Caveat Emptor: Huge Debt and Circular Financing Deals Dominate AI Build-Outs

By the Curmudgeon

Introduction:

AI infrastructure spending has become one of the most important (if not the most important) capital-market theme of the year, with data centers and related buildouts absorbing staggering amounts of capital and sending ripple effects across semiconductors, engineering firms, utilities, and energy producers. The boom has also become a major macroeconomic force, with JPMorgan Asset Management estimating that AI may be contributing roughly 1.1% to U.S. GDP.

Major cloud hyperscalers are projected to spend between $690 billion and $725 billion on capital expenditures in 2026, driven heavily by AI infrastructure. Total global AI spending across all categories is forecasted by Gartner to hit $2.59 trillion.

Big Tech Capital Expenditure (CapEx) in 2026:

· Amazon: Guided for roughly $200 billion in total capex build-out.

· Alphabet (Google): Estimated between $175 billion and $185 billion, backed by accelerated data center upgrades.

· Microsoft: Operating on an estimated $150 billion annualized run rate with a heavy infrastructure focus.

· Meta Platforms: Ramped up targets to between $115 billion and $135 billion.

· Oracle: Projected around $50 billion for cloud and server expansion.

The major financial question(s) should now shift from how big the potential AI opportunity is to where the money is coming from, how it is being financed, and how much leverage the system is quietly accumulating.

The Rise of AI Debt Financing:

A growing share of AI infrastructure expansion is being funded through debt -- bond issuance, structured leases, private credit, and other forms of borrowed capital. That pushes the AI sector away from asset-light software economics and toward old-fashioned capital intensity. Once capital expenditure consistently outruns free cash flow, the equity story becomes hostage to refinancing conditions, credit spreads, and the market’s willingness to keep funding the buildout at acceptable terms.

In a new research note from Goldman Sachs strategist Amanda Lynam today, she estimated that $489 billion of AI-related debt has been issued this year. That's already above Goldman's estimate of $322 billion for how much of this debt came to market last year. About 40% of this year's AI-related debt supply has been issued directly by hyperscalers. Data center financing and other parts of the tech ecosystem have also contributed. Check out this chart (IG=Investment Grade debt):

Chart Credit: Goldman Sachs

If the AI buildout increasingly depends on borrowed money, lease backstops, and strategic vendor funding, then the relevant question becomes not whether AI demand exists, but what happens when financing conditions tighten. In that scenario, the real pressure may show up first in bonds, structured credit, and data-center financing vehicles before it shows up in earnings.

Oracle is the clearest public-market cautionary tale because it has relied the most on debt to fund its AI build-outs. By contrast, Alphabet, Microsoft, Amazon, and Meta still have stronger balance sheets and more robust internal cash generation, but even they are now materially more capital intensive than they were a few years ago. That shift does not automatically make them bad investments. It simply means they no longer deserve to be analyzed as quasi-free-cash-flow monopolies.

Circular Funding Rules the Roost:

Meanwhile, an increasing number of transactions now exhibit circular funding characteristics, in which chipmakers, cloud service providers, and AI developers are financing one another’s growth in ways that can inflate demand visibility without necessarily creating durable end-market cash flow.

These deals are called circular because the money does not just fund growth; it also helps guarantee future purchases of chips, cloud capacity, or model access from the same firms providing the capital. That can accelerate AI infrastructure buildout, but it also raises questions about demand validation, accounting optics, and whether reported “market demand” is partly financed demand.

AI circular funding deals reminds me of a dog chasing its tail:

Image Credit: ChatGPT

…………………………………………………………………………………….

AMD – Anthropic Arrangement:

Today’s AMD-Anthropic announcement is a clean example of the AI ecosystem’s growing circularity: AMD is effectively financing a major customer while that same customer commits to buying tens of billions of dollars of AMD silicon over time. The structure looks strategic on the surface, but from a market perspective it also functions as demand support, balance-sheet reinforcement, and competitive positioning rolled into one.

Anthropic reportedly plans to deploy up to 2 gigawatts of AMD Instinct MI450 capacity, with the first gigawatt coming in the first half of 2027, while AMD will invest up to $5 billion in Anthropic, with that investment tied to deployment milestones. That matters because it turns the deal into a staged financing arrangement rather than a simple supplier contract.

