US Hurts Itself by Blocking Nvidia's China Exports
By Tom Kagy | 05 Feb, 2025

Feeding China's addiction to Nvidia GPUs is the smart way to boost US AI progress while weakening China's.

The ultimate business model is to cultivate addiction, then feed it.

Nvidia has succeeded in doing just that.  It has hooked global tech firms, big and small, on its GPUs as the path to achieving their AI dreams. The US firm derives 56% of its revenues from outside the US, with 17% coming from China, making it an outsize component of US export prowess.

In 2024 Nvidia revenues jumped 228% to over $60 billion, and it's on track to shoot past $100 billion in 2025.  Its $3+ trillion valuation keeps it in a tight 3-way horserace with Apple and Microsoft as the world's most valuable company.  

So where's the beef?  Short-sighted protectionists and jingoists are pushing to cut Nvidia's China exports as a way to help the US preserve its AI lead.  They point to DeepSeek AI models passing those from the likes of OpenAI and Anthropic in cost-efficiency as a cautionary tale about the danger of selling the latest tech to rival nations.

That myopic view misses the big picture.  In reality Nvidia's exports are helping to preserve US AI leadership while prolonging China's position as a follower.  It was the Biden administration's misguided restrictions on Nvidia's GPU exports that actually spurred Chinese firms to close the AI technology gap.

DeepSeek's recent efficiency edge over US rivals resulted from developing clever training and inference shortcuts at the application (software) level to achieve order-of-magnitude savings of processing and memory requirements for training and running its recent AI models.  

It achieved the feat using last-generation Nvidia A100 chips (not the current-gen H100 used by top US-based AI models), suggesting that advanced Nvidia hardware (GPU) wasn't the primary factor in DeepSeek's success.  The power differential between the Nvidia A100 and the H100 is perhaps 40% at most, and is more than overcome by efficiency gains from DeepSeek's software tweaks.  This suggests that even if DeepSeek had no access to Nvidia GPUs, its founder and chief engineer Liang Wenfeng would likely have found software shortcuts to overcome the hardware deficiencies of Chinese-made GPUs.  

Media coverage following DeepSeek's January 2025 leap up the Chatbot Arena AI leaderboard has created the impression that it's unique in having improved on US AI performance.  In November 2024 an AI model called Hunyuan-Large built by Chinese social media giant Tencent scored higher on several benchmarks than the most powerful version of Meta’s Llama 3.1.  What made this feat notable is that Hunyuan-Large was trained using Nvidia A20 GPUs which are even less powerful than the A100 GPUs used by DeepSeek, and certainly far feebler than the H100 state-of-the-art GPU used by Meta.

“They're clearly getting much better use out of the hardware because of better software,” said Ritwik Gupta who authored research by the Berkeley Risk and Security Lab which advises the Department of Defense’s Defense Innovation Unit. 

Software improvements that can squeeze leading-edge performance out of GPUs, American or Chinese, is the wildcard in the AI competition.  Like Bill Gates at the dawn of the PC era, Liang Wenfeng is a coding wiz with the vision and drive of a tycoon.  He's not alone.  Rather he's part of a burgeoning generation of Chinese AI researchers set to make its mark in the coming decade.  Their hands-on effort as AI pioneers in a Chinese market denied access to Nvidia chips will likely have a multiplier effect on Chinese AI innovations.  

Motivation provided by such underdog ambition is especially significant because China outproduces the US in AI talent, with nearly half the world's top AI researchers graduating from Chinese universities compared with only about 18% from the US, according to the MacroPolo think tank of the Paulson Institute.  Three years earlier China's share had been only a third while the US share was the same 18%. 

During the last decade the US AI sector was strengthened by an influx of talented Chinese students who studied at US universities, then stayed on.  Today 38% of US AI researchers come from China while 37% are Americans.  

Now, however, China's contribution to the US talent pool is declining as a majority of US-educated Chinese AI talent is returning home.  Three years ago 59% of the world's top AI talent was based in the US.  Today it's only 42%, reflecting the higher rate of return by Chinese AI PhDs.

