Every Financial Guardrail Just Reached the Same Conclusion

Northern Virginia's data centers are 99.7% full. Globally, vacancy sits at 6.7% — both record lows. CBRE's Pat Lynch made the state of the market plain.

My thoughts are we’re at least three years out before we catch up with the supply that equal the demand that's out there. — Pat Lynch

That sounds like good news. It is not. The fullness is the shape of the danger: the risk in AI infrastructure is not empty buildings but fully-utilized compute that cannot generate enough revenue per token to service the debt that built it. ^[1] Over the past several weeks, seven major institutional guardrails have independently arrived at the same diagnosis. None coordinated. None hedged. S&P Global downgraded Oracle to BBB, citing leverage from its AI build-out. Fitch was blunt about what lies ahead.

These companies are very large established players with, in many cases, pristine balance sheets, but they’re embarking upon almost unprecedented capital investment. — Moody's Ratings

The Bank for International Settlements published a study that drew an uncomfortable historical parallel.

Financial stability could also be at risk in the event of an AI bust. — Bank for International Settlements

Career analysts at the U.S. Treasury drafted a report drawing the same dot-com comparison but warning this one is more deeply embedded in the economy — a downturn would transmit through stock markets, chip manufacturers, utilities, and private credit simultaneously. The Bank of Canada's Senior Deputy Governor Carolyn Rogers described the precise contagion mechanism.

A cascading series of events could cause a sharp loss of investor confidence and lead to a spike in demand for liquidity or rapid asset sales. Funding markets could come under pressure, and stress could spread more broadly. — Carolyn Rogers

Moody's flagged the same leverage buildup. The Bank of England added its own warning. Seven guardrails, one conclusion, no coordination. ^[2][3] The mechanism they are all watching runs through a specific chain. On May 23, DeepSeek permanently cut its V4-Pro model prices by 75%, lowering API fees to as little as 0.025 yuan per million tokens. Third-party benchmarks show it costs 12 to 19 times less than OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 for equivalent tasks.

It is an efficiency gain being passed through. — Sanchit Vir Gogia

That price cut travels through a debt structure that has been layered three deep. At the startup level, GPUs themselves are loan collateral — Alpha Compute secured a $31.9 million non-recourse loan backed by Dell Nvidia GPUs in April. At the investor level, equity stakes are pledged as collateral — SoftBank is seeking a $10 billion margin loan against its OpenAI stake while already owing $40 billion due by March 2027. And in between sits circular financing: CoreWeave owes $14 billion in debt with $34 billion in upcoming lease payments, while Microsoft provides up to 70% of its revenue. Microsoft is both CoreWeave's biggest customer and its counterparty. ^[4][5][6] The revenue that services all this debt depends on token pricing. And token pricing is collapsing. Jefferies made the connection explicit this week.

The reason credit risk has become more of an issue is that it can no longer be assumed... that large language models will ever be profitable given the related ongoing collapse in token pricing. — Jefferies Group

China processed 36.39 trillion tokens in the week ending July 19, against 7.39 trillion for leading U.S. models. Corporate customers are pivoting to cheaper Chinese alternatives from Moonshot and Zhipu, reversing Jensen Huang's dictum that high token burn equals high productivity — as Moody's Vincent Gusdorf noted, the bills started piling up and companies realized the tools are expensive enough to require disciplined use. ^[7][8] The numbers beneath this are staggering. Big Tech's 2026 capex has reached $234 billion across Amazon, Alphabet, Meta, Microsoft, and Oracle. Meta alone is spending $125 to $145 billion on AI data centers — nearly equal to its projected 2026 EBITDA of $145 billion — and Mark Zuckerberg declined to provide a return-on-invested-capital timeline. OpenAI has contracted over $900 billion in compute commitments while projecting only $25 billion in 2026 revenue. Morgan Stanley reports hyperscalers doubled their gross leverage ratio from 0.9x to 1.8x in two quarters. ^{[9][10][11][12]} The White House has already removed the public cushion. David Sacks, the administration's AI advisor, answered the question directly.

We believe that governments should not pick winners or losers, and that taxpayers should not bail out companies that make bad business decisions or otherwise lose in the market. — Sam Altman

There will be no federal backstop. If the debt structure unravels, the losses stay with private creditors and equity holders. ^[13] The data centers are not empty. Northern Virginia is at 0.3% vacancy. Micron's high-bandwidth memory supply is sold out through 2028. Nvidia's $25 billion bond sale drew $85 billion in demand. The physical infrastructure is committed, the buildings are full, the machines are running. And the revenue per token still cannot service the debt that built them. The guardrails have spoken, the government has stepped back, and the private creditors holding the layered debt are the ones who will discover that fully-utilized infrastructure can still lose money. ^[1]