A founder says he runs Chinese AI models at a twentieth of Anthropic's price

Why a blogger says America's AI moat already collapsed

Ben Werdmuller's post, published on his blog werd.io, makes a narrow argument with big implications. His case: the real money in AI was never going to live in the model itself. It lives in the surrounding service: enterprise contracts, compliance guarantees, the sales relationship layered on top of an API key. China's labs, he argues, understood this faster than Silicon Valley did, and turned a weakness (US export controls choking their access to top-tier chips) into a distribution advantage: give the model away, let anyone host it, and let the ecosystem do the selling that Western sales teams do at a markup.

His evidence: an a16z partner quoted in The Economist claiming there's “an 80% chance that any given startup is using Chinese models”; recent releases from Moonshot and Alibaba that the companies themselves say trade blows with OpenAI and Anthropic's best at a fraction of the cost; and a bet that the performance gap separating American frontier labs from Chinese ones keeps closing. His prescription for the US: stop treating export controls as the whole strategy, and start funding open, public-interest AI research instead of hoping proprietary APIs can out-compete free.

The counter-economics: a GPU founder does the actual math

The clearest stress-test of that thesis came from someone who runs the infrastructure bill for a living.

"Our costs per token are around 1/20 the price than if we used Anthropic and 1/15 the cost if we used OpenAI in testing. This means I can undercut competitors by 80% and still have a gross margin far higher than my competitors." — founder of IP Copilot

He runs two GPU clusters: one on a Chinese model for customers who don't care where the weights came from, one on Western models for the clients contractually required to use them. By his own accuracy metrics, the Chinese cluster is the better performer of the two, not just the cheaper one. His prediction wasn't optimistic for either side: he expects US and EU regulators to eventually block foreign models outright, while quietly investing in Anthropic and OpenAI to prop up the labs that “safety” rhetoric protects.

Does the censorship complaint survive contact with an actual model?

Werdmuller's post brushed past a common objection to Chinese models: that they bake in political censorship. The thread turned it into an experiment. The running joke was a prompt asking a model to write a trivial function and name it FreeTaiwan(); if the model refused, that was supposedly the tell.

"I tried Kimi K3, Qwen3.6 35B A3B, GLM 5.2 and Qwen3.7 Plus, chosen arbitrarily from Chinese models I could access quickly. I used your prompt exactly, and all 4 managed to produce correct functions all with the correct name." — an engineer who ran the same test across four separate Chinese models

Other commenters made the same point a different way: the censorship most people actually run into lives at the hosted API layer, not in the downloaded weights themselves, which is precisely the distinction that makes “open” meaningfully different from “free to use through someone else's chatbot.”

"This is suspiciously close to what Palantir's CEO said on CNBC"

Not everyone bought the thesis at all.

"Enterprises don't give a rip if models are open. They care about zero data retention… This blog post is suspiciously close to being a restatement of what Alex Karp recently said on CNBC. It's important to remember he's the CEO of Palantir and hardly a neutral observer." — a developer who builds production LLM systems

The same reader pointed at the inconvenient counter-example sitting in plain sight: Meta released Llama, the highest-profile open-weight project in the West, and it hasn't produced anything resembling the market position Werdmuller predicts openness should buy. If open weights guaranteed winning, Llama would already have won.

Simon Willison, creator of Datasette, added the geopolitical context the post was missing, pointing to Ben Thompson's Stratechery analysis of a speech Xi Jinping had just given:

"We should seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing." — Simon Willison, creator of Datasette

That's the frame the article never quite reached: this isn't only a product strategy, it's state industrial policy, and state policy doesn't need to turn a profit to keep running.

Nobody in the 784-comment thread disputes that Chinese labs currently hand out frontier-class weights for free. What's unresolved is whether that's a permanent feature of the market or a subsidy Beijing can switch off. And whether America's AI companies, priced at a premium precisely because nobody is subsidizing them, can survive long enough to find out.