Why China's AI Export Curbs Won't Slow Open-Weight Models

China's draft export curbs target the same AI model weights that are already winning on usage and price, a contradiction worth watching rather than a closed case.

Why China's AI Export Curbs Won't Slow Open-Weight Models: key figures
ShareXLinkedInEmail

THE RUNDOWN

1

China's draft rules single out large models and agents, but the open-weight models already winning share, like DeepSeek's and Moonshot's, are hosted mostly by US intermediaries today.

2

A licensing contingency is now a real diligence question for any startup whose stack depends on a Chinese open-weight model's continued availability.

3

Washington's own open-weight debate complicates the picture: a US ban on using Chinese open models would hit the same companies currently benefiting from cheaper inference.

4

Export control is becoming a lagging indicator of where compute economics already pointed, not a leading one.

The VC Read

Value Add VC analysis

The real divergence to track is between API access and weight hosting -- almost everything restricted here is the latter, and almost everything US startups actually touch is the former. We'd bet the carve-out survives final text, because killing API access would hurt US-based inference resellers more than it hurts Beijing.

Analysis

Our read: China's draft export curbs on AI model weights are targeting a category that's already winning on its own terms, and no amount of catalogue language changes that the economics got there first. The draft restricts AI models above 2 trillion parameters and bans agents for sensitive settings, but the open-weight models actually moving the market, DeepSeek's and Moonshot's among them, are already handling 71.5% of tokens on gateways like Vercel's -- mostly through US-based hosts, not direct exports Beijing could easily unwind.

That's the contradiction worth sitting with: export control assumes the thing being restricted still has to cross a border to matter. Model weights already crossed that border months ago, got rehosted by US infrastructure providers, and are now priced into products those providers sell domestically. A registration rule that took effect this week doesn't claw any of that back. If anything, it's a lagging acknowledgment that the compute-economics argument -- cheaper, good-enough open models win developer share -- already played out before the policy caught up, as The Register reported.

“That's the contradiction worth sitting with: export control assumes the thing being restricted still has to cross a border to matter.”

For founders and allocators, the practical diligence item isn't "is this company exposed to China" -- it's "does this company's cost structure assume continued, unrestricted access to a specific open-weight model's license terms." That assumption got meaningfully riskier this week, even if nothing changes for months.

Room for disagreement: a draft with a month-long comment period and an explicit API/SaaS carve-out is closer to a negotiating position than binding policy, and Beijing has walked back similar moves before -- the US's own June restriction on Anthropic's foreign-user access was reversed within months. If the carve-out survives final text, most US companies renting access through an API never touch the restricted category at all, and this whole debate is academic for everyone except the handful of firms hosting raw weights directly.

ShareXLinkedInEmail

Key Sources

2 sources

THE WIRE in your inbox— Tech, startup & VC news with The VC Read, a few times a week. Free to subscribe, no spam.

No spam, unsubscribe anytime. Sent via Substack · Privacy · or subscribe on Substack ↗