Nvidia’s Jensen Huang recently said China is “not behind” the US in artificial intelligence. Other US tech giants might disagree.
At Google I/O, the search giant unveiled its latest Gemini AI model, claiming it has the world’s fastest AI, with a token output speed ten times that of China’s DeepSeek.
Token output speed, roughly how fast a model generates responses, is just one of several benchmarks used to evaluate AI capabilities.
Other major US AI players are also boasting about the superiority of their latest models, from GPT developer OpenAI to Elon Musk’s xAI to Anthropic, the startup behind Claude.
In China, meanwhile, Alibaba says that its latest Qwen 3 model series, released in April, requires significantly less computing power than its peers while outperforming several major rivals in benchmark tests ranging from mathematical reasoning to coding proficiency.
The irony, however, is that while the AI race is hotter than ever, it may no longer matter so much who has the most advanced model.
Arthur Lai, head of Asia technology research at Macquarie, says the next battleground for AI is in applications.
“Technologies like DeepSeek may enable distributed training, model sharding, or decentralized fine-tuning, which can reduce per-company cost and compute needs,” Lai told Nikkei Asia, adding that demand for computing power is continuing to rise due to the broader adoption and growing number of AI models.
“The high potential growth area will be agentic AI,” he said.
Lai’s sentiment was echoed by Jason Corso, a professor specializing in computer vision and AI at the University of Michigan. Corso said it is difficult to tell which company or country is actually “ahead,” because there is no accepted benchmark for evaluating the impact those AI models have made in the real world.
“The US may be ahead in overall model capacity and throughput, but this doesn’t matter much if the model capacity and throughput are not being transformed into actual value,” he said.
While Google and others could claim their models boast the “fastest intelligence” or can solve questions in the International Mathematical Olympiad, those achievements only “measure one small facet of intelligence, and clearly not direct utility to the end user,” he said. “Sure, it looks like things are getting better and improving, but is the speed of token generation seriously meaningful from a utility perspective?”
DeepSeek sent shock waves across the AI sector when its claims of a breakthrough in lower-cost AI development led to questions about the billions of dollars being spent to buy chips, build data centers, and train foundational models without yielding meaningful returns.
“Right now, I think we have a little bit of a capability overhang with reasoning,” Microsoft CTO Kevin Scott said at the company’s annual developer conference in May. “The models are more powerful than what we collectively are using them for at the moment.”
One answer to the question of what to do with that “capacity overhang” could be AI agents.
As the market recalculates how many Nvidia chips and data centers the world actually needs, companies in both China and the US are hoping that this emerging segment can convince investors and users that artificial intelligence is more than just a buzzword.
In simple terms, an AI agent is a digital assistant, a system that humans can delegate their tasks to without having to be involved in the execution.
The rise of reasoning AI models, such as DeepSeek’s R1 and GPT-4o, paved the way for AI agents by enabling AI systems to think more like a human. Reasoning models are a subset of so-called thinking models, which aim to replicate human thought more broadly.
Gartner forecasts that by 2028, 33% of enterprise software applications will feature so-called agentic AI, a significant rise from less than 1% in 2024. Additionally, AI agents are expected to handle at least 15% of daily workplace decisions on their own.
In the past few months, established tech giants like Microsoft, Google, Alibaba, Tencent, and ByteDance, as well as a number of startups, have all shown off their latest AI models and applications. Perhaps more importantly, they have also laid out their plans for using those models to develop truly useful AI assistants.
At Google I/O, the company said it has started integrating its Gemini AI assistant into its software and hardware ecosystem, including new AI glasses produced in partnership with Samsung. The aim, it said, is to create a proactive, personalized AI agent that can see, understand, and perform tasks for users.
“Our recent updates to Gemini are critical steps towards unlocking our vision for a universal AI Assistant. This is our ultimate goal for the Gemini app,” a Google spokesperson told Nikkei Asia.
Similarly, Microsoft laid out its plan for an “open agentic web,” a network of AI agents that can talk to each other through Microsoft’s ecosystem and make decisions and perform tasks and functions for users and organizations.
In China, Manus has made waves by claiming it has the world’s “first general AI agent” that can do everything from screening resumes to researching real estate and analyzing stocks. The application was even hailed by Chinese media as the “next DeepSeek” breakthrough.
Other Chinese startups are also trying to make a splash with AI agents. Startup Fellou, which is launching its AI browser in both the US and Japan, claims that its AI agent can not only handle open data from the public domain but also access local files or perform tasks on social media platforms, such as leaving a comment on the latest post of a particular account.
“We are aiming for the US market first because Americans are more willing to pay,” Xie Yang, founder of Fellou, told Nikkei Asia, adding that his product is based on platforms including Claude, ChatGPT, and DeepSeek.
“The US is paving the way for us because they have strong research capabilities and they will undoubtedly lead in the AI field in the short to medium term,” Xie added. “However, I believe China’s AI applications might be stronger because there are more application developers in China, and overall, they are less inclined toward research.”
Chinese tech giants also see AI agents as the next key battleground, sparking a flurry of releases in recent months.
Tencent recently upgraded its AI agent development platform to better enable companies to build their own smart AI agents, which can reportedly handle tasks from coding to data analysis without manual input. For its own AI agents, Tencent plans to transform them from quick response tools into tools that can ultimately interact with other applications and external APIs to offer comprehensive user support.
Alibaba launched its all-in-one AI assistant Quark in March. Its “super box” interface can perform a wide range of tasks, including deep search, content creation, image generation, and coding.
In April, ByteDance launched its latest thinking model Seed-Thinking-v1.5, which the company claims outperforms DeepSeek’s R1 in generic tasks and is also more cost competitive. Its AI-powered productivity platform Coze, which is only available in China, meanwhile, can automatically analyze questions to generate output like web pages, power point presentations or text documents.
RELATED ARTICLE
Not everyone is convinced that the era of AI agent is near, or that there are meaningful ways to rank where US and Chinese companies stand in the AI race.
“The posts I saw about agentic AI are primarily marketing hype,” said the University of Michigan’s Corso. “There has been little evidence that any new advances have been made, partially due to the lack of reasonable benchmarks available that map to meaningful user outcomes and the lack of actual understanding of the user needs.”
And while the US-China tech competition continues to heat up, some still believe AI advancement should be a global effort.
“We’re not in a world of zero-sum games,” Yossi Matias, head of Google Research told Nikkei Asia in an interview.
Matias declined to comment on how Google’s AI efforts compared to those happening in China.
“If we see innovation coming from various places, it’ll actually lead to more advancements everywhere. So I’m really excited about what we see from everywhere,” he said.
This article first appeared on Nikkei Asia. It has been republished here as part of 36Kr’s ongoing partnership with Nikkei.