“The data shows just how critical Chinese-born researchers are to the United States for A.I. competitiveness,” said Matt Sheehan, an expert on Chinese AI at the Carnegie Endowment for International Peace.

“We’re the world leader in A.I. because we continue to attract and retain talent from all over the world, but especially China,” he told Time.

The abundance of Chinese AI talent isn't an accident but the product of China's push begun in 2018 to add over 2,000 undergraduate A.I. programs, including over 300 at its elite universities, according to Damien Ma, managing director of MacroPolo.

As was the case with conventional computing, AI advances are driven by thousands of energetic young developers uncovering software strategies to squeeze more work out of whatever hardware happens to be available.  This large, fast-growing army of Chinese AI developers, if deprived of normal market access to Nvidia GPUs, will be even more motivated to devise ways to overcome advantages conferred by advanced Nvidia GPUs. 

Meanwhile US developers supplied with an abundance of expensive cutting-edge GPUs will have neither the freedom nor the incentive to innovate.  Instead they will be constrained to follow well-beaten application development pathways.  The end result: a leveling of the playing field, or even a tilting in favor of the underdog, as we saw with DeepSeek and Tencent.

The foolishness of restricting Nvidia chip exports to China is even more glaring in light of rapid developments of processor technologies that will obsolete conventional GPUs in a decade or less.  As western chipmakers run up against physical limitations in etching logic gates on silicon wafers, new technologies are removing prior limits on processor performance.  What will remain constant as the key driver of computing advances is the software developed to apply processing power to useful tasks.

On the five-year horizon are energy-efficient photonic chips that use light instead of electricity to speed up calculations while cutting energy use.  For example, in late August 2024 Chinese researchers unveiled Taichi-II, the second generation version of an optical AI chip released just four months earlier.  In that brief span the Taichi team achieved a 40% improvement in data processing accuracy and a million-fold increase in energy efficiency, using a single watt of electricity needed to perform 160 trillion operations.  

Several Taichi-II chiplets were linked together to simulate a network of 14 million artificial neurons.  That's a far cry from the 86 billion neurons of a human brain but about 30 times the processing power of a conventional FPGA board for conventional chip development.  With any luck Taichi, or one of several dozen other photonics chipmakers, could obsolete today's GPUs.  The Taichi-II chip is considered promising enough to be a leading candidate on which to build AI systems capable of achieving AGI — artificial general intelligence, or human-level intelligence, the long-term goal of DeepSeek's Liang and many other leading Chinese AI researchers.  

And only a handful of years later we are likely to see practical quantum computers.  Currently the most advanced quantum processor unit (QPU) is Google's Willow chip released in early December 2024.  By exploiting the particle-wave duality of quantum mechanics to manipulate single electrons QPUs can essentially perform nearly unlimited layers of calculations simultaneously.  

With only 105 qubits (building blocks of quantum processors) the Willow superconducting chip can perform in under five minutes benchmark RCS (random circuit sampling) computations that would take Frontier, the world's fastest supercomputer, 10 septillion years (10 to the 25th power).  In short, quantum computers will be able to solve in minutes problems that remain beyond the reach of conventional computers.

Even Nvidia is preparing for the advent of practical quantum computing, starting with its CUDA-QX collection of libraries and tools to help quantum computing researchers and developers use existing GPU programming to develop quantum computing applications.

US national security policymakers must remember the adage, "There's more than one way to skin a cat," when considering the likely impact of export controls in tech, the area that always attracts the fiercest innovators.  Export controls will merely fire up an entire generation of young Chinese AI developers while keeping US developers complacent and on the safe, easy road to obsolescence. 

So let Nvidia and other US tech suppliers book tens of billions in GPU exports to China while they can.  About a third of their profits will be plowed back into R&D, helping to feed the next generation of US innovation while leeching resources and motivation from China's AI development.