The Alignment Tax

Microsoft's equity stake in Anthropic means this is financial self-defense as much as principle — protecting a $4B+ investment from a government designation that would crater its value. Google and OpenAI's support is similarly self-interested: a legal precedent that lets the Pentagon blacklist AI companies for ethical constraints is a weapon any future administration could use against them. When financial interest and ethics align this neatly, calling it a values test overstates the moral clarity.

ARC-AGI-3 was designed by François Chollet specifically to be hard for LLMs — it measures a narrow definition of reasoning (novel game generalization) that may not be relevant to the work most businesses care about. GPT-5.4 scored 83% on GDPVal (OpenAI's own benchmark — directionally correct, not independent validation) and solved an open math problem this same week. The 0.26% result says AI can't learn new games from scratch; it says nothing about whether AI can do your accounting, write your code, or draft your legal brief — tasks it's already doing at scale. ARC-AGI-3 tests one specific cognitive capability that happens to be hard to fake with pattern compression.

Adobe at 12x forward P/E and Salesforce at a 10-year price low look like deep value to any investor who believes the software layer survives AI. These companies have data moats, customer lock-in, and compliance requirements that AI agents can't easily replicate — try telling a regulated financial institution to swap Workday for a custom agent. And 95% of generative AI pilots fail to reach production scale (Deloitte, March 2026). The seat compression thesis assumes enterprises will replace software with AI agents; the data shows most can't even get agents off the ground.