An upstart Chinese AI lab is challenging the Silicon Valley conventional wisdom on what it takes to build a leading AI system.
DeepSeek made headlines in December with an open-source model that performed comparably to similar US AI models, at what the lab said was a fraction of their development cost and time. Then, last week, DeepSeek released a reasoning model, R1, that purportedly performs on par with OpenAI’s landmark o1 reasoning model—at a much lower price.
It’s the latest example of how some Chinese tech companies are seemingly outmaneuvering restrictive chip export rules to produce AI systems that could potentially erode US dominance. Alibaba and ByteDance each released their own models in recent months.
Much of the focus of this competition is around AI reasoning, the ability of generative systems to use logic and problem-solving to elevate their capabilities beyond question-and-answer machines and onto more complex tasks. OpenAI made a breakthrough on this front last September, when it first unveiled its o1 model.
Copychat? There’s also some question of whether DeepSeek is piggybacking on OpenAI’s progress by training on ChatGPT outputs—which would amount to a violation of OpenAI’s terms of service. While the Chinese lab hasn’t been open about its training data, TechCrunch reported in December, the DeepSeek V3 model has been caught referring to itself as “ChatGPT,” a sign it could’ve trained on its competitor’s outputs, either intentionally or as part of a large public dataset.
Costs remain: S&P Global tech lead David Tsui said he wouldn’t necessarily expect AI spending to decrease if US tech giants decided to replicate some of DeepSeek’s efficiency measures, which are detailed in an accompanying research paper.
“I’m sure they’re looking at how to improve their models, including spending more efficiently,” Tsui said. “But it doesn’t mean that just because they can spend more efficiently, they’ll spend less. I think they’ll spend less for what they thought would be the output, but they’re going to continue to spend a lot of money on AI so that they can get to their ‘destination’ sooner, and then continue to monetize and commercialize AI.”
Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
Ongoing restrictions: DeepSeek is also building its models despite US rules that have blocked Chinese access to Nvidia’s most sophisticated AI chips. Tsui said this success will likely lead the US to maintain and even “broaden” these trade restrictions as it seeks to cement its dominance.
“[That might mean], number one, enforcement of the current existing restrictions, and number two, maybe broadening the scope of what can be shipped—not just leading-edge, but maybe leading-edge plus additional, even lower-spec-type chips,” Tsui said.
While US-based Nvidia still has a wide lead in chip performance, IDC research director Brandon Hoff also said China is building its own chips that will eventually help to close the gap.
“Nvidia and the US [will be] dominant for quite some time, but China will catch up,” Hoff said. “They’ll start having maybe not as good performance, not as big of models, but they will be able to catch up with homegrown GPUs and homegrown CPUs.”
Open vs. closed: DeepSeek—as well as Alibaba’s recent Qwen reasoning model—have also notably performed competitively while releasing under open-source licenses, an issue that has divided US tech companies.
“I do think it’s a positive development that you’re seeing highly performing open source models,” Mozilla Foundation President Mark Surman said. “America absolutely needs to invest more in its own open-source AI leaders.”