A Big Shift in the AI Race

What It Covers

The AI race enters a realignment phase as Anthropic's Washington standoff over Mythos and Fable five drags into its fifth day, SpaceX's post-IPO valuation hits $2.6 trillion after Cursor acquisition, and leaked OpenAI financials reveal $13 billion in 2025 revenue with profitable inference margins.

Key Questions Answered

• •Government Relations as Core Business Function: AI companies at Anthropic's scale must treat federal relationship management as a primary operational priority, not a secondary concern. Anthropic took over an hour to return a critical government call, delayed providing Project Glasswing recipient lists by days, and reportedly had CEO Dario Amodei reachable through Amazon's Andy Jassy before US officials — all contributing to regulatory escalation.
• •Jailbreak Regulation Reality: Regulators demanding a complete jailbreak fix before restoring model access may be requesting something technically impossible. Former OpenAI board member Helen Toner states that fully eliminating jailbreaks in large language models is widely considered unfeasible within the research community, making compliance-based negotiations structurally difficult for Anthropic to resolve on any defined timeline.
• •AI Inference Profitability Signal: OpenAI's leaked audited financials show the core token-selling business generates solid margins — $13 billion in 2025 revenue against $7.5 billion in direct costs. Stripping a one-time $30 billion non-cash accounting charge from the corporate restructuring, actual operational losses were approximately $8 billion, with $73 billion in cash currently on the balance sheet.
• •NeoCloud as SpaceX's Revenue Engine: SpaceX's Colossus one and two supercomputer facilities transformed the company's economics by leasing compute to Anthropic and Google, making NeoCloud its top revenue source almost overnight. This infrastructure monetization strategy, combined with the $60 billion Cursor acquisition at a $4 billion annual run rate growing 7x year-over-year, repositions SpaceX as a full-stack AI competitor.
• •Cursor's Efficiency-First Model Strategy: Cursor's Composer 2.5 model matches Claude Opus and GPT performance at roughly one-tenth the cost by focusing on post-training optimization using a Kimi K base. A new model trained from scratch with 10-20x more compute is releasing within weeks, signaling a potential shift from pure efficiency toward competing at the frontier state-of-the-art level.