China rations H200 chips: your cheap AI vendor is throttled

The cheap Chinese model you route overflow work to isn't cheap because it's magic. It's cheap because a handful of well-capitalized labs are competing hard on price. This week we got a reminder that their raw material — compute — sits behind a government valve. On July 8, 2026, reporting from The Information (picked up by Reuters and Bloomberg) said Beijing plans to let firms including Alibaba, ByteDance, and DeepSeek buy Nvidia H200s again — but on a leash.

What actually happened

This is a loosening, not an opening. According to Reuters and the South China Morning Post, the reported terms come with hard constraints. The total approved quantity is expected to land under 200,000 units — less than half of what Chinese firms requested this year. Companies have to submit quantities and intended use for case-by-case approval. The chips are cleared for AI model training only; for inference, Beijing wants firms to prioritize domestically produced processors. And they can process publicly available data only — not sensitive Chinese customer information.

Read that as rationing, not a reversal. China is threading a needle: give its national champions enough frontier silicon to keep training competitive models, without becoming dependent on U.S. hardware or exposing sensitive data to it. Note the caveat that matters for planning — this is a reported plan still under review, not a finalized, announced policy.

Why it matters for your business

We've written before that routing to cheaper models is a legitimate cost lever. It still is. But this is what "cheaper" is standing on: a supply of training compute that a government meters by the unit, caps below demand, and can retighten at will. The price advantage you're arbitraging is downstream of a policy decision made in Beijing — one you don't get a vote on and won't get advance notice of.

That doesn't mean avoid these models. It means don't marry one. The operator's insurance is the same as it's always been: a portable model layer. Keep your prompts, your evals, and your routing logic vendor-agnostic, so "the cheap option got expensive" or "the cheap option got throttled" is a one-line config change, not a re-platforming project. If a supply ceiling on a chip you'll never touch can move your inference bill, the model provider is a dial — build so you can turn it.

Sources: Reuters via TradingView — China plans to let top AI firms buy limited Nvidia H200 chips, Bloomberg — China to Let AI Firms Buy Nvidia H200 Chips, South China Morning Post.