The U.S. government’s effort to curtail AI development in China through tightened controls of chip exports, including restrictions on the most advanced offerings from Nvidia, didn’t stop DeepSeek from creating its generative AI app efficiently, and at a level that rivals the best the U.S. has to offer from companies such as OpenAI.

While the details on just how DeepSeek did it remain incomplete, and its success doesn’t mean export controls don’t have a place in markets and national security policy, it does show that a focus on stopping the competition can’t keep pace with innovation. Now, the debate is underway over just how far the U.S. government should go in the future in blocking access to U.S. chip technology.

President Biden’s Department of Commerce issued its rules to “regulate the global diffusion” of AI chips and models in the administration’s waning days. The rules already have been heavily criticized by tech companies, including Nvidia, as well as policy experts. A Brookings analysis argues the the AI diffusion rules seek to create “a centrally planned global computing economy.”

“A decade from now, we will look back and recognize how quixotic it was for the U.S. government of the mid-2020s to attempt to limit the ability of people in 150 countries to perform fast multiplications,” wrote John Villasenor, a nonresident senior fellow at Brookings and professor of electrical engineering, law, public policy, and management at UCLA.

In any technology war, questions about what countermove the U.S. should make next inevitably run up against the awareness that any notion of controlling innovation through measures like restricting exports is not guaranteed to work – and may even backfire. Among the risks cited by Brookings: spurring the development of a global AI ecosystem anchored outside the U.S.; pushing more nations into building stronger technology ties with China; and allowing non-U.S. makers of advanced chips to grow global market share at the expense of the U.S. companies behind the original innovations.

“I worry that we will have a knee-jerk response to ratchet up controls heavily, before we fully think through the trade-offs,” said Martin Chorzempa, senior fellow at the Peterson Institute for International Economics.

There is a 120-day comment period that ends on May 15 on the AI diffusion rules, unless Trump reverses or revises the rule before then. While the president has spoken in general about the need to protect the U.S. technological lead, he has not specifically addressed this rule. It’s unknown what stance the current administration will take – expanding, curtailing, or overturning chip export rules already in place.

Vice President JD Vance said at an AI summit in Europe on Monday that the U.S. will safeguard its AI and chips, and block efforts to “weaponize” them.

“Some authoritarian regimes have stolen and used AI to strengthen their military intelligence and surveillance capabilities, capture foreign data and create propaganda to undermine other nations’ national security,” Vance said in an address at France’s AI Action Summit in Paris. “I want to be clear, this administration will block such efforts, full stop,” Vance said. “We will safeguard American AI and chip technologies from theft and misuse, work with our allies and partners to strengthen and extend these protections and close pathways to adversaries attaining AI capabilities that threaten all of our people,” he added.

Trump’s first-day signing of an executive order to “identify and eliminate loopholes in existing export controls,” suggest he could take a hard line. It said the government will “assess and make recommendations regarding how to maintain, obtain, and enhance our Nation’s technological edge and how to identify and eliminate loopholes in existing export controls – especially those that enable the transfer of strategic goods, software, services, and technology to countries to strategic rivals and their proxies.” 

The tech sector was quick to do its outreach to the new administration, with several major CEOs at the inauguration, and Nvidia CEO Jensen Huang meeting with President Trump at the White House in recent weeks for a discussion that included chip restrictions to China.

Trump also called Deepseek a “wake-up call for our industries that we need to be laser-focused on competing to win.”

Particularly relevant to Deepseek in the AI diffusion rules are controls surrounding closed AI model weights, essential to the training process that develops how AI systems think and respond to queries. 

“In part, DeepSeek was able to get around the speed limit imposed on chips allowed for sale to China in 2022, but banned in 2023, when the U.S. realized that the limit imposed was the wrong one,” said Chorzempa.

When the U.S. put controls on China in 2022, Chorzempa explained, they set a specific parameter concerning the speed of communication between chips. It was thought that if you control the power of an individual chip that might not be enough, because if you bring enough less powerful chips together, it’s possible to have supercomputer-like capabilities at a level the U.S. government didn’t want China to obtain. It appears from what DeepSeek described in its R1 paper that the company was able to overcome that speed limit.

“Experts in the technical community in at least early 2023 were pointing out that other restrictions were required to have an effective control as the technology evolved,” Chorzempa said.

In 2023, the U.S. government added additional layers of restriction that made the Nvidia chips DeepSeek says it trained the model on no longer legal for export. Tightened controls could be further strengthened by subsequent initiatives from the Trump administration. 

But through a combination of having a limited number of advanced chips available and innovation spurred on by that limit, DeepSeek was able to build a better, and potentially cheaper, mousetrap. 

“DeepSeeks seems to have optimized heavily with clever software and hardware engineering to sort of neuter the speed limit meant to hold those chips back,” Chorzempa said. 

AI rivals will continue to do more with less

There are other aspects to the evolving AI race which show gaps that are narrowing for other reasons.

“The story is really about the gap being closed between open source and closed source models,” said Alexandra Mousavizadeh, CEO of Evident, an AI consulting firm. “Now the open source models are getting much closer to the capabilities of the closed ones, and we see the price driving down to zero,” Mousavizadeh said.

DeepSeek has already shown that you don’t need maximum computing power, and you can you use open-source as alternative when building a viable LLM, and in fact, according to Mousavizadeh, these factors can be a driver of innovation. 

“We’re seeing that limits forced them to use scientific methods and systems that compress data onto a much smaller pool that uses much less power using mixed expert models,” she said. AI rivals can “do more with less,” Mousavizadeh added.

Continuation of such controls is also likely to push China in further developing its own chips. 

“You can’t really gatekeep,” Mousavizadeh said, noting that there is lots of sharing that occurs in the open source environment, “regardless of governmental policy.”

If DeepSeek’s success leads to export controls on advanced chips intended to slow Chinese AI efforts that become even stricter, it should also be clear they are no silver bullet. “They’re not a way to duck the competition between the US and China,” wrote Dario Amodei, CEO of gen AI startup Anthropic, in a blog post last week. “In the end, AI companies in the US and other democracies must have better models than those in China if we want to prevail. But we shouldn’t hand the Chinese Communist Party technological advantages when we don’t have to.”

His issue isn’t with the AI researchers in China, but the government to which they are ultimately beholden. “In interviews they’ve done, they seem like smart, curious researchers who just want to make useful technology,” Amodei wrote about DeepSeek. “But they’re beholden to an authoritarian government that has committed human rights violations, has behaved aggressively on the world stage, and will be far more unfettered in these actions if they’re able to match the US in AI.”

To be sure, there are many reasons to be wary of doing anything to contribute to China’s AI advances and successes like DeepSeek, from national security concerns about data sharing with the Chinese government, to ongoing hacking risks, to Chinese AI apps becoming popular enough to be used by Chinese intelligence to learn about Americans and American industries, and to sow division among the public.

Palantir Technologies CEO Alex Karp told CNBC’s Sara Eisen in a recent interview that “we have to run harder, run faster, have an all-country effort.”

“The second-mover can move very quickly, especially if we’ve already done the innovation,” Karp said, describing DeepSeek as derivative of U.S. models with “improvements at the margins.”

He expects a “huge policy discussion” to make sure innovations are not exported, but Karp added that in the end, “the real advantage goes to the first mover as long as the first mover is running hard. … We have the lead, we have to focus on making sure we keep it. Our adversaries are gonna copy anything they can.” 

Palantir’s rise — its shares soared 340% last year to lead the S&P 500 — didn’t come by trying to stop others, and in that there may be a lesson. “We don’t focus on the competition,” Karp said. “We focus on how do we execute.”