The AI Industry in Wartime

Anthropic recently rolled out a new policy: in certain scenarios, using Claude requires uploading an ID document plus a real-time selfie for identity verification. People in China are in an uproar, with many acting shocked.

Honestly, this is completely normal. From my perspective, this is actually pretty mild.

This Was Never Peacetime

Earlier this year, I told my team exactly this: while you can still use overseas models, don't get hung up on money or account issues. Just start using them. Master the intuition and methodology of collaborating with AI first. In as little as three months, you could run into problems and lose access. You have to approach this with a sense of urgency.

Since 2023, the AI industry has been on a wartime footing.

Look at when ChatGPT launched in 2023: GPT-3.5 was free, but Chinese IPs were blocked from Day 1. You needed an overseas phone number just to register a free account, and China blocked it domestically too. Things never loosened up after that—only got tighter.

By September 2025, Anthropic officially announced a ban on all "China-controlled companies" using Claude, regardless of where they operated—if Chinese entities held over 50% equity, you were out. Mass account bans began that November. By February 2026, things got even more direct: Anthropic publicly accused three Chinese AI companies—DeepSeek, Moonshot AI, and MiniMax—of using roughly 24,000 fake accounts to distill Claude's models, generating over 16 million conversations in total.

Just a few days ago, Peter Steinberger, founder of OpenClaw, also got banned. It wasn't targeted—most likely a false positive from anomaly detection—and he was later reinstated. But incidents like this are unsettling.

So I'm telling you: don't be so shocked. These model companies' intentions haven't changed since Day 0—they've been blocking you by every means possible. But mass bans cause collateral damage, and handling appeals puts enormous strain on their infrastructure. How many people do they even have managing this? Real-name verification is actually a fallback solution.

Chips and Models: Two Sides of the Same Coin

The chip side is just as turbulent.

In January 2025, the U.S. Bureau of Industry and Security issued new rules that completely choked off exports of NVIDIA's flagship H100 and H200 GPUs to China. By April, even compliance chips were banned. In July, the H20 was quietly unbanned on the condition that NVIDIA hand over 15% of its revenue to the U.S. government. By December, even the previously banned H200 was released, this time at a 25% cut. Huawei's chips were also banned by the U.S. in reverse. Back and forth.

The model side is the same. In February 2025, the U.S. Congress proposed the "No DeepSeek on Government Devices Act," prohibiting federal employees from using DeepSeek on government devices. The Department of Commerce, the Navy, and other federal agencies followed suit, as did states including New