investors are worrying that hyperscalers will have to take on massive amounts of debt to plug the gap between the money they make and the money they are spending on Nvidia chips.

Baker's argument is that the credit markets are overreacting because they are fundamentally mispricing Big Tech's future revenue.

His logic relies on the mechanics of cloud contracts. Right now, spot prices to rent GPU compute are reportedly at least twice as high as the older, contracted rates signed back in 2024. Hyperscalers are currently honoring those old contracts, which means they are severely "underearning."

But contracts expire. When those early deals roll off, the hyperscalers will simply reprice their renewals closer to the current, much higher spot rates.

To understand why this matters, you have to look at the math that is currently spooking the credit markets. Consensus estimates suggest hyperscalers will add 25 to 35 gigawatts of capacity by 2028. At $60 billion per gigawatt, that is roughly $1.5 to $2.2 trillion in total CapEx. Current estimates project they will generate $1.3 to $1.4 trillion in operating cash flow, leaving a scary funding gap of up to $700 billion.

But when you plug Baker's repricing assumption into that math, the financial panic disappears. If Microsoft, Google, and Amazon can just hike their rental prices by 2x as contracts renew, their operating cash flow will go through the roof. The funding gap vanishes. The hyperscalers will self-fund the entire AI revolution out of their own organic revenue.

For a tech investor looking at a single company's income statement, a 2x price hike is a no-brainer. The spreadsheet balances, the debt requirement evaporates, and the only real limit to human ambition is how fast you can pour the concrete for the data centers.

Money Has to Come From Somewhere

If you are a global macro trader, however, that price hike is not a simple no-brainer. It is a systemic liability.

Constan's objection is rooted in the principles of double-entry bookkeeping. Money is not created out of thin air. If Microsoft and Google are suddenly collecting hundreds of billions of dollars in new, repriced cloud revenue, someone else in the economy has to actually pay that bill.

This is where the bottom-up tech view tends to collide with the top-down macro view.

If AI labs and thousands of enterprise software companies have to pay twice as much for compute next year, their own cash balances go down. Where do they get the money to cover the difference? They either have to fire their own employees, charge their own customers higher prices, burn through their venture capital, or go to the banks to borrow more money.

Constan's point is that you cannot look at Microsoft's cash flow in a vacuum. At a scale of $2 trillion, you cannot rely entirely on organic cash flow, because the system as a whole does not magically generate trillions of dollars of new liquidity overnight.

If AI capex is genuinely that large, someone, somewhere, ultimately has to save less, sell assets, or borrow more.

And that is why the macro view cares deeply about credit spreads. If the AI ecosystem requires the broader economy to take on massive amounts of debt to afford these new GPU rental rates, then the cost of borrowing money matters immensely. Rising bond yields and widening credit spreads are not just market noise; they act as a financial governor. When capital becomes too expensive, marginal startups fail, enterprises cut their IT budgets, and the hyperscalers are suddenly forced to slow down their spending.

Is There Enough Pie?

At its core, the entire disagreement boils down to this:

Baker is effectively saying: Don't worry about hyperscaler financing. Cloud revenue and operating cash flow are going to explode as AI demand grows and old compute contracts reprice upward.

Constan replies: That may be completely true for the hyperscalers. But you're treating their future cash flow as though it appears out of thin air. Their customers have to actually generate the money to pay them.

Both frameworks are entirely coherent. They just measure different things.

The tech investor sees the AI boom as an engine of net-new economic expansion. End-user demand is already proven. Enterprises and consumers are willingly paying for software that makes them more productive. The cash velocity of the ecosystem is accelerating fast enough to make the entire infrastructure build-out self-financing.

The global macro trader sees an "adding-up" problem. Across the AI ecosystem, the chipmakers, the model builders, the hyperscalers, and the enterprise buyers all expect to extract enormous financial value from this transition.

But macroeconomic arithmetic is a zero-sum game until proven otherwise. Somebody's surging revenue is somebody else's surging expense. Unless AI accelerates global GDP growth enough to create a massive amount of new wealth, it is mathematically impossible for every layer of the stack to extract outsized returns at the same time. There simply isn't enough pie.

For the bottom-up investor, the only constraint on the AI boom is the laws of thermodynamics—how fast can we generate the electricity? For the top-down macro strategist, the constraint is the price of money—how much financial friction can the economy tolerate before the ecosystem chokes on its own compute bill?

Both investors are looking at the exact same $2 trillion build-out. We are about to find out which end of the telescope is actually in focus.

Signal Stack

The operating reality beneath the headlines.

• TSMC Sees 45% Sales Surge as AI Demand Stays Strong (CNBC) – High-performance computing—where TSMC books its AI chip sales—now accounts for 66% of the company's total revenue, meaning the world's most important chip manufacturer has become, in effect, an AI infrastructure company that happens to also make other chips.
• Wall Street Endorsed Jensen Huang's 'Big Concept' for AI (CNBC) – Nvidia and six of Wall Street's largest firms—Goldman Sachs, BlackRock, Blackstone, KKR, Apollo, and Brookfield—say they're willing to raise $500 billion or more to finance AI data center construction, marking a shift from corporate balance sheets funding the buildout to Wall Street treating GPU clusters as a securitizable asset class in their own right.

You received this message because you are subscribed to ARPU newsletter. You can unsubscribe using the link below.

If a friend forwarded you this message, sign up here to get it in your inbox.