Private Credit Is Financing the AI Datacenter Boom. Here's the Concentration Risk Nobody's Pricing.

TL;DR: Private credit funds are underwriting the AI datacenter boom by lending against GPUs and hyperscaler contracts, and the industry is pitching it as diversified, asset-backed credit. It is not. It is concentrated exposure to a handful of borrowers and tenants, secured by hardware that Nvidia is racing to make obsolete, a risk Moody's flagged in its "Power without delivery" report. If you allocate to a fund with this exposure without tracing the collateral chain, you are not buying a diversified credit sleeve. You are buying a bet on Nvidia's product roadmap and one or two corporate counterparties.

How Private Credit Got Into the GPU Business

I want to start with the mechanics, because the marketing language obscures them on purpose.

A "neocloud" is a company that buys large quantities of GPUs, the specialized chips that train and run AI models, and rents out that compute capacity to AI labs and enterprises. CoreWeave is the largest and most visible example. Crusoe and Lambda run a similar playbook: buy Nvidia hardware, build or lease datacenter space, sign customers to multi-year compute contracts, repeat.

Neoclouds cannot fund this with equity alone. Fitting out a datacenter with the latest chips can now cost more than constructing the building around it, according to Optio Investment Partners, a Swedish private credit firm building a lending business around exactly this gap, as reported by Debtwire. Banks are reluctant lenders here because regulatory risk-weights punish them for holding fast-depreciating hardware collateral, and that reluctance created an opening for private credit.

CoreWeave's own history shows the structure evolving fast. In 2023 the company raised $2.3 billion in debt collateralized directly by Nvidia H100 chips, led by Magnetar Capital and Blackstone, according to Reuters. By August 2026, per the company's own disclosures, it had secured more than $30 billion in debt and equity capital for the year, including a single $3.1 billion facility that drew $19 billion in investor demand. Hyperscalers run a parallel version of the same trade at larger scale still. Meta's $27 billion-plus special purpose vehicle for its Hyperion datacenter campus in Louisiana, with Blue Owl-led investors reportedly holding an 80% stake, set the template, with Pimco leading roughly $26 billion of the debt tranche after beating out Apollo and KKR, according to Bloomberg's account of the bidding war. Morgan Stanley estimates private credit could supply more than half of the roughly $1.5 trillion needed for global datacenter buildout through 2028, an $800 billion opportunity, according to Reuters. Here is the pitch, roughly as it's presented to allocators: asset-backed private credit, senior in the capital structure, collateralized by hardware with real resale value, backed by revenue from an investment-grade tenant. Every phrase is technically true in specific instances. Strung together as a category description, it is misleading.

The GPU Depreciation Curve Nobody Wants to Underwrite Honestly

A depreciation curve is the rate at which an asset's accounting value, and its real resale value, declines over its useful life. For a building, that curve is gentle: decades, sometimes generations. For a GPU, the debate over that curve is a live fight between short sellers and the companies buying the chips, and it matters enormously to anyone lending against them.

Nvidia now releases a new flagship GPU architecture roughly every year, up from a two-year cadence previously. Google, Oracle, and Microsoft depreciate AI server equipment over as long as six years in their financial filings. Amazon cut the useful life on a subset of its servers from six years to five after finding "an increased pace of technology development," according to CNBC's reporting on the depreciation debate. CoreWeave has used a six-year depreciation schedule since 2023.

Michael Burry, the investor who shorted subprime mortgages before the 2008 crisis and disclosed a short position against Nvidia in late 2025, argues the real number is two to three years, not five or six. He estimates the gap could understate industry-wide depreciation expense by $176 billion between 2026 and 2028, inflating Meta's and Oracle's reported profits by roughly 21% and 27% respectively by 2028, per his analysis as covered by CNBC. Even Nvidia CEO Jensen Huang has conceded the resale-value point: when Blackwell chips started shipping in volume, he said, "you couldn't give Hoppers away."

I am not going to claim Burry has the exact number right. He might not. The point for a credit investor is narrower and less debatable: nobody has a reliable track record for GPU longevity, because the current AI compute buildout is only about three years old. Haim Zaltzman, a lawyer at Latham and Watkins who works on GPU financings, put it bluntly to CNBC: "Is it three years, is it five, or is it seven? It's a huge difference in terms of how successful it is for financing purposes." When collateral value depends on a number nobody can verify, "asset-backed" is doing a lot of marketing work relative to what it actually secures. CoreWeave's own facilities now stretch to maturities of five and a half years and beyond, and if the chips backing them are functionally obsolete before the loan matures, the lender's collateral cushion is thinner than the headline loan-to-value ratio implies.

