La crise des subprimes de l'IA : pourquoi votre mairie paie l'addition

If you want to understand the next financial crisis, stop looking at Nvidia’s P/E ratio and start looking at the water bill in Covington, Georgia.

While the financial press obsesses over whether AI is a bubble, a far more dangerous game is being played in the municipal bond market. The “AI Boom” is no longer being funded just by venture capital equity; it is largely being constructed on a foundation of opaque Private Credit and aggressive municipal tax incentives that transfer the risk of failure from Silicon Valley balance sheets to local school boards.

This week, the Bank for International Settlements (BIS) released specific warnings in Bulletin No. 120, titled “Financing the AI boom: from cash flows to debt”. The report highlights a structural shift that has gone largely unnoticed: the explosive growth of private credit lending to AI infrastructure, which has surged from near zero to nearly 8% of all private credit volumes. A trajectory set to hit $600 billion by 2030.

The warning is clear: The “risk” isn’t staying in the private market. Through complex “PILOT” (Payment in Lieu of Taxes) structures and milestone-based infrastructure deals, the volatility of the AI sector is being securitized by the stability of American townships.

The Covington Mechanism: How to Offload Risk

To understand the macro threat, you have to look at the micro deal. Consider the recent agreement between Amazon Data Services and the City of Covington, Georgia.

On paper, it looks like a win: Amazon is building a massive data center campus. But the structure of the deal reveals who is actually holding the bag.

Amazon agreed to pay $100 million upfront to build water-reuse infrastructure. In exchange, they received a Payment in Lieu of Taxes (PILOT) agreement that abates their property taxes for 16 years. Instead of paying the estimated $530 million in standard property taxes, Amazon will pay just $301 million in PILOT fees.

Here is the “Subprime” math:

• Amazon Savings: ~$229 million in taxes.
• The Cost: A $100 million check for water pipes that they need anyway to cool their servers.
• The Loser: The local Board of Education (BOE). Because PILOT payments are distributed differently than property taxes, the schools, which rely on property tax revenue, are projected to forgo roughly $237 million over the life of the deal.

The city gets a new water plant. The tech giant gets a 40% discount on its operating costs. The school board gets a budget crisis in 2028 when the enrollment numbers don’t match the revenue shortfall.

The “Escrow” Trap

The new trend in 2026 municipal deals is the “milestone” or escrow payment. Cities, burned by previous vaporware projects, now demand cash upfront. Amazon’s $100 million water payment is the golden example.

It sounds prudent. But strictly speaking, it transforms the municipality into a project finance partner. The city is no longer a neutral regulator; it is a stakeholder in the success of the data center. If the “AI Bubble” bursts and Amazon halts construction or provisions the servers at 20% capacity, the city is left with oversized infrastructure designed for a tenant that isn’t consuming it, and a tax base that never materialized.

The Private Credit Shadow

Why are tech giants, sitting on record cash piles, using complex municipal deals and private credit?

Financial Opacity.

BIS Bulletin 120 identifies a critical pivot: AI companies are moving from “cash flow” funding to “debt” funding. But they aren’t using transparent corporate bonds. They are using Private Credit (loans issued by non-bank funds like Blackstone, Apollo, or Ares) that do not require the same public disclosure as bank loans or public bonds.

1. The Shell Game: Tech Firm X creates a “Special Purpose Vehicle” (SPV) to build the data center.
2. The Funding: The SPV borrows billions from a Private Credit fund.
3. The Collateral: The loan is secured by the future lease payments of the GPUs, not the parent company’s assets.
4. The Result: If the data center fails, the SPV goes bankrupt. The parent company walks away. The Private Credit fund takes a haircut.

But the Municipality? They are left with a half-built concrete shell and a 16-year hole in their budget.

$$\text{Systemic Risk} = \text{Opaque Debt} \times \text{Public Liability}$$

The BIS estimates that private credit exposure to AI could reach $600 billion by 2030. This is “Shadow Banking” 2.0. In 2008, the toxic asset was the subprime mortgage hidden in a CDO. In 2026, the toxic asset is the AI Lease Agreement hidden in a Private Credit bundle.

