Jordan Schneider on ChinaTalk: DeepSeek was made in America, US robotics manufacturing gap, and reading China's truth signal

Summary

Jordan Schneider runs ChinaTalk, a podcast and newsletter on US-China technology and policy aimed at the intelligence and policy community. His recent analysis touches three core topics: DeepSeek's origins, China's robotics manufacturing edge, and the effectiveness of US chip export controls.

DeepSeek's American roots

The US inadvertently created the conditions for DeepSeek's rise. Before 2017, China's best STEM students followed a clear path: top performers attended MIT or Stanford, the next tier went to second-rank US universities, and the remainder studied in China. Three factors collapsed that pipeline. Rising wages and startup opportunity in China made returning home attractive. The first Trump administration created a chilling environment for Asian-Americans. COVID travel restrictions made studying abroad logistically difficult. The result is that DeepSeek's engineering core consists almost entirely of people under 30 who studied at top Chinese universities and would have previously chosen the US.

DeepSeek's annual training run costs roughly $300 million, according to Nathan Lambert's estimate. That figure is well within reach for a well-funded quantitative hedge fund. More structurally, chip constraints forced DeepSeek to optimize for efficiency earlier than US labs did. Western labs are only now hitting that wall as training runs approach $100 billion.

Robotics manufacturing at scale

The US does not build robots at scale. China's edge is not primarily about labor costs but rather 25 years of accumulated manufacturing knowledge, supplier networks, and learning curves that allow Chinese firms to produce drones roughly 15 times cheaper than US competitors. Industrial robotics represents the nearer-term market over the next three years rather than humanoids, but Schneider frames the broader gap as a structural problem with no obvious solution.

Export controls have significant leakage

Ant Group claims it trained a model entirely on domestic Chinese chips. Schneider is skeptical. Huawei stockpiled chips manufactured at TSMC through a shell company before the US discovered the arrangement and stopped it. Claims of all-Chinese chip training likely refer to TSMC chips obtained under false pretenses rather than domestically produced semiconductors from SMIC. The real constraint remains EUV lithography tools, which ASML manufactures and China cannot yet produce or legally import. Recent headlines about a Chinese ASML competitor are largely overblown. Meanwhile, the Bureau of Industry and Security delayed closing memory chip loopholes until the week before the Trump administration took office, giving China enough stockpiled memory to sustain training runs for at least two more years.

The talent advantage at risk

Talent ties all three threads together. Over 75% of top AI research engineers and founders at leading US AI firms were not born in the United States. Historically, 85% of immigrants who earned STEM PhDs have remained in the country, creating the pipeline that won the Cold War and built Silicon Valley. Cutting NSF and NIH funding, pressuring universities, and making the US less welcoming to foreign scientists threatens this compounding advantage. Schneider sees no clean answer to the manufacturing gap but argues the talent magnet is irreplaceable and currently being dismantled.