This is why this deal belongs in the broader category of circular AI financing. AMD’s earlier OpenAI agreement reportedly involved stock-linked warrants, while this Anthropic transaction is the opposite direction: cash from the supplier into the customer, plus a very large customer commitment back to the supplier. In other words, the industry is increasingly replacing plain-vanilla demand with engineered demand.

……………………………………………………………………………………..

Other Examples:

The best-known examples are:

· Nvidia’s direct and indirect stakes in AI cloud providers like CoreWeave and Lambda which buy its GPUs.

· The broader OpenAI funding and infrastructure stack involving Microsoft, Oracle, SoftBank, and Stargate partners.

A few other circular funding deals include: Nvidia’s proposed investment in OpenAI, Microsoft and Nvidia’s support for Anthropic, Oracle’s exposure to OpenAI-linked demand, and a meaningful part of CoreWeave’s ecosystem. Nvidia investing in CoreWeave while CoreWeave buys Nvidia GPUs, and CoreWeave’s major customer contracts with OpenAI creating demand for the infrastructure it is financing.

These deals all point in the same direction: the same dollars are being recycled through a small club of AI participants, with each deal helping validate the next one.

…………………………………………………………………………………………

Conclusions:

Debt and circular deals is why the current AI build out cycle is not a normal technology upgrade but is much more akin to a speculative infrastructure boom.

The important point here is not the headline deal size, but the capital structure. AI infrastructure demand is increasingly being underwritten by the vendors themselves, which means the same dollars are often traveling through multiple layers of the ecosystem: chipmaker to model developer, model developer to cloud and data-center operators, and back again in the form of hardware orders, lease commitments, and financing support.

For investors, the key risk is that these structures can make revenue look more durable than it really is. A customer order backed by supplier capital is not the same thing as open-market adoption, and a multi-year deployment plan is not the same thing as free cash flow. If AI monetization slows, the market may discover that much of today’s “growth” was financed growth, not self-sustaining growth.

The bigger takeaway is that AI is beginning to look less like a software cycle and more like a highly financial shenanigans infrastructure cycle. That does not mean the technology is a bubble; it means the financing around it is becoming speculative enough that investors need to separate genuine end-demand from the recycled capital supporting it.

……………………………………………………………………………….

Sidebar- Historical Precedent- Vendor Financing 1998-2001:

Today’s AI circular funding arrangements are very similar to the vendor financing during the dotcom/fiber optic boom in the late 1990s. For example, Cisco Systems was selling routers and switches to shaky telecom and internet service providers [1.] while extending credit and lease-style financing to them. That effectively helped them buy Cisco gear with Cisco-backed money.

Note 1. Cisco lent money to cash-burning telecom carriers, ISPs, CLECs, and broadband operators such as Metrocomm, Rhythms NetConnections, and HarvardNet. Those firms often had weak balance sheets and little or no real earnings power, yet Cisco financing let them keep ordering product, which made demand look stronger than it truly was.

By 2000, Cisco executives said financing and leasing deals were about 10% of its annual revenue, and by early 2001 Cisco reportedly had about $1.3 billion of financed-lease exposure across 735 customers, including a meaningful slice tied to higher-risk borrowers.

………………………………………………………………………………………………

Wishing you good health, success and good luck. Till next time….

The Curmudgeon
ajwdct@gmail.com

Follow the Curmudgeon on Twitter @ajwdct247

Curmudgeon is a retired investment professional. He has been involved in financial markets since 1968 (yes, he cut his teeth on the 1968-1974 bear market), became an SEC Registered Investment Advisor in 1995, and received the Chartered Financial Analyst designation from AIMR (now CFA Institute) in 1996. He managed hedged equity and alternative (non-correlated) investment accounts for clients from 1992-2005.

Victor Sperandeo is a historian, economist and financial innovator who has re-invented himself and the companies he's owned (since 1971) to profit in the ever-changing and arcane world of markets, economies, and government policies. Victor started his Wall Street career in 1966 and began trading for a living in 1968. As President and CEO of Alpha Financial Technologies LLC, Sperandeo oversees the firm's research and development platform, which is used to create innovative solutions for different futures markets, risk parameters and other factors.

Readers are PROHIBITED from duplicating, copying, or reproducing article(s) written by The Curmudgeon and Victor Sperandeo without providing the URL of the original posted article(s).