°FAI: Karen Hao Says the US-China AI Race Is Fake.

Karen Hao wrote Empire of AI, and she has been watching OpenAI longer than almost anyone. She spent three days inside the company in 2019 and came away unconvinced. I have followed her work since I started the Fahrenheit AI series, mostly because she reports on the parts nobody else covers: where the data comes from and who does the labor.

The claim of hers that stuck with me came from a talk she gave at CUNY last October. She said the US-China AI arms race is “completely fake.” In her words, “it is a narrative that has specifically been leveraged by Silicon Valley extremely effectively to ward off regulation.” Her evidence is the last ten years. The industry told Washington not to regulate them, to block China’s access to chips instead, and America would leave China in the dust. She says we got the opposite: “the gap between US and China AI capabilities has never been narrower.”

DeepSeek

The part that matters for a print shop is how that gap closed. The chip ban meant Chinese labs had to get clever , that is where DeepSeek came from, and then Qwen from Alibaba. Both are open source, and both are free. Hao told Mishal Husain on Bloomberg that DeepSeek proves “it is in fact possible to produce the same exact kind of AI with significantly less resources.” At CUNY she mentioned that grad students at UVA now do all their research on Qwen “because it is just the best model out there that is free and small.”

I have been running small models locally for a while, and I wrote about it in Two AI Tools That Changed My Workflow and again in the RAG post. A separation tool or a setup-time estimator does not need a frontier model. What it needs is a model that runs on the shop computer and never sends your customer’s art anywhere.

Open vs Closed

Hao’s point is that the real dividing line is open versus closed, not America versus China. “I feel comfortable if that open source model is coming from China or if it’s coming from the US because it’s open. It’s scrutinizable.” John Borthwick, who invests in open-source AI, agreed, and added that Google, which wrote the original transformer paper, has mostly stopped publishing.

The background: in 2017 a team of Google researchers published “Attention Is All You Need,” the paper that introduced the transformer architecture. Every large language model since, including GPT (the T stands for transformer), Claude, Gemini, DeepSeek, and Qwen, is built on it. Google gave that away openly, the way research had always worked. Borthwick’s point is that the company that made the whole field possible by publishing has now gone quiet, because there’s too much money at stake to keep handing competitors your best ideas. That’s the irony he’s pointing at, and it’s why it supports Hao’s argument that the open-source momentum has shifted to China.

“Mostly stopped publishing” is Borthwick’s characterization, and it’s an overstatement if read literally. Google still publishes papers and still releases some open-weight models (its Gemma family, for instance). What’s changed is that it no longer publishes the research behind its frontier models the way it published the transformer paper.

Contradictions

Here is where I have to be honest about my own contradiction. I owe my health to one of the closed everything machines she is criticizing. I went through a health crisis that modern medicine could not get me through. The AI model did not discriminate. It read what I gave it, took it seriously, and helped me find a path when the doctors would not. I cannot disparage that. So I hold both ideas at once. Could I have gotten through it with an open-source model? I don’t know. I could have tried. US-China AI arms race is another scare tactic in something we are all struggling to understand.