The BiPolar Chips Wars

To a company like Nvidia with a nearly $3 trillion valuation, even an unexpected cost of a few billion dollars might be somewhat small in scale, but the U.S. chipmaker’s announcement that new export controls would likely cost it up to $5.5. billion caused a 6% drop in its stock value the very next day. Moreover, the news caused more than just a ripple of concern throughout the global semiconductor chip industry.

Nvidia’s financial hit comes from a new Trump administration rule requiring the company to acquire a special license in order to sell its H20 chips in China, adding another hurdle in accessing one of the world’s largest tech markets and the U. S.’s foremost competitor in what has become a white-hot race for artificial intelligence.

The new license requirement is intended to prevent the chips from being “used in, or diverted to, a supercomputer in China,” according to Nvidia’s filing. It’s just the latest attempt by the Americans to slow China’s AI development and preserve its advantage.

The perennial question hanging over U.S. restrictions on Chinese tech over the past eight years has been how well they are actually working. Significant milestones in China this year—such as the launch of the advanced AI model DeepSeek-R1 and advances in semiconductor chips from tech giant Huawei—have reignited that contentious debate. Some experts and policymakers are now questioning whether or not it’s too late to keep China from catching up to U.S. (much less surpass it) AI technology, and whether the United States should instead pursue a more collaborative approach with Beijing on AI development and regulation is being hotly debated.

Nvidia created the H20 as a workaround for U.S. government restrictions on another one of its chips—the H800, which the Biden administration banned the company from selling to China in October 2023. The H800 had also been created in response to earlier restrictions by the same administration on Nvidia’s sales.

Essentially, the first round of chip controls came and they set this bar, and then Nvidia said: ‘OK, we’ll build the fanciest thing we can that’s allowed, and sell a bunch of them, because we’ve just been told we could sell those’—and then a bunch of people in Washington got angry, as if this was a sort of unpatriotic thing to do. Nvidia’s orientation is pretty consistent with that of most businesses— pursue the profit prerogative wherever it takes you: in their case it’s build increasingly advanced chips and sell as many as they can to whoever they can get them to.

Nvidia’s graphics processing units (GPUs)—a type of semiconductor circuit that the company invented in 1999—have recently exploded in popularity because of their critical role in training artificial intelligence models such as OpenAI’s ChatGPT and its competitors. They have also made the company’s products a prime target of export controls by successive U.S. administrations intent on curbing China’s access to advanced technology.

In 2019 Trump 1.0 began that effort, restricting Huawei from access to semiconductor chips and other U.S. technology by placing the company (and other Chinese firms) on the Commerce Department’s so-called “entity” (or ‘hit’) list. Joe Biden’s successive administration broadened the fight in 2022, imposing export restrictions on chips and chipmaking technology to China and continuing to periodically expand those restrictions all the way up until the end his term in 2021 when the old boss (Trump) became the new boss.

One of the big uncertainties hanging over Trump’s return to the White House was how his past hawkishness on China pre-Biden would manifest itself post-Biden. While there have been some reversals in positioning (see: TikTok) and some continuations (see: trade war), his second-term strategy to curb China’s semiconductor industry point to more of the same.

But this time around, Trump is facing a slew of recent reminders from China of its continued—and, to Washington, alarming—progress.

None of those have been sharper than DeepSeek, whose R1 language model—released in late January—showcased capabilities rivalling those of U.S. leader OpenAI but at a fraction of the cost and computing power. DeepSeek’s debut sent shock waves through Washington, though experts still debate the extent to which it actually constituted a dreaded “Sputnik moment” for American AI.

More pointedly, DeepSeek’s debut raised questions for U.S. policymakers about the effectiveness of export controls. That’s because its success actually came on the back of legally acquired American chips—the company trained its model largely using Nvidia’s H800 and H20 GPUs. DeepSeek was prescient enough to stockpile enough H800s before Biden clamped down on the chips in 2023. To everywhere who’s ever played whack-a-mole, the U.S. government was swinging its regulatory hammer a beat too late.

At the same time, Chinese AI companies’ inability to access the most cutting-edge U.S. chips may have paradoxically turbocharged their innovation by forcing them to be more resourceful, as was the case with DeepSeek.

And DeepSeek isn’t alone. China’s top AI models are rapidly closing in on their U.S. equivalents despite the restrictions. In March, Kai-Fu Lee, the Beijing-based CEO of investment firm 01.AI and a leading AI expert, told Reuters that Chinese AI companies now lag behind U.S. firms by only three months in core technologies. Tech-wise that margin is less than paper thin and far from reassuring to American interests eager to perpetuate the technology edge.

Chinese tech giants have also been racing to pump out their own versions of advanced chips. Huawei is preparing soon to launch its new Ascend 910c AI chip, which Chinese companies are expected to use to replace H20s. The company is also testing the 910d, which it hopes will supersede the power of Nvidia’s H100—one of the previously banned chips—for model training. Taken together, China’s advancements haven’t exactly been a glowing testament to U.S. export restrictions. Yet many experts argue that the policy still has legs.