The Unitranche Illusion: Why Covenants Disappeared

To understand why this is accelerating, you must look at the deal structures themselves. In traditional bank lending, a construction loan for a $1 billion facility would come with strict “covenants” (rules requiring the borrower to maintain certain cash ratios or occupancy levels). If the tenant leaves, the bank can seize the asset immediately.

Private Credit deals in the AI sector are increasingly using “Covenant-Lite” Unitranche structures.

A Unitranche loan blends senior debt (low risk) and junior debt (high risk) into a single instrument with a blended interest rate, often floating at SOFR + 550 basis points. Because these funds are desperate to deploy capital (the “Dry Powder” problem), they are waiving traditional protections.

• No “Key Man” Clauses: If the star AI researchers leave the tenant startup, the loan remains valid.
• PIK Toggles: Borrowers can choose to “Pay In Kind” (pay interest with more debt) rather than cash, masking default risks for years.
• EBITDA Add-Backs: Borrowers are allowed to add “projected” future AI efficiencies to their current earnings, artificially lowering their debt ratios.

This creates a “Zombie Facility” risk. A data center SPV can be technically insolvent (generating zero cash) but appear healthy on the private credit fund’s books because it is paying its interest with new PIK debt. The municipality sees a functioning building and assumes the tax revenue is safe. In reality, the tenant is a ghost, and the landlord is a shell company solely kept alive by accounting tricks.

When the PIK toggle expires (usually after 18-24 months), the default is sudden and total. The private credit fund seizes the keys, but they are financial engineers, not facility managers. They will look to liquidate the asset for pennies on the dollar. The city, which built roads, sewers, and schools anticipating 16 years of PILOT payments, is subordinated to the lenders. They get nothing.

The “Stranded Asset” Scenario

The danger isn’t that AI is “fake.” The danger is that it is physical and specific.

Unlike software, an AI data center is a hyper-specialized industrial facility. It requires:

• Liquid cooling loops that are useless for standard warehousing.
• Power density (100kW+ per rack) that is overkill for standard logistics.
• Reinforced flooring that is unnecessary for commercial retail.

If the AI hype cycle cools, or if the “Demand Shortfall” scenario outlined by the BIS occurs, these buildings cannot be easily repurposed. You cannot turn a 500MW liquid-cooled gpu cluster into an Amazon fulfillment center without gutting it. The cost to “retrofit” a failed AI data center back to a standard warehouse often exceeds the cost of new construction.

History provides a clear precedent. In the late 1990s, companies buried billions of dollars of “Dark Fiber” under American cities, betting on internet traffic growth that wouldn’t arrive for another decade. The fiber companies went bankrupt. The cities that subsidized them were left with lawsuits.

In 2026, the “Dark Fiber” is the “Dark Data Center”—a massive, tax-abated fortress that consumes water and power but generates no profit. See our analysis on The Aluminum Moment for more on the physical constraints.

What Happens Next?

The “AI Capex Bubble” is currently being subsidized by two groups who can least afford the risk:

1. Pension Funds: Who are the biggest investors in Private Credit funds? Public employee pension systems chasing 10% returns.
2. Local Taxpayers: Who are subsidizing the construction costs? Residents of towns like Covington, GA.

If the music stops, Nvidia and Microsoft will be fine. They have the cash reserves to weather a winter. But the Newton County Board of Education cannot print money. The pension fund retirees cannot renegotiate their benefits.

The “Subprime AI Crisis” won’t look like a stock market crash. It will look like a pothole that doesn’t get fixed, a school bus route that gets cancelled, and a property tax bill that quietly goes up 15% to cover the difference.

The tech industry is effectively shorting the American municipality. And right now, City Hall is taking the other side of the trade.