A week ago, the world woke up to the news that a seldom-heard-of Chinese startup had created an AI model—DeepSeek R1—that could rival ChatGPT at a fraction of the cost. US tech stocks plummeted: chipmaker Nvidia closed out last Monday down 17%, and over $1T was knocked off tech stocks overall.
Turning point: Since then, the waves set in motion by the DeepSeek launch have continued to ripple throughout the tech world, even as stocks recovered slightly.
- OpenAI accused DeepSeek of stealing its property to build the algorithm (the irony).
- Trump said DeepSeek was a “positive” development (yes, really) but that it should be a “wake-up call” for American AI firms.
- Some AI execs and experts have accused DeepSeek of underreporting the cost and compute it really took to train the model.
Race to automate: The launch raised a worrying question: could the US be losing its advantage in the AI arms race? And, perhaps even more importantly, does this mean that China could outpace the US in AI-enabled and automated weapons systems, even with tight semiconductor export controls in place?
Janet Egan, a senior fellow with the Technology and National Security Program at the Center for a New American Security, said that it’s not quite time to panic. “If you talk to people who are experts in the field, they aren’t as worried and they’re not as surprised,” Egan said, “US AI leadership is not just about the model itself.”
DeepSeek was in all likelihood trained on Nvidia chips already in China before US export controls were tightened in late 2023 (with maybe a couple more smuggled in), said Egan. That means that we’ve yet to see the effect of these controls and China’s limited access to more advanced chips. “As we progress more and more into the future, we’re going to see a greater capability gap emerging with regards to the advanced data centers and training clusters needed to train these AI models,” she said. “Right now, it looks like China’s closed on the heels, but actually, there’s a lot that’s misread into that development.”
Cheap and cheerful: Former DIU Technical Director for AI Jared Dunnmon says that the primary risk posed by DeepSeek is similar to the one posed by TikTok – millions of Americans have downloaded the app, which means millions of Americans’ data is on servers that could be accessed by the CCP.
DeepSeek is also comparatively cheap and open source – that means that actors (malicious or not) around the world are now able to build and use an AI model that rivals those build by OpenAI, Anthropic, and Meta. Dunnmon says it’s unclear whether there are guardrails in place on the DeepSeek models to keep users from using them to do dangerous things. Overall, he thinks that cheaper AI will mean its use in more things.
“You’ll see the military, just like and the defense and security apparatus, just like the rest of the world, leverage the fact that we can now run workloads that previously were much more expensive for an order of magnitude cheaper,” he said.
Stop the spread: Dunnmon said that chip controls and compute limitations will indeed limit China’s AI advancement, but that doesn’t mean there isn’t anything to worry about. “The better question is, is China going to be able to offer good enough, very inexpensive AI systems to the rest of the world, and take an enormous amount of market share and mind share by doing that?”
If the US wants to curb this spread, he said, some portion of AI investment (read: Stargate) should be focused on building cheaper, “extremely capable” open source models to rival LLMs like R1. “The idea would be that you would continue to have slightly better models continuously, and continue to drive kind of traffic and capital into the Western computing ecosystem,” he said.