September 30, 2025 · JustSayAI

Anthropic Bans Chinese Firms: Security Theater, Revenue Grab

Word is Dario (Anthropic CEO) once worked at Baidu’s Silicon Valley lab. Either he was dazzled by Chinese engineers’ grit, or he got scammed buying GPUs on Baidu—either way, the grudge is now policy.

I. Why the anti-China ban? Moralizing is the excuse, money is the motive

Anthropic’s new “no Chinese-controlled companies” clause wraps itself in national-security speak—model distillation! data leakage!—but only rubes buy that. This is a naked revenue play, a kneeling oath to the Beltway.

OpenAI open-sourced a lightweight model last month and walked straight into a Pentagon contract. Anthropic wants the same buffet. Enterprise deals? Sure, it’s winning some. But the trillion-dollar gov/DoD trough requires louder loyalty theatrics. The ban is the ticket, timed to Anthropic’s sprint from $5B to $10B ARR and a valuation moon-shot.

II. Claude ascends, agents ride its coattails

Cursor’s viral 2024 and Manus’s spring hype look like app-layer magic; they’re really Claude overflow. When Claude 3.7 shipped, Cursor’s valuation doubled overnight. Claude 4.1 lifts general-agent success rates again. Under the hood: the base model keeps leaping.

Add Anthropic’s Model Context Protocol (MCP) and the “brain” learns to wield tools—search, IDEs, docs—turning agents from chatty theorists into hired guns. Lesson: every agent party is just runoff from a stronger foundation model.

III. Too late: Chinese labs already ship 80% clones

The punch lands on China, yet the long-term bruise is Anthropic’s. Overnight Zhipu launched one-click Claude migration—same API shape, price tag sliced in half and denominated in yuan. The turf is ready.

Big Chinese shops will grumble through a stack swap, then feed data, users and revenue into domestic models. An involuntary, government-sized internal market just got born.

Can local LLMs fully replace Claude? Real talk: 80% yes. For everyday coding and content, Kimi costs me a few dozen RMB a day; Claude bills make you question life choices. On the remaining 20%—gnarly debugging, multi-file orchestration—Claude and GPT-4o still edge ahead. But capabilities converge; user habit is the final moat.

Brute-force scale plus rock-bottom cost is the real miracle. Once devs get used to solving 80% of problems at one-tenth the price, a new, steel-spined AI paradigm sticks.

Cost is the ultimate productivity.

Anthropic’s grandstanding may unlock the very market it wants to starve. Every slammed door forces open a dozen windows—windows labeled “Made in China, 50% off.”

Viewpoints condensed from my podcast: People’s Park Talks AI

👉Subscribe to “JustSayAI Daily Brief” · twice a day · one-click listen: https://justsayai.org/newsletter (science上网 required)