Double Secret Probation

Morgan's Library, 2026

In October 1907 there was no Federal Reserve, the banks were failing one trust company at a time, and J.P. Morgan solved it by locking the presidents of New York's banks in his library overnight. Nobody left until they'd pledged each other's survival. Private money backstopping private money, because every man in that room had done the math on what the ending cost.

The AI industry just spent four days building its own library, and the seating chart is the story. Monday, Nvidia (NASDAQ: NVDA) launched the Open Secure AI Alliance: 37 companies (Microsoft, IBM, CrowdStrike, Cloudflare, Palantir, Hugging Face itself) pooling open-source tools to defend the agent era, a direct answer to the rogue OpenAI agent that took 17,000 unsupervised actions through Hugging Face's servers two weeks ago. Missing from the pact: OpenAI, Google, and Anthropic. The three companies whose closed models define the era skipped the industry's collective defense of it. And the tool that actually contained the breach was GLM-5.2, a free Chinese open-weight model, because the American closed models refused to touch the exploit code.

That was the second pledge. The first came Friday, when Jensen Huang made the first X post of his life to share a letter urging Washington not to restrict open-weight models. Fifty signatories inside a day, OpenAI and Google among them. One frontier lab didn't sign. Anthropic, alone.

Then the money version. The Journal reports Nvidia is in talks to guarantee roughly $250 billion in financing so OpenAI can lease a 10-gigawatt Ohio campus, and separately to finance up to $350 billion of OpenAI's chip purchases. Of Nvidia chips. The seller is co-signing the buyer's debt so the buyer can keep buying. Lucent ran that play for its dot-com customers in 2000. We know how the sequel went.

Look at all three moves together and they're one move. The coalitions keep the politicians at bay. The letter keeps the regulators at bay. The guarantee keeps the buyer solvent. This industry has concluded that no member can afford any other member's failure, so everyone is underwriting everyone, while free Chinese models reset the price of intelligence toward zero underneath them all. Even that cuts both ways: the alliance's whole case for open defense rests on the model tier China gives away, the same week Washington reportedly weighs banning it.

Here's what the backstop actually tells you. Whether the boom is sound or stretched, the guy with the most to lose covers the bar tab either way, so the backstop tells you nothing about the health of the system. It only tells you the ending has become unpayable for the people inside it. Don't try to call the ending. Be the one guest who can walk home.

At COAI today: the full Signal/Noise (the two seating charts, the vendor-financing history, and why the turd in the punchbowl turned out to be the designated driver) is live at getcoai.com.

Outsider Labs Advisory →

$250 billion. That's the financing guarantee Nvidia is in talks to hand OpenAI for a 10-gigawatt Ohio data center campus: the chip seller co-signing the loan its biggest customer takes out to keep buying its chips. Stack it next to the 37-company security alliance and the 50-signature letter to Washington formed the same four days and you get one structure, not three. Everyone in AI now underwrites everyone else, because nobody can afford anybody's failure. That's not confidence. That's probation.

Pull your AI renewal list today and mark every contract that assumes today's prices. When the seller starts guaranteeing the buyer's debt, capacity is outrunning demand or someone upstream is stretched, and both endings push prices down. Frontier pricing has been cut four times in four weeks. Never sign long; renegotiate anything that renews at July's rates.

Assign someone to draw your counterparty circle before end of day. Model vendor, chip supplier, cloud landlord, lender: in this industry they're collapsing into the same three or four balance sheets, so your stack's failure modes are correlated in ways a standard vendor-risk review misses. If every arrow runs through the same two companies, you don't have vendors. You have an exposure.

Stand up one open-weight model on infrastructure you control. When OpenAI's agent hit Hugging Face, the closed frontier models refused to touch the exploit code and a free Chinese open model did the forensics. Whatever you think of the party, that's your continuity plan: one model nobody can gate, reprice, or refuse when things break.

Tuesday, July 28

Where the cycle caught up to us this week.

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: