The US government has for years actively tried to curb China’s access to semiconductor chips, a key component in generative AI models. Instead, those export limits may have fueled the innovation that led to DeepSeek‘s R1 — a large language model that’s disrupting the domestic AI industry and the booming economy built around it.
Brian Colello, a tech analyst for Morningstar, said the quote “constraints lead to creativity” came to mind.
“These Chinese models were processor-constrained, so it led to some creative techniques in training, and the DeepSeek model has come out with better-than-expected performance given the processors that it’s been trained on,” he told Business Insider.
DeepSeek disruption
DeepSeek, a China-based AI startup, dropped the app version of its R1 model last week. The model appeared to rival those from major US tech companies, like Meta, OpenAI, and Google — but at a much lower cost.
DeepSeek said it spent nearly $6 million in computing power to train its new system, a fraction of what US tech companies have spent on their models.
DeepSeek said its models were trained with fewer and less powerful semiconductor chips than their competitors typically use.
Since 2022, US sanctions have made it illegal for manufacturing leader Nvidia to sell some of its chips to China, including its most advanced chips. The sanctions aimed to limit China’s advancements in AI and military technology.
“Sanctions forced DeepSeek to use H800s, which were less powerful than H100s,” Patrick Moorhead, the CEO of Moor Insights and Strategy, told BI of the Nvidia chips DeepSeek has used.
“In a roundabout way, sanctions initiated in the Biden administration motivated DeepSeek to be more creative in how it trained and ran models,” he added. “No one should be surprised, as ‘necessity is the mother of invention.’”
Murky training and computing costs
Some experts and analysts who spoke to BI expressed skepticism over DeepSeek’s claims about the cost of the models and the number and type of chips they were built on. However, it remains unclear exactly what semiconductors were used to train and deploy DeepSeek.
Still, some analysts said the startup showed that it’s possible to do more with less when it comes to AI.
Deutsche Bank analysts Adrian Cox and Galina Pozdnyakova wrote of DeepSeek in a research note published Monday: “They’ve had to squeeze more value out of their software and methods such as chain-of-thought reasoning and using several models at once, instead of just throwing more computing power at the problem.”
Chris Miller, author of the 2022 book “Chip War,” told BI the DeepSeek models are impressive but that costs in AI have come down dramatically since 2023, so he did not find the company’s latest paper especially surprising.
He also said the idea that DeepSeek was working on a “shoestring budget” was not true, saying the company used a “very narrow definition of training costs.” Miller said it’s “pretty clear that the training cost is an order of magnitude higher” than DeepSeek suggested.
Ineffective chip restrictions
Alexandr Wang, the CEO of Scale AI, said during a January 23 CNBC interview that DeepSeek had 50,000 H800s, which Miller said would be a “substantial number. ” While that number is still much less than what US firms have, Miller said, it’s likely a lot more than US export officials would’ve wanted a single Chinese firm to accumulate.
Zongyuan Zoe Liu, a senior fellow for China studies at the Council on Foreign Relations, told BI that the developments at DeepSeek suggest AI development in China “seems to be at least on par with the US.”
However, she said, “we’re still at the beginning of the race” for AI dominance.
“It certainly serves as a good reminder for American policymakers that technology restriction may not work, depending upon what the end goal is,” Liu said.
Several experts said they thought the latest developments with DeepSeek could lead to even more semiconductor sanctions on China but would not necessarily stop further innovation.
“The US could put sanctions on China all day long,” Colello said, “but there’s always the threat: What if China comes up with some breakthrough anyway?”
The post How the US may have unintentionally helped create an AI monster in China appeared first on Business Insider.