Chinese AI companies have continued to do whatever they can to buy U.S. chips, which proves their superior quality and performance. Before Trump brought the hammer down on H20 chips, Chinese companies had placed orders for 1.3 million of them, totaling more than $16 billion. The H20s were highly sought after because they are specifically designed for inference—the actual running of a trained model, which is becoming an increasing focus of AI innovation as AI is used more widely. Depriving Chinese companies of these chips could present a real stumbling block. And for the next round of AI advancements, some experts argue that the sheer volume of advanced chips is still a difference maker.

China finds a lot of ways to come up with innovative developments that maybe are less compute intensive, but they still haven’t quite worked around the fact that the U.S. is still the lead in compute, and the Americans still have access to the most chips and the most computing resources. That scale still really matters but it’s not an absolute advantage, and it’s by no means a permanent advantage.

Even though the U.S. policy of restricting China likely has an expiration date, proponents argue that it is worth pursuing as long as possible for one reason above all else: the military implications. Both the Trump and Biden administrations have pointed to the potential for AI to confer new military advantages to China as the primary driver of U.S. policy. In short, a national security component that can only be ignored at peril.

There is still much debate about how significantly AI could enhance China’s military capabilities. Hard facts are few as the opacity of the People’s Liberation Army (PLA) has made it challenging for researchers to assess the PRC’s progress and plans. Strategists’ concerns range from the more prosaic—including AI models being applied to increase supply-chain efficiency for ammunition and other battlefield resources—to the more nightmarish, such as AI being used to control vast swarms of drones in an invasion of Taiwan (shades of the ‘Terminator’ movies).

Due to China’s military-civil fusion policy, which calls for harnessing cutting-edge commercial technologies to strengthen its military, pro-U.S. controls supporters say that it is necessary for the United States to continue targeting the flow of advanced chips to China in general. But another camp argues that the costs are too high—especially considering that the policy will likely only buy the U.S. a limited amount of time before the Chinese catch up.

The most visible cost is the blow to corporations’ bottom lines. Nvidia, with a market capitalization rivalling that of the United Kingdom’s GDP, certainly isn’t the most sympathetic victim, and some AI scholars have argued that soaring demand for the company’s chips in Western nations means that it can easily compensate for any lost revenue from the China market.

But others warn that Washington’s restrictions will eventually cost U.S. companies, which will be increasingly cut out of the Chinese market as its ecosystem becomes more mature, technologically advanced and independent. If U.S. companies do indeed see an overall hit to their revenue, that could reduce their research and development budgets and cause them to lose ground to Chinese competitors over time.

Of greatest concern to those who question the logic of export restrictions is that AI safety and security conversations have been displaced by the all-out effort to win the U.S.-China race. During the Biden administration, there were efforts to simultaneously curb China and collaborate on safety standards, with some success. In the final months of the administration, both sides agreed to maintain human control over nuclear weapons. For critics, the U.S. focus on restrictions is undermining further safety talks. Even for those who support continued restrictions, Washington’s lack of plan for a future of AI parity with China is a concern.

For now, the Trump team has indicated that those discussions are not a priority. U.S. Vice President J.D. Vance said that “The AI future is not going to be won by handwringing about safety,” in a speech at the February AI Action Summit in Paris.

The next big test for American ambitions to outpace China will be the extent to which the U.S. can bring the rest of the world—particularly traditional allies and partners—into the fold. The imminent challenge for Trump on that front is finalizing another Biden holdover as the previous administration pushed its “Framework for Artificial Intelligence Diffusion” out the door a week before Trump’s return to the White House.

The framework, more commonly known as the AI diffusion rule, divides countries into three tiers of access to advanced U.S. AI technology. Top of the tiers feature 18 close U.S. allies who enjoy near-unrestricted access (including Canada, Germany, and Taiwan) while the bottom tier includes roughly two dozen arms-embargoed adversaries (i.e. China, Russia, North Korea, and Iran) where chips exports are totally banned. Most other (more than 150) countries are in the second tier, which will be subject to strict licensing requirements for advanced chips and software parameters critical to developing AI models and data centers.

The Biden administration included a 120-day public comment period that kicked the rule’s implementation down the road into the hands of the Trump administration, setting May 15th as the deadline for countries and companies—many of which, including Nvidia, Microsoft, the United Arab Emirates, and the Czech Republic, have lobbied against it—to comply. Whatever final shape that rule takes under Trump will be seen as a bellwether for U.S. AI strategy going forward.

Trump 2.0 has thus far provided few windows into its thinking. The Trumpian instinct to try to simplify policy (the administration is reportedly considering scrapping the country tiers altogether) as much as possible could run counter to the president’s broader China containment strategy.

The choice is binary. One can try to contain China or help American industry. There does not seem to be any easy middle ground between those two solitudes.