AI infrastructure had a week of two halves. On Tuesday, Modal raised $750M at a $15.75B valuation with $1B in annualized run rate — the largest AI-infra equity round of the week. The rest of the week belonged to the debt markets: CoreWeave's shares fell ~16% in a single session, its bonds plunged to prices implying ~16% yields, and its credit-default swaps hit record wides near 1,000 basis points. The AI credit boom is not over — but it is being repriced, fast.
What "AI credit" means
The neocloud model — companies like CoreWeave, Nebius, Nscale, and Applied Digital that rent out Nvidia GPU capacity — requires buying the chips before the customers arrive. A GPU data center is billions in capex for hardware that depreciates in years. Neoclouds can't fund that from equity alone, so they finance it with debt: unsecured corporate bonds, secured project-finance debt, GPU-collateralized loans, and sometimes vendor financing. "AI credit" is that whole stack of borrowing — the leveraged bet that GPU demand will stay strong enough, long enough, to service the interest.
It worked while AI enthusiasm was limitless and rates were falling. This week, lenders started asking what happens if it doesn't.
What the bond market is saying
The numbers from mid-October are stark:
- CoreWeave's 9.75% 2031 bonds traded as low as 78 cents on the dollar — a 16.39% yield — and its 2032 notes hit 16.5% yields. Morningstar reports these as fresh lows for data-center bonds generally, triggered partly by developer Cipher Digital's new debt raise fueling AI-overspending anxiety.
- S&P rates CoreWeave B+ — firmly junk. The company carries roughly $50B in debt, raised its 2026 capex forecast to $35–39B in August, and is still forecast to lose ~$2.1B this year on ~$13B in revenue. Its gross leverage ratio sits near 6.3x EBITDA.
- A 16% yield is distress pricing: investors demand a steep return because they doubt refinancing at normal rates will be possible. CDS at ~1,000 bps implies the market sees a serious probability of default within five years.
- Gartner warned that 80% of neoclouds could disappear by 2030, flagging over $3.5T in AI infrastructure spending exposed to impairment.
- The equity market joined in: Oracle fell ~9%, Nebius ~10%, Dell ~11% in the same selloff.
- And the IPO pipeline is freezing: Nvidia-backed Nscale stumbled as Microsoft pulled out of its UK data-center arrangement, its valuation was cut ~40%, and details emerged of extreme backlog concentration — $43.8B of contract ceiling from Microsoft and $44.6B from Anthropic make up ~85% of a $103.4B "total contract value," versus just $2.6B in actual operating contracts. Fellow listing hopefuls SB Energy, Firmus, Vitro, and AirTrunk all delayed or scrapped offerings.
CEO Mike Intrator called the reaction "irrational" in an internal memo, arguing CoreWeave's contract backlog covers its obligations. The market's reply: backlog doesn't pay coupons — cash does, and at 16% yields, every refinancing is a wound.
Why the two halves coexist
This is the important nuance. Modal's $750M raise and CoreWeave's 16% bond yield are not contradictory — they're the same market sorting itself out. Equity is still chasing software-layer AI infrastructure with proven unit economics (Modal's $1B run rate at ~200 employees). Debt markets are repricing hardware-owning neoclouds that must constantly re-borrow to buy more chips. The market is learning to distinguish "infrastructure margin" from "financed hardware resale." One gets a 3.4x valuation step-up in five months; the other gets a junk yield.
What founders building on rented GPUs should do
- Diversify compute supply. If a single neocloud is your only GPU source, you are exposed to its credit risk, not just its uptime. Multi-cloud — across neoclouds, hyperscalers, and aggregators — is now a resilience decision, not just a cost one.
- Understand reserved vs. spot. Reserved capacity is a contract with a company that may be financially stressed; spot is market-priced but volatile. Match your workload's fault tolerance to the right mix, and know what happens to your reservations if a provider restructures.
- Model unit economics without cheap GPUs. If your gross margin depends on subsidized compute, you don't have a margin — you have a subsidy. Price as if compute costs 2–3x more and see what's left.
- Watch the 2030 horizon. If Gartner is even directionally right about neocloud consolidation, the surviving providers will have pricing power. Build portability now, while switching costs are low, rather than during a crunch.
The AI build-out isn't stopping — the money is still enormous. But the era of free leverage for GPU capacity is ending, and the cost of capital is becoming the binding constraint on AI infrastructure. Founders who build as if compute is abundant and cheap will be fine; founders who build as if it will always be financed cheaply may not be.
Sources: Morningstar; CNBC via InvestingLive; Reuters via IDNFinancials; Finimize; Gartner; biggo finance desk. Bond yields and ratings as reported mid-October 2026.
Read Next
General Translation Raises $27 Million Series A Led by a16z to Build Agent-First Localization Infrastructure
LangChain Raises $300 Million Led by CapitalG at a $3 Billion Valuation to Build the Agent Stack
Mount Thor Raises $45 Million to Run Cloud Fleets of Macs for AI Agents
OpenAI Reportedly Raises $15 Billion at a $1.25 Trillion Valuation
Q-CTRL Raises $290 Million Series C Led by Salesforce Ventures, Reaching Unicorn Status
SE3 Raises $30 Million Series A Led by Accel



