The Investor

Today's through-line: the AI build-out rests on fixed obligations (loans, long-dated power contracts and take-or-pay capacity commitments) funded by revenue that is becoming variable and, at the model layer, falling. Power is where timing risk bites first. a16z reports that ERCOT's large-load queue grew from 63 GW at end-2024 to 474 GW by June 2026. Texas has paused energization approvals for every data center of 75 MW or more, leaving 17 projects stranded. Any Texas energization date in your models now depends on a state update to the governor on October 19 and an audit in December. Separately, MIT Technology Review reports that Finland halted Google's €13B data center on environmental grounds. That is a distinct permitting risk, not the Texas queue problem spreading to Europe.

SpaceX is seeking about $40B to buy Nvidia chips. The Information reports the split as roughly $10B of bank loans and $30B of investment-grade debt led by Apollo, with no equity. Anthropic's agreement to pay SpaceX up to $84.5B through 2029 effectively anchors the credit, so your Anthropic-dependent holdings now sit inside a lender's collateral chain. Reuters has not independently verified the underlying Financial Times report, and the deal is not expected to close until 2027.

Mistral's preview pricing for Large 4, a 1-trillion-parameter open-weight model, is $1.36 per million input tokens and $4.18 per million output tokens. AI Breakfast compares that with $10 and $50 for OpenAI's GPT-6 Astra. Morning Brew notes the cheaper models target the enterprise and government workloads OpenAI and Anthropic lean on as they prepare to go public. Mistral's benchmarks are self-reported, and the wider release is due October 27.

The Information reports that Anthropic found Zhipu AI's open-weight GLM-5.3 can run autonomous cyberattacks comparable to Anthropic's own Mythos model, and anyone can download it and strip its safeguards. JPMorgan's Jamie Dimon told Bloomberg that Mythos lifted global cyber risk 10-fold. CSO reports Google paused some bug bounty submissions because machine-generated reports are swamping triage. For your security deals, value is moving from finding flaws to validating and fixing them.

a16z's policy brief reports that the D.C. Circuit upheld the Department of War's exclusion of Anthropic's Claude from procurement. Anthropic also faces an FTC probe and is a defendant in a Sherman Act suit. Those three exposures attach to Anthropic specifically, so a single 'frontier lab' risk bucket no longer works in your IC memos. California has also set a statutory calendar for third-party AI auditors, with registration rules due January 1, 2028.