1 hour ago
670
Top artificial intelligence researchers in China are now saying something that goes against the upbeat headlines of the past year. Their country probably won’t catch up to the United States anytime soon. The problem comes down to computer chips.
“The truth may be that the gap is actually widening,” Tang Jie, who started the Chinese AI company Zhipu, told a conference in Beijing last weekend. “While we’re doing well in certain areas, we must still acknowledge the challenges and the disparities we face.”
The chip shortage became clear when Nvidia rolled out its new Rubin hardware in January. The company named several American firms as buyers, but left out every Chinese AI developer. American rules stop Nvidia from selling directly to China.
Chinese businesses have started talking about renting computer power from data centers in Southeast Asia and the Middle East to get their hands on Rubin chips, people who know about these talks told WSJ. This follows last year’s efforts to reach chips in Nvidia’s Blackwell line.
These workarounds through other countries are mostly legal. But they mean Chinese AI developers end up with fewer chips and more headaches than their American rivals who have deep pockets.
Industry leaders put odds of catching up at 20% or less
At the same conference, Justin Lin, who runs the development of Alibaba’s AI model called Qwen, was asked whether any Chinese company could jump ahead of OpenAI and Anthropic in the next three to five years. He guessed the odds at 20% or less.
American export controls have scared many Chinese firms away from building cutting-edge AI, which needs huge amounts of computing power. Instead, they focus on putting AI to work in everyday products. Meanwhile, American companies keep buying the newest chips to push forward.
“A massive amount of compute at OpenAI and other American companies is dedicated to next-generation research, whereas we are stretched thin,” Lin said. “Just meeting delivery demands consumes most of our resources.”
UBS analysts figure that China’s biggest internet companies spent about $57 billion on capital projects last year, with much of it going to AI. That amounts to roughly one-tenth of what American companies spent.
Still, nobody is writing off China yet. Developers like DeepSeek have proven they can do a lot with little. Two other AI firms, Zhipu and MiniMax, raised more than $1 billion together through stock offerings in Hong Kong this month. MiniMax shares more than doubled from their starting price.
“Despite a more challenging operating environment, investors continue to price in the possibility of technological catch-up or breakthrough,” said Alyssa Lee, a longtime tech investor now working at an AI startup. “That optimism itself speaks to the level of innovation Chinese companies have demonstrated.”
DeepSeek closes the gap through efficiency
DeepSeek grabbed attention in America a year ago with a strong AI model. Since then, it has shared methods to make AI development more efficient, and some Western researchers have picked them up. This month, DeepSeek put out two papers describing a new training setup that lets developers build bigger models with fewer chips, plus a memory design that helps models run better.
Models from DeepSeek and Alibaba have closed the gap with top American models to just four months, down from seven months on average in recent years, according to Epoch AI. Many leading Chinese models are open source, meaning anyone can download and change them. This raises the profile of Chinese companies while top American models stay closed off.
But DeepSeek has hit bumps. When building its new main model last year, it tried chips from Huawei and other Chinese makers. The results fell short, so it switched to Nvidia chips for some work, people familiar with the project said. The company made progress and plans to release the model in coming weeks.
“The primary bottleneck is chip-manufacturing capacity,” said Yao Shunyu of Tencent at the Beijing event. Yao recently left OpenAI to lead Tencent’s AI efforts.
H200 chip approval unlikely to change the game
Washington’s recent decision to let Nvidia sell its H200 chip to China probably won’t change much, industry insiders said. The H200 sits two generations behind the Rubin line and has become too weak for training top AI models. Companies are still waiting for Beijing’s approval to buy the chips, with Chinese officials drafting rules to regulate purchases, as reported by Cryptopolitan previously.
Nvidia’s China business keeps facing political hurdles. Revenue from China dropped 45% from a year earlier to about $3 billion in the most recent quarter. Yet overall, Nvidia hit $57 billion in third-quarter revenue, up more than 60%, and became the first company worth $5 trillion last fall.
The longer-term worry for Nvidia is that Chinese companies might build open-source software that works on many chip types, not just Nvidia’s. Much of Nvidia’s edge comes from its CUDA software platform, which locks developers into using its chips.
“That’s the real nightmare scenario,” said Seaport analyst Jay Goldberg.
If Chinese developers, forced to use domestic chips, create software tools that gain worldwide adoption, it could punch a hole in Nvidia’s competitive moat.
Nvidia CEO Jensen Huang sees it differently. “As I have long said, China is nanoseconds behind America in AI,” he wrote on X in November. “It’s vital that America wins by racing ahead and winning developers worldwide.”
The smartest crypto minds already read our newsletter. Want in? Join them.
English (US)