The Fine Print of the AI Buildout

Programming note: ARPU will return next Thursday to look at the US-China AI race.

Project Finance Comes to Silicon Valley

We talk a lot here about how the AI era has murdered the concept of "capital-light" software. But for decades, that was the entire foundational appeal of the technology sector: you hired some smart engineers, gave them MacBook Pros, wrote some code, and sold it a million times at an 80% gross margin.

Because there were no factories to build or physical assets to finance, tech companies rarely had to touch the bond market. Debt was something messy, reserved for airline executives, pipeline operators, and real estate developers.

Now, Big Tech is building the largest physical infrastructure project in human history. They are pouring millions of tons of concrete, ordering customized gas turbines, and securing power grids capable of lighting up midsize countries. By all rights, this should turn the tech giants into debt-heavy industrial utilities.

Except tech CFOs are much too clever for that. Instead of ruining their pristine balance sheets with conventional corporate debt, they have quietly borrowed one of the most sophisticated financial tools in Wall Street history: project finance.

The masterclass in this new playbook was Meta's "Hyperion" data center in Louisiana—a facility the size of 1,700 football fields.¹ Here is the Financial Times explaining how Meta pulled it off:

...instead of just issuing bonds off its own balance sheet, Meta formed a joint venture with Blue Owl called Beignet that would develop and own Hyperion. Meta owns just 20 per cent of Beignet, but made a rock-hard commitment to lease Hyperion for at least 20 years. That guarantee allowed Beignet to issue an amortising $27bn bond, but this debt doesn't actually appear as debt on Meta's balance sheet, even if it is on the hook for the payments.

The Art of the Uncommenced Lease

To understand why this is a stroke of financial elegance, you have to appreciate the beauty of corporate nesting dolls.

If you ask the bond investors who lent $27 billion to Beignet where their money is going, they will tell you they are taking "Meta risk." But if you ask Meta's accountants where the debt lives, they will tell you it belongs entirely to a separate entity named after a deep-fried pastry.

Meta did not invent this concept—energy companies have used Special Purpose Vehicles (SPVs) to finance oil pipelines and power plants for half a century. Big Tech has simply adopted that same playbook at an unprecedented scale, leaning directly into the specific mechanics of US accounting rules.

Under GAAP rule ASC 842, an operating lease does not hit a company's balance sheet as a liability until the lease actually starts. In plain English: if you sign an irrevocable, 20-year contract to rent a data center, but the developer is still waiting for power transformers to arrive, the debt does not officially exist. It is an "uncommenced lease," which means it lives entirely in Footnote 14 of the quarterly report.

And tech companies have signed a staggering number of them. Here is the Wall Street Journal on the scale of these off-balance-sheet commitments:

Nine top tech companies had some $3 trillion of off-balance-sheet commitments mostly related to AI, according to a Wall Street Journal analysis of footnotes in their most recent securities filings. Those obligations are growing faster than traditional 'capex,' which totaled about $600 billion over the past year they reported, and were about triple what the companies owe under their outstanding leases and long-term borrowings.

Of that $3 trillion, Goldman Sachs estimates that $1.5 trillion consists of lease commitments, with roughly $1 trillion tied to leases that haven't even started yet.

It is the ultimate corporate win-win. Private credit giants like Blue Owl, Blackstone, and KKR get to deploy tens of billions of dollars backed by the creditworthiness of the world's most profitable companies. And Big Tech gets to build a continent-spanning network of supercomputers while keeping its net leverage ratio at a modest 0.5 times—compared to 0.8 times for the broader tech sector, according to Morgan Stanley data cited by the FT.

Take-or-Pay to 2054

Leases, however, are only half of the tab. The other half consists of purchase obligations—the iron-clad promises tech companies make to buy chips, custom hardware, and electricity years in advance.

As reported by the WSJ, purchase commitments across the top tech giants have ballooned to nearly $1.9 trillion.

Alphabet offers the most eye-catching example: in the span of just three months, its contractual commitments more than doubled, exploding from $332 billion at the end of March to $811 billion by the end of June.

In its 10-Q filing, the company revealed that $707 billion of this consists of long-term supply agreements for hardware and "energy service agreements" with power utilities.

Crucially, some of these power agreements run as far out as the year 2054, and almost all of them include strict "take-or-pay" provisions. If Google uses the electricity to train a superintelligence, it pays the utility company. If enterprise AI demand cools off and half the server racks sit idle, Google still pays the utility company.

There is a rather humorous duration mismatch at play here. An AI accelerator sits on a five-to-six-year depreciation schedule. But to convince a utility to build a substation, Google has to sign an irrevocable power contract for the next 28 years.

The New Weight of Big Tech

This does not mean Big Tech is hiding trillions of dollars of secret debt. These commitments are disclosed, and many will eventually appear on the balance sheet as leases commence or purchases are delivered.

The more important point is that the usual headline numbers no longer tell the whole story. CapEx tells us what Big Tech is spending today. Conventional debt tells us how much it has borrowed. Neither fully captures the leases, power contracts, and purchase agreements already signed for tomorrow.

That is the deeper shift. Big Tech has not simply become more capital intensive. It is committing years of future cash flow to physical infrastructure whose economics are still being worked out.

[1] We recently published a YouTube explainer breaking down the engineering and physical constraints behind Meta's Hyperion project. You can watch it here.

Signal Stack

The operating reality beneath the headlines.

• AI Cloud Firm Nebius Taps Bond Market With $4.5 Billion Convertible Offering (Bloomberg) – Nebius is raising $4.5 billion in convertible bonds—its second such raise this year after a $4.3 billion offering earlier in 2026—with the 2030 tranche priced at a 0% to 0.5% coupon, meaning investors are accepting almost no yield in exchange for equity upside on a data center operator whose shares are up nearly 150% year-to-date.
• Hyperscalers Tap External Financing as AI Capex Outruns Cash Flow (FactSet) – FactSet frames Alphabet's $84 billion equity raise as a race to secure capital before markets are asked to absorb the anticipated OpenAI and Anthropic IPOs. This implies that hyperscalers are now competing not just for chips, power, and land, but for finite investor capacity in the capital markets themselves.

📺 On Our Channel

Does Microsoft Have an AI Endgame?

In 2022, a private Satya Nadella email revealed a fear at the center of Microsoft's AI strategy: becoming a "very thin layer" on top of Nvidia—buying the chips, running the cloud, distributing the products, while the technology that mattered belonged to someone else. Four years later, Microsoft still depends heavily on Nvidia for chips, and on OpenAI for frontier models. We broke down what has to go right for Nadella to escape that thin-layer thesis.

Watch ARPU's deep dive on YouTube (28 Mins)

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.