In the Gaps
Investment-grade AI credit is one leveraged, correlated bet on eight companies — own it only at a spread that pays you to read the documents and watch the flows.
Polar opposites, now counterparties
The most cautious pool of capital in finance now funds the most risk-tolerant: record annuity inflows need investment-grade paper at the exact moment venture-scale AI capex needs funding.
Insurance — built to avoid the rare loss
Venture — built to be wrong most of the time
100+ digital-infrastructure financings closed in the past twelve months — roughly half private, ~50 public deals — issued across seven debt markets (corporate bonds, project/construction, 144A/private placement, GPU/ABF, high-yield, chip SPVs, ABS/CMBS), each engineered to earn an IG rating.
All the risk converges on eight names
- Recurrence. Oracle’s every lease is a bet on OpenAI’s $300B compute commitment; Meta risk is available in many formats.
- Correlation. Nvidia has reportedly committed up to $100B to OpenAI, which is buying $300B of compute from Oracle on Nvidia chips — a closed triangle. Microsoft is OpenAI’s largest investor and reportedly Anthropic’s landlord; Amazon holds an Anthropic stake while supplying OpenAI.
- Coincidence. All eight — Meta, Oracle, Microsoft, Amazon, Google, Nvidia, plus look-through OpenAI and Anthropic — are issuing into every IG market simultaneously. CoreWeave borrows via a GPU facility, two term loans and ~$5B of high-yield/convertibles while others sell notes against its leases; Nvidia appears on the map four ways at once.
26 mapped financings, LTM. Share of the book touching each name — columns overlap and sum past 100%, which is the concentration.
Same transaction, opposite labels
The loop under stress: if the backstop holds, the corporate absorbs the loss and downgrade pressure rises; if any “out” triggers (performance, delay, force majeure, law, weather, cost), the corporate walks and the lender owns asset risk. Conditional backstop ≠ unsecured corporate debt.
Data-center leases signed but not yet commenced (Amazon, Meta, Alphabet, Microsoft, Oracle) already exceed those five firms’ entire adjusted debt — total commitments approach $1T.
Deconsolidated off Meta’s balance sheet and financed with single-A paper — but wrapped in a residual-value guarantee, so it still trades like Meta recourse. Legally equity; economically recourse.
If enthusiasm cools, what exactly are we holding?
Most of the spend capitalizes the asset with the lowest recovery value — the chips.
- Rent tells the story. An H100 that rented at $7–10 per GPU-hour at launch was down to $2–4 within two years (Hashrate Index); architectures ship roughly annually (Hopper → Blackwell → Rubin) — a chip financed today is a generation behind before its loan is half amortized.
- Nobody really knows residual value. The market assumes three-year residuals anywhere from 10% to 60%, and the vendors control the obsolescence clock — CoreWeave renewing expired H100 leases at 95% of the original rate may be scarcity, not a floor.
- Vendor backstops are circular — the same vendor-financing loop that turned the 2000 telecom fiber glut into a risk-correlation story.
Value retained after the vendor’s move: IBM $233,900 → $71,650 — lessor Itel recovered ~19% of a $310M residual claim. A380 ~$197M → ~$45M part-out. Solar $3.25/W → $0.72/W. H100: 35–51% retained after ~3 years — still declining.
The collateral (chips), the tenant’s cash flow (rents), the credit support (vendor volume backstops) and the funding (insurance capital) are not four independent risks — they are one bet on the AI capex cycle.
Read the documents, not the deck
| As marketed | As documented |
|---|---|
| “Take-or-pay” offtake from a brand-name hyperscaler | A services agreement with uptime and force-majeure outs — not hell-or-high-water |
| Investment-grade GPU financing at +200 bps | Delivery date certain; 99.9% uptime; chips “as installed” — the site did not exist at pricing |
| Hyperscaler protection behind the cash flow | A step-in right, not an obligation |
Borrower: a former bitcoin miner turned “AI cloud provider.” Miss any “what must go right” item — build, operator track record, grid power, uptime — and the end customer doesn’t have to pay.
Insurers will not touch this risk as a property liability — yet they buy it as an asset when it arrives rated investment grade.
Done correctly, a massive opportunity. Done incorrectly, a severe hangover — and the difference between the two is documentation-level diligence.
Watch the flows, act early
Investment-grade AI credit is one leveraged, correlated bet on eight companies — own it only at a spread that pays you to read the documents and watch the flows.
- Record annuity sales keep pushing hundreds of billions into insurance balance sheets that must buy IG paper — even as insurance ROEs compress.
- If these inflows slow, the marginal buyer of the entire structure disappears — and when flows reverse, liquidity is what matters most.
- Rates 3.11% → 6.66% (+355 bp) cut turnover −36% (4.4% → 2.8% of stock); half of outstanding mortgages still carry <4% (20% <3%, FHFA 1Q26); each 1pp rate gap cuts sale probability 18%; lock-in prevented >1.7M sales (2Q22–2Q24).
- Feb-26 dip below 6%: 5.4M borrowers gained ≥75 bp refi incentive; 1Q26 first-lien refis $242B, 2× prior year (avg −97 bp / −$257 a month). The first-order effect of cuts is household cash flow, not inventory — the refi market thaws before the resale market heats up.
Reduce risk before the stress arrives, not during it.
Underwrite downside scenarios, diversification, and alignment of interests
Cycles are easy to see; timing is hard.
Buy assets and cash flow · risk is exponential but often priced as linear · watch the flows
The worst stresses sit where no rating criteria required anyone to look.
Stress-test the risks outside the models — that is the case for reading the documents