Borrowing costs across the whole AI financing chain still sit close together, even though the risk on each link in it is not remotely the same.
Equity investors have spent the cycle separating winners from losers, and their returns show it. Credit has not. Lotfi Karoui writes that lenders to a hyperscaler and lenders to a company that rents out graphics processing units are being paid almost the same premium for taking on very different exposures — and that the reason to care is what happens when the supply of that debt grows.
"For credit investors, the AI question is not simply who wins the race, but who is left holding the risk."
Karoui is a managing director and multi-asset credit strategist at PIMCO, where he also co-heads client solutions and analytics, and this is the firm's own published view for the clients who buy the debt he is describing.
The full episode is covered here so you can skip it.
Here are the 7 arguments that matter.
👤 Speaker: Lotfi Karoui, Managing Director and Multi-Asset Credit Strategist at PIMCO, and Co-Head of Client Solutions and Analytics
📰 Published: 14 September 2026 on the PIMCO Pod
🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 11 min
Key Takeaways
The credit market is treating the whole AI buildout as a single trade, while the equity market has already stopped doing that
AI-related debt has underperformed the broader investment grade and high yield indices so far this quarter
Lenders take the downside of the buildout and almost none of the upside, which makes avoiding losers more important than picking winners
The financing chain has four rungs, and utilization and resale-value risk shifts onto the lender at every step down it
At the bottom rung the lender is exposed directly to how much the chips are used, what they rent for, and who rents them
A bust could be good for hyperscaler bondholders, because capex falls and free cash flow recovers
The condition is speed: sunk power costs, signed leases and contracted equipment orders can keep spending going anyway
The lease protections on a data center leased to a neocloud are weakest in exactly the scenario they exist for
A bankruptcy court can cap a landlord's claim well below what the rest of the lease is worth
Even the good scenario is mostly an equity outcome — a longer capex cycle means more bond supply, not tighter spreads
1. One Trade, Not Many
Karoui opens on the difference between the two questions the same buildout poses. "For equity investors, the question is relatively simple: Who wins the AI race? For credit investors, it is more nuanced: Do spreads adequately compensate for the broad set of risks embedded in financing the buildout?"
His answer to whether the market treats the buildout as one trade is that so far it does. "Spread dispersion – differences in borrowing costs among issuers – across the financing chain remains limited despite sharp differences in underlying risk," he wrote
He sets that directly against the equity market, "where performance has become increasingly differentiated"
The one place the difference does show up is in the level rather than the spread between issuers: "In both investment grade (IG) and high yield (HY) credit, AI-related debt has underperformed broader indices quarter-to-date, according to Bloomberg index data"
The note's own headline framing is that this is temporary — issuers, structures and exposures are about to get more varied, and pricing should follow
2. Financing Without Upside
The structural point underneath the whole note is that a lender's payoff is capped by contract while the risks are not.
"For debt investors, the proposition is fundamentally asymmetric: They finance the AI buildout without directly participating in much of its economic upside," Karoui wrote
What a bondholder can earn is fixed in advance: "Returns are largely contractual, driven by coupon, principal, and, at most, some spread compression."
The risks are not. He lists five that are not capped: leverage, execution, utilization, technological obsolescence and refinancing
The practical consequence he draws is a change of job description: "Credit investors finance the AI buildout but capture relatively little of its upside, making risk mitigation more important than winner selection."
Even the optimistic case does not remove the question, on his reading: "Even if adoption, monetization, and capex reinforce one another, debt investors still need to ask whether spreads adequately compensate them for obsolescence, re-contracting, and refinancing risk."
3. Why Dispersion Arrives
Karoui's reason for expecting borrowing costs to separate is a supply argument rather than a credit-quality one, and it rests on the gap between what the buildout costs and what its sponsors can fund from cash.
The funding gap is what keeps the new debt coming: "Given that the current funding gap for AI capex will likely persist, debt issuance will likely continue to expand, bringing a wider variety of issuers, structures, and risk exposures to market."
"That variety should create more room for differentiation," he wrote — more kinds of borrower means more chances for the market to price them apart
He points to two figures in the published note showing AI capital spending still absorbing enormous amounts of capital while the funding gap persists
The second thing time brings is information: investors should gain greater clarity on where the economic value of the buildout ultimately accrues
4. The Four Rungs
The core of the note is a taxonomy of who is exposed to what, arranged from safest to riskiest. Karoui's framing of it is the sharpest line in the piece: "For credit investors, the AI question is not simply who wins the race, but who is left holding the risk."
Hyperscaler corporate debt is the top rung, and the buildout barely reaches the creditor. "Diversified businesses, strong cash flows, and large balance sheets largely insulate creditors from the economics of any single data center, model, or GPU cluster," he wrote
Where the risk sits instead: "Utilization and technology risk remain inside the enterprise rather than passing to lenders"
Debt on a data center leased to a hyperscaler moves the analysis to the lease. The site typically serves one internal customer, and so, he wrote, "Creditors primarily underwrite the tenant's balance sheet and lease, not utilization at a specific site."
Construction risk depends on completion obligations, and lease protection depends on guarantees, termination rights and delay remedies
The tenant absorbs obsolescence: "Once operational, utilization and technology risk largely remain with the tenant, which can replace obsolete GPUs while continuing to use the same shell and power infrastructure."
