You can look at a single piece of silicon and see the future of finance. Over the last two years, a niche corner of lending has exploded into a multi-hundred-billion dollar market. We are not talking about data centers, those familiar physical shells. This is about what goes inside them, the specialized chips like GPUs and TPUs that power artificial intelligence. The scale of capital required to feed the AI boom has forced Wall Street to invent entirely new financial structures. In doing so, it has effectively created a new asset class.
The numbers are staggering. Goldman Sachs estimates the total bill for the AI buildout, covering chips, power, and data centers, could reach $7 trillion between 2026 and 2031. Hyperscalers like Google, Meta, and Amazon are projected to spend over $1 trillion on capital expenditures next year alone. This scale of investment cannot be funded by traditional bank loans. It demands capital from every available source: private equity, sovereign wealth funds, public bonds, and private credit. As Dan Dees, co-head of global banking and markets at Goldman Sachs, noted, it requires “new joint venture structures, some of which have yet to be invented.”
These deals are massive because the chips themselves are breathtakingly expensive. Jake Streit, a managing director at SMBC, puts it in perspective. A one-gigawatt data center might cost $10 billion. The chips to fill it can cost $35 to $45 billion. SMBC estimates about $130 billion in chips have been financed to date. Streit believes that figure will approach half a trillion dollars within three to four years. Theoretically, a single chip financing deal could reach $100 billion. We are moving far beyond the realm of simple equipment leasing.
The initial barrier for investors was uncertainty. What is the useful life of a chip when technology evolves every six months? What is its residual value as collateral if a loan goes bad? To solve this, deal architects have borrowed techniques from more mature markets like infrastructure and aviation finance. The most critical innovation has been the corporate guarantee.
Consider two landmark deals. In March, an $8.5 billion GPU financing for CoreWeave closed, backed by a tight master services agreement with Meta, an investment-grade company. In June, a $35 billion TPU deal for Anthropic closed, using the chips as collateral with a residual value guarantee from Broadcom, another investment-grade firm. This means Broadcom promises to make lenders whole if the chips lose value. These guarantees have allowed portions of these deals to secure investment-grade ratings, opening the floodgates to a much wider pool of institutional capital.
This is why many now see chip financing as distinct from data center financing. Rob Horn, global head of infrastructure and asset-based credit at Blackstone, draws a clear line. A data center has a useful life of 20 to 30 years. The technology inside it evolves rapidly, with a much shorter assumed lifespan. The debt is structured accordingly, with terms kept short to match contracts and conservative estimates of the chip’s life. The whole structure, as Matt Eagan of Loomis Sayles describes, is designed to deliver financing to an end-user who might not be creditworthy on their own. It is becoming a new standard.
Rating agencies are scrambling to define it. S&P Global categorizes these deals under “digital infrastructure,” developing new methodologies that blend aspects of infrastructure and aviation finance. KBRA uses the term “AI/digital infrastructure.” Fitch applies a flexible framework focused on cashflow coverage and sponsor linkage. For lenders like Brookfield’s Eric Wittleder, the underwriting focus is on the investment-grade contracted cash flows from the hyperscaler. That is what makes it infrastructure finance, not mere equipment rental.
But with explosive growth comes new risk. The model often relies on a highly-rated chip maker guaranteeing the debt of a less-established AI company. Raj Joshi of Moody’s points out a dynamic where chip companies are entering the market to “unlock” demand. The danger is that a guarantor like Broadcom or Nvidia could overextend itself across too many deals. If several AI clients falter simultaneously, the guarantor could face obligations it cannot meet. Lenders would then be stuck with specialized chips whose value is hard to quantify.
Riley Thompson of Mitsubishi HC Capital America highlights another complexity. The ultimate user of the compute power can be unclear, especially if the primary customer resells capacity to third parties. The source of repayment should be the user, with the guarantee as a backstop. If the user’s business model is opaque, the risk calculations fall apart. Lenders simply do not have the manpower to conduct deep legal reviews on every novel guarantee agreement in this fast-moving market.
The wild card is the chips themselves. Their rapid evolution defies traditional asset valuation. “Where else do you see a product improving 50 times every six to 12 months?” asks Joshi. Yet, paradoxically, this has not always led to steep depreciation. Some older chips have retained or even increased in value. Bernie Margulies, CEO of American Compute, explains why. Newer chip generations are so power-hungry they require data centers that do not yet exist. Older, less power-intensive chips are often the only viable option for many companies, creating a robust secondary market.
This secondary market provides a cushion, but lenders are not counting on it. The dominant structure ensures debt is fully paid down within the contract life, often just a few years. The potential residual value of the chip is a bonus, not a foundation. As Horn at Blackstone notes, the underlying compute has retained its value better than expected. But the deal is built on contracted cash flows.
We are witnessing the birth of a hybrid. It is part asset-backed security, part project finance, and part corporate credit. It demands expertise in semiconductor technology, structured credit, and contract law. For investors, it offers a new way to gain exposure to the AI megatrend through diversified risk tranches, from investment-grade debt to higher-yield private credit. For the economy, it is the essential plumbing that funds the engines of artificial intelligence. The machines are learning, and Wall Street has just taught itself a new language to pay for them.
- AI buildout could reach $7 trillion
- Hyperscalers projected to spend over $1 trillion
- Initial barrier for investors: uncertainty
- Two landmark deals in GPU and TPU financing
- Rapid evolution of chips defies traditional valuation
- Hybrid financial structures are emerging
| Deal | Amount | Purpose | Backer |
|---|---|---|---|
| CoreWeave GPU Financing | $8.5 billion | GPU capital financing | Meta |
| Anthropic TPU Deal | $35 billion | TPU capital financing | Broadcom |
