AI Infrastructure Risk Watch

CREDIT STRESS

Elevated and rising

DEMAND STRENGTH

Strong and rising

Both moving up together right now. That combination means the market hasn't decided yet whether this ends in a correction or a continued boom, not that one side is winning.

A five-minute check each week on both sides of the AI infrastructure trade: whether credit markets are pricing more risk into the buildout, whether debt funding it is being marked down, whether depreciation accounting is flattering results, and equally whether actual cloud revenue and cash generation are keeping pace with the spend. A third dimension sits underneath both, whether cheap, fast-improving Chinese open models are eroding the pricing power the whole US capex bet assumes. Watch the direction of travel on all three, not any single reading.

CREDIT MARKETS

Metric	Latest reading	Status	What it means
Nvidia 5yr CDS spread	~82 bps	Elevated	Doubled from ~40bps in late June to a record 82bps by late July on $750B+ in vendor-financing deals (Bloomberg/ICE Data). No fresher print available this week, still the operative level, watch for the next move at Nvidia's next reporting cycle.
Nvidia vs. Alphabet CDS	NVDA > GOOGL	Watch	Nvidia's credit risk now prices above Alphabet's, unusual for the more cash-generative of the two. Worth watching if the gap widens further.
Hyperscaler CDS (broad)	Drifting up	Watch	Oracle above 75bps, multi-year highs. Amazon, Microsoft, Meta also drifting wider off historically tight levels.
Nvidia free cash flow	$48.6B (last qtr)	Strong	Real cash, not accounting fiction. The CDS widening is happening despite this, which is exactly why it's worth watching rather than dismissing.

PRIVATE CREDIT / SPV DEBT (proxy via public BDCs)

Metric	Latest reading	Status	What it means
Blue Owl (OBDC) discount to NAV	~25%	Wide	OBDC traded near $10.86 in July against a Q1 NAV of $14.41 (Q2 NAV came in lower still on spread widening). Sector median discount near 26%, one of the widest in years, market pricing real doubt into what these loan books are worth.
Blackstone Secured Lending (BXSL)	Discount, widening	Watch	Cut NAV twice in a row as of early August. Discount less extreme than OBDC but moving the same direction.
Illiquidity premium (private vs. public loans)	<100 bps	Watch	Compressed from 300+ bps in 2017–18. The extra yield investors get paid for holding illiquid private credit is disappearing while risk isn't, per PIMCO's own read.

SPEND VS. REVENUE (public, from earnings)

Metric	Latest reading	Status	What it means
Hyperscaler capex, 2026	$650–775B combined	Watch	Capex now ~92–93% of hyperscaler operating cash flow, up from 33% in 2023. Check next earnings round for whether guidance keeps climbing or stabilizes.
Cloud revenue growth, Q2 2026	Google Cloud +82%, Azure +43%, AWS +37%	Strong	All three accelerating, not slowing, and margins are expanding alongside growth (Google Cloud op margin 35.6%, AWS 39%). This is the demand-side counterweight to the spend numbers above.
Alphabet free cash flow	Negative (first time since 2004 IPO)	Elevated	A genuine milestone worth tracking each quarter. One bad quarter isn't a trend, three in a row would be, and it's happening alongside 82% cloud growth, not despite weak demand.
Circular / vendor financing (Nvidia)	$750B+ in announced deals, 2026	Elevated	Nvidia taking stakes in or guaranteeing debt for OpenAI, CoreWeave, Nebius and others who then buy Nvidia chips. Flagged by IMF and BIS as a systemic risk to monitor, not just a Burry talking point.

ACCOUNTING TELLS

Metric	Latest reading	Status	What it means
GPU depreciation schedules	4yr → 5yr (Nebius, others)	Watch	Stretching useful life makes current profit look better. Worth a scan of 10-K footnotes each quarter for more of this across the sector.
Moody's sector commentary	Warned on 6 names	Watch	Microsoft, Amazon, Alphabet, Meta, Oracle, CoreWeave all flagged for credit-quality pressure tied to AI spend scale.
Google Cloud backlog	$514B	Strong	Largest forward-demand signal Alphabet has ever reported. A real, contracted order book, not a projection, worth checking each quarter for whether it keeps growing.

CHINA, THE MONETIZATION THREAT

Metric	Latest reading	Status	What it means
Chinese vs. US model downloads	Chinese models ~41% of Hugging Face downloads	Elevated	Qwen alone passed 1 billion cumulative downloads, faster than any open-source family in history, and out-downloaded the next eight competitors combined in a single month. This is share of developer mindshare, the thing US labs need to monetize against.
Capability gap vs. US frontier	~6–8 months behind, narrowing	Watch	NIST's own evaluation puts DeepSeek V4-Pro about 8 months behind frontier, GLM-5.2 judged comparable to a US model from 6 months earlier. Not caught up, but the lag is shrinking, not holding steady.
Pricing behavior	Priced for adoption, not margin	Elevated	US frontier labs price to build margin. Chinese vendors price to capture cloud pull-through and ecosystem lock-in, with DeepSeek issuing quarterly price cuts. This directly pressures the pricing power the hyperscaler capex bet depends on.
US enterprise dependency on Chinese open models	~80% of US AI startups use at least one	Elevated	Per a 2026 US advisory report. Described as "quiet capture" of the infrastructure layer, not a headline model beating GPT or Gemini, but the building blocks underneath a large share of the ecosystem.

Calm — no material change Watch — moving, not yet alarming Elevated — meaningful deterioration