What is left for the lender is concentration in one tenant and the risk of having to re-let the building when the lease ends
Debt on a data center leased to a neocloud looks identical and behaves differently: "Creditors indirectly underwrite the neocloud's ability to resell compute capacity profitably, which depends on utilization, pricing, customer retention, and access to capital."
The mismatch he names is between three different clocks: "Leases can outlast GPUs, GPUs can outlast customer contracts, and debt can outlast both."
That makes execution and refinancing central, because capacity has to be re-let at good prices while the hardware is being replaced
Neocloud debt is the bottom rung and holds everything at once: "This is the furthest point on the spectrum: direct exposure to utilization, pricing, customer concentration, technology-refresh, and refinancing risk, with the fewest contractual protections and the least asset coverage."
5. Where the Risk Migrates
Karoui is careful that the four categories are a map rather than a set of walls, and that the links between them cut in the lender's disfavor.
The categories bleed into one another through vendor financing, hyperscaler equity stakes in neoclouds and residual-value backstops
The caution he adds is that "a hyperscaler guarantee can be softer than it appears"
The overall movement is one-way: "Still, the direction of travel is clear: Moving from hyperscaler debt toward neocloud debt, utilization and residual-value risk migrate steadily from the hyperscaler balance sheet to the creditor."
His conclusion from that is a preference for the top rung in both extreme outcomes — he writes that hyperscaler credits may offer the cleanest balance of risk and return in the best case and the worst case alike
In the good case the spread cannot tighten much, because success prolongs the capex cycle and keeps issuance high, but scale, cash flow and strategic control support the credit
In the bad case, balance-sheet flexibility and fast spending cuts protect the lender — "provided that discipline arrives quickly enough"
6. If the AI Trade Busts
The bust scenario produces the note's most counterintuitive claim, which is that a failure of the investment case can be good news for the people who lent against it.
"For hyperscaler creditors, the first-order effect is disappointing returns on AI investment. Paradoxically, the second-order effect could be credit-positive: Capex falls, free cash flow recovers, and the issuance pipeline shrinks," Karoui wrote
"The same shock that undermines the AI investment case could improve hyperscaler credit metrics" — but only if the spending actually stops
What could stop it stopping: competitive dynamics, sunk power costs, lease commitments and contracted equipment purchases
He separates better credit ratios from better conditions for bond prices: "Hyperscaler capex is increasingly debt-funded rather than cash-funded, so a pullback that strengthens leverage and coverage ratios could still coincide with elevated net issuance if near-term maturities need refinancing."
Data centers leased to hyperscalers should hold up, provided the contractual protections hold
The hard case is a data center leased to a neocloud, and the question is whether the tenant can keep paying a 10- to 15-year lease if utilization and compute economics deteriorate. If it cannot, and if a bankruptcy caps the landlord's claim well below the remaining contract value, recovery depends on re-letting the building and on what its power access is worth in a market that probably has too much capacity
"In other words, contractual protection is weakest exactly when it is needed most," he wrote
"Neocloud debt would absorb the largest share of the downside, as falling utilization weakens profitability while faster obsolescence erodes GPU asset values"
7. If Adoption Accelerates
The upside case reverses the order of who benefits, and Karoui is explicit that it is not the mirror image of the bust.
The bottom rung gains the most: "Neocloud credits would see the greatest cyclical improvement: Higher utilization supports pricing and cash flow, leverage declines passively, and refinancing gets easier."
Data centers leased to neoclouds improve with them: "Neocloud-tenanted data centers benefit too, as leases that once looked like leveraged bets on future demand begin to resemble durable contracted cash flows."
The limit is that a good outcome does not change the instrument. "Strong demand mitigates utilization risk; it does not eliminate obsolescence, re-contracting, or refinancing risk," he wrote
The specific ways it can still go wrong: chips can become obsolete before leases mature, customer contracts can roll before the lease or the debt, capacity can be re-let on worse terms, and the debt still has to be refinanced
Data centers leased to hyperscalers gain least, because their contracts already priced in the good outcome
Hyperscaler bonds could actually lag in the upside case: success validates more investment, extends the capex cycle and keeps issuance high
"Shareholders capture the economic upside; bondholders mostly face another round of supply"
Bonus Insights
Karoui's headline for the whole note is a timing claim rather than a rating one — one trade now, many trades later — and the thing that turns one into many is the arrival of more issuers, not a change in the underlying technology
The note is credited to Karoui with two contributors, Michael Puempel and Gabriel Cazaubieilh
The word "paradoxically" is the author's own, applied to the idea that hyperscaler credit could improve in a downturn — he flags the claim as counterintuitive rather than leaving a reader to notice
Nothing in the piece names an individual company. The entire argument is built on four categories of borrower, which is itself the point: the question is structural position, not corporate identity
Karoui's bottom line is that the AI buildout is being financed as if it were one credit and is in fact four, that the lender's share of the outcome is capped in the good case and open-ended in the bad one, and that the widening supply of AI debt is what will force the market to start pricing the difference.
Products, Companies & Tools Mentioned
PIMCO (Publisher of The Credit Market Lens and Karoui's firm; the note is its own published view for clients)
Bloomberg (Source of the index data behind the claim that AI-related debt has underperformed both the investment grade and high yield benchmarks this quarter)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: