Opinion

Banks Rush to Dump AI Data Center Loans: Who's Left Holding the Bag?

Hackers News threads have been buzzing about a wave of AI data center loans being sold off in the secondary market at deep discounts before the ink is even dry. Behind the chatter is a potentially troubling signal of asset bubble dynamics converging on AI infrastructure.

Font size

A thread on Hacker News yesterday climbed to 75 points with 35 comments. The headline was blunt enough: "Banks are now trying to offload the AI data center debt they just underwrote".

Short version: a bank underwrote a loan for an AI data center project, and before the ink dried on the paperwork, tried to flip the debt in the secondary market.

This isn't an isolated case. Reuters and Bloomberg have been tracking similar moves. Platforms that monitor loan syndication flow, such as LBJ Assistant, have already carved out a dedicated section to track secondary discounts on AI data center projects.

Loans Signed and Already on the Block?

The event itself boils down to one sentence: a bank originated a project loan, then wanted to unload it cheaply in the secondary market. Three layers worth unpacking:

  • The project: an AI data center build-out, usually with capex in the multi-billion-dollar range.
  • The behavior: the bank kicks off a secondary sale almost immediately after underwriting.
  • The result: buyers only step in at a discount.

Why is this interesting? Secondary loan sales are routine. What's unusual here is the speed of the exit and the size of the haircut. Those two signals together usually mean the underwriter doesn't want to hold to maturity.

High-frequency keywords in the HN comments: underwriting party, so much capital, capex, distributing the risk. The short version: the underwriter is spreading the risk because even they think the duration is too long.

The Financial Mechanics: Syndication Isn't New

Loan syndication is a 50-year-old move for banks. A $3 billion project loan is too big for one bank, so a few form a club, slice it up, or package it into a CLO and sell it to insurers and pension funds. When the underlying asset blows up, everyone takes a piece of the loss. Textbook stuff.

The issue is timing and price.

Normal cadence: underwrite → hold for a few months to observe performance → sell depending on market conditions. That gives lenders time for credit research and lets the loan season.

Current cadence: underwrite → package and sell almost immediately. The underwriter seems to be scrambling to dump the bag.

Price signal: syndicated loans typically trade at 98–99 cents on the dollar in the secondary market. When you start seeing frequent 90-cent or lower quotes, the underwriter is telling you something with that price — "I'm not sleeping well holding this."

The original story used the phrase "stands out." Not because bank loan resales are unprecedented, but because the AI data center paper is moving unusually fast.

Why AI Data Centers Specifically

Time to run the numbers.

A 100MW hyperscale data center:

  • Land + construction: $200–500M
  • Servers + networking gear: $1–2B
  • Power + cooling + redundancy: $500M–1B
  • Total capex: $2–3B

On the revenue side: a 10-year MSA (master services agreement) with one hyperscale customer can generate $3–6B in revenue. Sounds great. Three problems though:

  1. Depreciation speed. GPUs depreciate over 3–5 years. At year four, you're holding a book entry that pays the bills only on paper.
  2. Power risk. A 100MW facility spends $100–150M per year on electricity — that's 5–7% of total capex. A 5% swing in power prices moves equipment depreciation by a few percentage points.
  3. Uncertain useful life. A 10-year MSA sounds stable, but AI workloads evolve so fast that betting on GPUs still being relevant in 2030 is a stretch.

On the bank's balance sheet: a $2–3B loan, 10-year tenor, project finance structure (repaid from cash flow, not a parent guarantee). For the bank, that's long duration + uncertain cash flow + concentrated depreciation risk = no desire to hold.

So the decision to flip the paper right after underwriting is economically rational.

Who's Buying

The secondary market has a few buyer types:

  • CLOs (Collateralized Loan Obligations). Pool leveraged loans into a security and sell to yield-hungry institutions. These are the main bidders.
  • Private credit. Blackstone, Apollo, KKR, the usual names. Returns have been juicy at 15–20% in recent years.
  • Insurance capital. Matches the duration, moderate return target.
  • Pension and sovereign funds. Parked in the infrastructure asset class.

These buyers have risk appetites that can absorb long-duration, low-liquidity project loans. But watch for a specific combination: "CLOs and insurers both pulling back."

Widening bid pricing in the loan market is a signal: buyers are demanding bigger discounts. They're not uninterested — they just require much better terms to engage.

When long-duration capital starts demanding extra risk compensation, the asset's price has decoupled from fundamentals.

How This Differs from the 2000 Dot-Com Bubble

2000 dot-com2024–2025 AI infra
Risk vehicleEquityDebt
Capital providersRetail + VC + IPOBanks + CLOs + private credit
Leverage signalStock price + IPO valuationLoan secondary discount + resale speed
Failure pathStock crash + delistingDefault + restructuring
First wave exposedRetail + VCCLOs + private credit

Back then, it was an equity bubble — Cisco at $410B market cap, pets.com tripling on its first day. Risk sat in the stock market and cleared through IPOs and share-price collapses.

Now it's a debt bubble — risk is moving off bank balance sheets through loan resales and CLO packaging, landing with yield-seeking institutions.

One more 2024 data point: in the high-yield bond market, AI-related issuance (solar, semis, infrastructure) has been visibly climbing. That's the same kind of signal as the rising IPO share during the dot-com era — a "hot asset class" reaching saturation.

Editorial Take

Why does this matter? Because we're watching a transfer signal in the AI infrastructure financing chain — not "will it blow up" but "when it does, what's the contagion path."

My read diverges from the mainstream narrative:

The buyers won't blow up immediately. CLOs have structural protection: asset isolation plus senior/sub tranching. Private credit doesn't even report performance. No default wave is visible in the next 12–18 months. But 2027–2028 is the key observation window. Four reasons:

  1. Loans underwritten in 2024–2025 enter their refi window.
  2. Second- and third-gen GPUs ship in volume, rewriting the residual value of older hardware.
  3. If power prices and cooling demand overshoot estimates, opex pressure transmits through the model.
  4. If AI workload growth undershoots expectations, MSA cash flows may come in below projections.

What makes this structurally different from past episodes?

Versus dot-com: back then it was retail + IPO, so the unwind was fast. This time it's institutions + CLOs + private credit, so the unwind is slower but more continuous.

Versus 2008: back then it was subprime + residential + retail. This time it's investment-grade + industrial + AI. The underlying assets may be healthier, but duration and concentration risk are higher — large single tickets, long tenors, and the collateral is increasingly the same kind of bet.

The real uncertainty: the trajectory of AI workloads. If inference grows faster than training, data center utilization will tighten well beyond current forecasts. If the reverse happens — generative AI growth slows and hardware iteration decelerates — the cash flow models on these loans need to be rewritten.

My call: no signs of imminent default, but the froth is visible. Watch the 2027 refi queue more than any macro indicator — that tells you what's actually happening.


Sources:

Source: https://startupfortune.com/banks-are-now-trying-to-offload-the-ai-data-center-debt-they-just-underwrote/