Single-Tenant Risk Wearing an Infrastructure Costume

The second leg of the risk is counterparty concentration, and it may matter more than depreciation, because it can break a deal even if the hardware holds its value fine.

Most large AI infrastructure financings are underwritten against one thing above all: the strength and duration of the tenant's contract. CoreWeave's February 2026 $8.5 billion facility was backed specifically by a Meta contract worth up to $14.2 billion plus a separate agreement worth more than $5 billion, according to Bloomberg. That is not diversification. That is a facility whose repayment capacity rests largely on Meta's continued willingness to pay for compute it contracted for years earlier, in a market where AI model architectures and compute needs keep shifting.

CoreWeave's August 2026 $2.6 billion facility carries an approximately five-year maturity while the underlying customer contracts it finances average only about three years, per the company's own disclosure. Lenders are explicitly underwriting renewal risk, betting the customer re-ups or CoreWeave re-leases the capacity, because the contract in hand doesn't cover the full loan term. That turns an "asset-backed, contracted cash flow" deal into an equity-like wager dressed up as senior secured debt. Most neocloud offtake agreements with AI developers run just 12 to 18 months, and rating agencies apply steep haircuts to any revenue projected after that term ends, according to Optio's David Lindström, as reported by Debtwire. He compared the current state of GPU financing to "where oil was before we had a futures market," meaning there is no way yet to hedge against a customer walking. Layer the private credit fund on top and you get a four-link chain: fund investor, private credit fund, neocloud borrower, hyperscaler customer. An accredited investor in a BDC with datacenter exposure is two steps removed from a contract they have never read, signed by a customer they cannot evaluate.

Fund managers market this exposure as diversification away from sponsor-backed middle-market corporate loans, and on paper that's fair: different sector, different collateral, often investment-grade-like ratings. But diversification within a portfolio is not diversification within a loan. A fund can hold ten different datacenter loans and still be exposed to the same handful of things: Nvidia's product cadence, the same short list of hyperscaler counterparties, and the same assumption that AI compute demand keeps outrunning supply. That's concentration wearing a diversification costume. The law firm Bird and Bird flagged this directly in its analysis of GPU-based financing structures, warning that simultaneous collateral liquidations could trigger a market-moving oversupply that undermines the wider collateral pool.

Risk Factor	How It's Marketed	What It Actually Is
Collateral	Asset-backed, hardware-secured lending	Collateral with an unresolved two-to-six-year useful-life debate and no long track record
Tenant base	Investment-grade hyperscaler contracts	Revenue concentrated in one or two named counterparties per deal
Sector diversification	New, uncorrelated asset class in a credit portfolio	Correlated exposure to Nvidia's roadmap and a handful of AI buyers across many loans
Duration match	Long-term, contracted cash flow	Some loan maturities now exceed contract length, betting on renewal

Is the Steelman Case Right, and This Is Actually Fine?

I should make the other side of this case as strongly as I can, because it is not a strawman.

Counterparty quality here is genuinely unusual for private credit. Most of this capital flows toward Meta, Microsoft, Amazon, Google, and Oracle, five of the strongest corporate balance sheets on earth. Ares Capital's review of its $29.5 billion BDC portfolio found 85% of its AI-exposed software holdings carry low disruption risk, with only 1% at high risk, disclosed alongside quarterly results and reported by Bloomberg. Structuring has improved too: limited guarantees, parent-level pledges, and collateral segregated in special-purpose subsidiaries protect lenders, and construction-only lenders can sidestep obsolescence risk almost entirely by taking short, three-to-five-year exposure instead of betting on decade-long operating economics, as wealth manager iCapital noted in its market commentary on the sector. Default rates in AI-linked private credit remain low today.

I take that case seriously. I don't think it's wrong that this generation of deals is better protected than the summary above suggests. I think it's incomplete. The steelman assumes today's contract terms are a reliable guide to year four or five of a loan. Neither assumption survived Optio's own observation that lenders are stretching maturities past contract length to win deals, or Moody's warning that many facilities are financed against future compute demand rather than established operational utilization.

Why the Blue Owl Episode Should Worry Datacenter Lenders Too

In February 2026, Blue Owl Capital, a direct lender where more than 70% of its loan book sits in software, sold $1.4 billion of loans to institutional buyers at 99.7 cents on the dollar to demonstrate confidence in its book. The sale instead triggered a selloff in Blue Owl and other alternative asset manager shares, partly because Blue Owl simultaneously replaced voluntary quarterly redemptions with mandatory capital distributions, according to CNBC's reporting. Blue Owl shares fell more than 50% over the following year, and economist Mohamed El-Erian publicly asked whether Blue Owl was "a canary in the coal mine."

