03 Aug, 2026

Log 024 - Margin Call

Leopold Aschenbrenner's AI hedge fund sold its public stock portfolio after a sharp reversal. Nvidia considered financing hundreds of billions of dollars of OpenAI infrastructure. A new open-security alliance formed without OpenAI, Anthropic, or Google. And a five-year-old Coldcard flaw exposed thousands of Bitcoin wallets.

In This Log

1. Aschenbrenner Sells His Public Portfolio (Markets)

On July 30, Situational Awareness sold its public stock portfolio to Ken Griffin's Citadel after heavy losses in AI infrastructure stocks. The fund had recently reported about $20 billion in assets under management. 1

Its assets were estimated at roughly $10 billion after the sale. Earlier reports had put the fund's peak as high as $45 billion, but that was not its size immediately before the unwind. 2

An investor letter sent on July 24 said the fund had returned 439% through June. It called the selloff a buying opportunity and asked clients for fresh capital.

The fund was concentrated in memory, data-center, and power companies. Several of those holdings fell more than 30% during July, and leverage made the losses harder to absorb. Situational Awareness kept its private investments, including a large stake in Anthropic.

Why it matters

This was the leveraged version of the AI infrastructure trade. The companies did not fail, but the fund could not wait for them to recover. A strong long-term view can still produce a forced sale when borrowed money sets the timeline.

Reality check

The 439% return came from an unaudited investor letter, and the asset figures are estimates from different points in time. Citadel bought the portfolio, suggesting it still saw value in the positions. The fund also kept private holdings that could recover a large part of the loss.

2. Nvidia Backs Its Buyers (AI / Markets)

On July 27, Nvidia was in talks to guarantee as much as $250 billion of financing for an OpenAI data center in Ohio. It was also considering financing up to $350 billion of OpenAI's chip purchases. 3

The same day, Nvidia invested in Ilya Sutskever's Safe Superintelligence and gave the lab access to Vera Rubin systems. The companies said the partnership would increase SSI's compute by an order of magnitude. 4

The talks followed a $500-billion-plus partnership announced with South Korea's SK Group. That plan covers a two-gigawatt AI factory and a long-term memory partnership, but it is still based on letters of intent. 5

The concern is simple. Nvidia sells the chips, invests in the labs that buy them, and may help finance the data centers where they run. That makes it harder to separate outside demand from demand supported by Nvidia's own balance sheet.

Why it matters

Nvidia is becoming more than a supplier. It is helping fund the companies and infrastructure that create demand for its products. If those customers struggle to raise money on their own, Nvidia's balance sheet may become part of the AI buildout.

Reality check

The OpenAI arrangements are talks, not completed deals. The SK plan is also preliminary. Vendor financing is common in capital-heavy industries, and the chips and data centers would still be real assets. The risk is that financing can pull future demand forward.

3. AI Security Splits Open and Closed (AI)

On July 27, Nvidia and 36 other organizations launched the Open Secure AI Alliance. The group will build open tools for finding, sharing, and fixing vulnerabilities in AI systems. 6

The alliance includes Microsoft, Hugging Face, CrowdStrike, and the Linux Foundation. OpenAI, Anthropic, and Google are not members. 7

The launch directly cited the Hugging Face breach. Closed tools had limited the investigation, while an open-weight model could be run inside Hugging Face's own systems to help contain the incident.

The split is not absolute. Microsoft is deeply tied to closed models and still joined the alliance. OpenAI and Google also signed a separate industry letter supporting open-weight AI, while Anthropic did not. 8

Why it matters

The argument over open models is now a security argument. Open systems let defenders inspect and run tools themselves. Closed systems give labs more control over access. The Hugging Face breach showed that each approach can also create a blind spot.

Reality check

The alliance has not shipped its main tools yet, and missing from the member list does not mean opposed to its goals. Most large technology companies use both open and closed models. The practical test is whether the group produces shared tools that defenders actually use.

4. $89 Million and a Five-Year-Old Bug (Crypto)

On July 30, an attacker swept 1,196 Bitcoin addresses in 41 minutes. Two more suspected waves followed, taking the estimated total to 1,367 BTC worth about $89 million across 4,585 addresses. Researchers said the activity was still ongoing. 9

The wallets were created with Coldcard hardware devices. A firmware change made in March 2021 sent seed generation to a predictable software random-number generator instead of the device's hardware generator. 10

The flaw reduced the estimated security of seeds to about 40 bits on Mk2 and Mk3 devices and about 72 bits on later models, instead of the intended 128 bits. An attacker could test possible seeds offline against public Bitcoin addresses.

Coinkite released fixes for the affected models and told users to create new seeds and move their funds. Installing new firmware does not repair a seed that was already created with the faulty code. 11

Why it matters

A hardware wallet is trusted to create a private key that nobody else can guess. Here the failure happened during that first step, where users had no practical way to check the result. The wallets stayed exposed for more than five years.

Reality check

This was a Coldcard firmware failure, not a break in Bitcoin's cryptography. Seeds strengthened with enough private dice rolls, or protected by a strong unique passphrase, have an extra defense. The $89 million total is based on transaction patterns, and researchers have not proven that every address came from the same bug.

Signals

Microsoft and Meta Split on AI Spending

Microsoft reported $90 billion in quarterly revenue, while annual Azure revenue passed $100 billion for the first time. 12 Meta grew revenue 28%, but expenses rose 55%, Reality Labs lost another $4.62 billion, and planned 2026 capital spending increased to $130-145 billion. 13

OpenAI Previews Astra in Washington

Sam Altman spent the week showing OpenAI's unreleased Astra model to lawmakers and White House officials. OpenAI also said an internal version helped produce advances on ten open problems in mathematics and theoretical computer science. 14

IBM Claims Three Quantum Advantages

IBM and research partners presented three experiments they say beat the best known classical methods on specific problems. The results used IBM Heron R3 systems, but the papers were still preprints and the claims remain open to better classical challenges. 15

Rocket Lab Wins Its Largest Launch Contract

Rocket Lab won a $266 million Space Force contract for 12 suborbital missile-defense launches, with options for six more. Most will fly from a new company location at the Pacific Spaceport Complex-Alaska in Kodiak. 16

Meta-Thread

Situational Awareness and Nvidia exposed two sides of the same financing question. One used leverage to concentrate an AI infrastructure trade and had to sell when the market moved. The other may use its own balance sheet to help customers keep buying chips and building data centers.

The security stories exposed a second problem. Hugging Face could not fully inspect the closed tools involved in its breach. Coldcard users could not see that their wallet seeds came from weak randomness. In both cases, the system looked secure until someone tested the assumption underneath it.

Next Log drops next week.

© 2026 AELIUM // Nothing here is advice // readable by humans and agents