▶ YouTube Summaries

• https://www.youtube.com/watch?v=OQMPhkmeru0
• Original title: The World's Evilest Company

Prime reacts to Palantir CEO Alex Karp's on-air CNBC rant calling the AI industry "effing insane", plus Palantir's 9-point "AI sovereignty" manifesto. The real thesis isn't that AI labs are losing money — it's that heavy per-token usage funnels enterprises' proprietary data and business strategy to the model providers, who can then use it to move into and capture those same industries. Prime largely agrees, and points at Anthropic vs. Cursor and vs. Figma as evidence.

Token maxing = spend maxing

Karp's "token maxing": labs incentivize maximal token consumption the way a boat store upsells a fishing boat into a mega-yacht. Enterprises burn tokens, get little value back, and — worse — hand over their IP in the process. Manifesto point 3: token maxing "hijacks your value orientation and decreases institutional fortitude", incentivizing disposable scripts over robust software via "the addictive feeling of false progress." Prime admits he feels this in his own looping/token-maxing experiments — lots of spend, little value.

The real cost is your data, not the bill

The core worry isn't the money wasted, it's that every prompt, answer, and re-prompt exposes your defensible edge ("alpha" = trade secrets, not crypto-bro alpha). Karp's "safe" framing: safety means knowing who stores your data and preventing the model from caching it and replicating your business.

"Why charge for tokens if it's so valuable?"

Karp's sharpest line: if AI could truly triple your revenue, labs would skip consumer products and just take 30% of the upside per client. That they instead sell tokens and run consumer offerings suggests the value to them is the aggregated data — seeing "shapes of data" across millions of users to spot which industries are hot and worth entering. Prime jabs at Meta's ~$2B/year AI spend as an example of value not adding up unless data is the point.

Evidence: Anthropic vs. Cursor and Figma

• Cursor once made ~40–50% of Anthropic's revenue; Anthropic reportedly told Cursor that Claude Code was "just a research effort" — then shipped it as a direct competitor.
• ToS often bar building "competing projects" (i.e. competing with Claude), giving the provider a lever to drop partners.
• Figma: per The Information, Anthropic "blindsided" partner Figma with Claude Design; Anthropic's chief product officer sat on Figma's board until three days before launch. Prime calls this being "stabbed in the front."
• Zooming out: Anthropic also trained on the world's books and knowledge — a pattern of ingesting others' data.

Takeaway

No smoking-gun proof that partner data directly drove these product moves, but the pattern fits: with effectively unlimited token spend, a lab can slop together a "close enough" competitor in any industry it can see demand for — because it holds every prompt and answer. Manifesto's closing: "Data retention is your treasure. Transfer it at your own peril." Prime, reluctantly, sides with Karp: the insane part is the potential to weaponize partners' data to capture their industries.