The \$3 Trillion Off-Balance-Sheet Spend, OpenAI's Safety Dissolution, and the Sacks-Amodei Ideological Clash

Highlights

Today’s AI landscape is dominated by eye-watering macroeconomic figures and deep institutional debates. From the Wall Street Journal’s revelation of $3 trillion in off-balance-sheet AI commitments to the quiet dissolution of OpenAI’s last remaining safety team ahead of its anticipated IPO, the tension between massive capital spending and corporate governance has reached a boiling point. Meanwhile, a high-profile debate between David Sacks and Dario Amodei highlights the profound ideological split over whether AI’s future should be decentralized or centrally regulated.

Top Stories

• The $3 Trillion Off-Balance-Sheet Spend: Nine top tech companies have committed an astonishing $3 trillion in off-balance-sheet obligations, mostly tied to AI infrastructure, which is triple their outstanding leases and long-term borrowings. Meta alone has committed nearly $700 billion in future spending, raising serious concerns about a potential “SPV bubble” and depreciating chip assets if closed-weight models fail to yield expected margins. (Source)
• AI Borrowing Distorts National Interest Rates: Nomura and Bank of America report that the sheer volume of debt issued by AI companies (now equaling 25% of US Treasury bond issuance) is actively pushing up the 10-year Treasury yield by 0.3 percentage points. Lenders are heavily favoring corporate bonds like Meta’s 7.5% data center bonds over 5.2% Treasuries, effectively making the AI arms race a direct competitor with the US government for capital. (Source)
• OpenAI Quietly Dissolves Final Safety Team Ahead of IPO: OpenAI has quietly dissolved its last remaining team dedicated to catching catastrophic risks, representing the third safety team to be disbanded in two years. Commentators note that this final abandonment of their 2023 pledge (to dedicate 20% of compute to safety) occurs just as the company prepares for an anticipated high-profile IPO, where internal hazard documentation would face intense S-1 regulatory scrutiny. (Source)
• Greg Brockman Defends OpenAI Departures Amid High Growth: Appearing on CNBC’s Squawk Box, OpenAI President Greg Brockman dismissed concerns regarding the company’s recent executive shakeups, asserting OpenAI is a “highly resilient organization” with a deep leadership bench. Brockman deflected questions on why key leaders departed and instead pointed to stellar financial growth, noting that July revenue grew 20% month-over-month, led by a 32% surge in the enterprise segment. (Source)
• Zero Train-Infer Mismatch: River API Outperforms Tinker on RL MoE Runs: Researcher Ashwinee Panda announced a breakthrough enabling reinforcement learning on large Mixture of Experts (MoEs) with zero train-infer mismatch. In benchmark runs, the open-source River API outperformed Tinker while training a Qwen3.6-35B model to play Wordle. (Source)

Articles Worth Reading

[David Sacks: The Battle Over AI Centralization and Regulatory Capture] (Source) In a comprehensive critique, venture capitalist David Sacks argues that Anthropic’s push for federal pre-deployment testing (which he dubs a “DMV for AI”) is a calculated move to create regulatory capture that protects incumbent labs while handicapping open-source models. Sacks critiques Dario Amodei’s framing of safety, arguing that concentrating gatekeeping power in a federal bureaucracy harms U.S. competitiveness against China and serves Anthropic’s commercial business model. Sacks cleanly summarizes the dispute: “Dario believes frontier AI is too powerful to distribute; we believe it is too powerful to centralize”. Amodei responded directly, defending Anthropic’s proposals (like CA SB 53) as championing small competitors and open-weights by exempting them from coverage, and noting support for the Trump administration’s pre-deployment testing and Demis Hassabis’ FINRA-like entity proposal.

[I Wrote 200 Lines of Rules for Claude Code. It Ignored Them All.] (Source) A power user shares a confession of writing 200 lines of rules and safeguards for Claude Code across 258 knowledge base files, only to find that the system ignored all of them. The article highlights a frustrating paradox: the more specific instructions you give to current LLM-centered systems, the less likely they are to follow them. Gary Marcus highlights this as evidence that LLMs suffer from endemic architectural limitations that cannot be fixed without a fundamental redesign.

[How Yana Bana Uses Codex to Build an AI-Native Everlane] (Source) Solo founder Yana Welinder showcases how she is building “Yana Bana,” an AI-native clothing brand, using Codex as her technical co-founder. The workflow takes sketches to runway-ready videos, uses computer use to operate specialized design software, and researches manufacturing vendors. It’s a fascinating, practical showcase of agentic workflows operating in the physical world, complete with a friendly competition between AI and human pattern makers to see who can produce the best design first.

📊 I could compile these off-balance-sheet commitments and interest rate data points into a structured spreadsheet model to map out the financial runways and borrowing impacts of these hyperscalers.