That episode was about software lending, not datacenters directly, but Blue Owl is also the lead equity investor in Meta's Hyperion campus SPV. Ares, Blackstone, and Apollo, all active in AI infrastructure financing, moved within weeks to cap redemptions on other funds after withdrawal requests surged, as reported by CNBC's coverage of the broader wobble. The lesson is not that datacenter lending and software lending are the same risk. They are not. The lesson is structural: these are illiquid loans held in vehicles that promise investors some liquidity, and when sentiment turns on any AI-adjacent corner of private credit, redemption pressure can force asset sales across a manager's entire platform, including datacenter positions that were fundamentally sound on their own terms.

A Due-Diligence Checklist Before You Allocate

If you are an accredited investor looking at a fund, BDC, or interval vehicle with AI infrastructure or GPU exposure, put these questions to the manager directly and do not accept a marketing deck as the answer.

What is the actual tenant count behind this exposure? If several loans all depend on Meta, Microsoft, or one or two neocloud borrowers, that is not three uncorrelated positions. It is one concentrated bet wearing three legal wrappers.

What depreciation schedule is embedded in the underwriting, and how does it compare to the loan's maturity? If the loan matures well after the GPU collateral is likely to be technologically stale, ask what the recovery assumption is in a default scenario and who verified it.

Does the loan term match, exceed, or fall short of the underlying customer contract term? A loan that outlasts its contracted revenue is a bet on renewal. Ask the manager to disclose that gap deal by deal.

Is this a construction loan or an operating loan? Construction-only exposure avoids most obsolescence risk by design. Long-dated operating exposure does not.

Has the manager restricted withdrawals anywhere else on its platform recently? Given the Ares, Apollo, and Blue Owl redemption caps of 2026, this is no longer hypothetical.

Who takes the first loss if the tenant does not renew, the borrower's equity or the lender? This is the single most important question, and the one marketing materials most reliably obscure.

The Bottom Line I'd Give a Client

I am not telling you to avoid private credit exposure to AI infrastructure entirely. The counterparties involved, Meta, Microsoft, Amazon, Google, and Oracle, are genuinely strong, and some of these structures carry real protections. What I am telling you is that the category is being sold as diversified, asset-backed, infrastructure-like credit when a meaningful share of it is concentrated counterparty risk wrapped around collateral whose useful life is a live, unresolved argument between Nvidia's own customers and a hedge fund manager with a credible track record of being early on this exact kind of accounting question. Treat every datacenter private credit allocation as a corporate credit bet on two or three named companies, not a new flavor of diversification, and price it that way.

Frequently Asked Questions

What is a neocloud, and how is it different from a traditional cloud provider like AWS?

A neocloud is a company that buys large quantities of GPUs and rents out that compute capacity to AI labs and enterprises, typically without the diversified software, storage, and consumer businesses that hyperscalers like Amazon or Microsoft run alongside their cloud units. CoreWeave, Crusoe, and Lambda are commonly cited examples. Because neoclouds depend heavily on GPU rental revenue and a small number of large contracts, their financial profile is far more concentrated than a diversified hyperscaler's.

What does GPU-backed or asset-backed private credit actually mean?

It means a loan is secured by a specific asset, in this case GPUs, the datacenter housing them, or the customer contracts generating revenue from them, rather than solely by the borrower's general creditworthiness. In theory that gives the lender something to seize and sell if the borrower defaults. In practice, the resale value of GPU collateral depends on assumptions about obsolescence that remain unsettled after only about three years of AI-driven demand.

Why does GPU depreciation matter so much for lenders specifically?

Depreciation determines both the accounting profitability of the borrower and the real-world collateral value backing the loan. Hyperscalers depreciate server equipment over five to six years in their financial statements, while short seller Michael Burry argues the real economic life is closer to two to three years. If Burry is closer to correct, loans structured with five-to-six-year maturities may be secured by collateral that has lost most of its resale value well before the debt matures.

Is private credit exposure to AI datacenters likely to cause a systemic financial crisis?

Most analysts describe the risk as significant but not systemic, citing lower leverage in private credit funds and business development companies compared with banks in 2008. The more realistic concern is idiosyncratic: individual funds taking outsized losses if a major tenant fails to renew or collateral depreciates faster than underwritten, combined with liquidity strain if investors rush to withdraw at the same time.

Author Disclosure: Jeff Barnes, MBA has no personal position in any company, fund, or platform named in this article. Angel Investors Network has no current commercial relationship with any party mentioned. AIN provides marketing and education services, not investment advice. Past performance does not guarantee future results. All investments involve risk, including loss of principal.