Another 'DeepSeek moment'? What China's Kimi K3 means for the global AI industry

Visitors at the booth for Moonshot's Kimi K3 during World AI Conference in Shanghai, July 17. AP-Yonhap

The launch of Moonshot AI's Kimi K3 has revived an intense debate that has raged in Silicon Valley ever since DeepSeek's shock breakthrough last year: whether China can overcome its limited access to advanced chips to match the performance of the United States' frontier artificial intelligence models.

Trillions of U.S. dollars might rest on the answer, as some see Chinese developers' success in increasing performance through architectural innovation as weakening the rationale for America's vast spending on AI infrastructure.

The arrival of K3 has sent shock waves across the global AI industry. The 2.8-trillion-parameter open-weight model performed close to the latest frontier systems from OpenAI and Anthropic, fuelling concerns in Silicon Valley that China had closed its model-development gap with the U.S. to weeks rather than months.

For some analysts, the launch is about more than a shift in the benchmark contest between U.S. and Chinese models. It offers a test of how China's advances in AI software could affect an industry built around US dominance in frontier models - and the computing hardware that powers them.

"For much of the past three years, the global artificial intelligence story has been framed as an American one. OpenAI, Google and Anthropic have dominated headlines, while investors have poured money into the companies supplying the chips, memory and data centres needed to power the AI boom," said Sunil Tirumalai, head of emerging markets and Asia equity strategy at UBS, in a research note.

"But a new question is beginning to emerge: what happens if Chinese AI models become much better - and much cheaper?"

Several investment banks have described K3 as evidence that Chinese laboratories are moving beyond their earlier reputation for producing cheaper but less capable alternatives to Silicon Valley's top systems.

Morgan Stanley said K3 represented an "all-round catch-up" in model scale, performance and pricing. Goldman Sachs described China's progression as moving from the cost efficiency demonstrated by DeepSeek to stronger intelligence and, with K3, a greater ability to charge for frontier-level performance. Bernstein said the model showed that China's leading laboratories could continue to keep pace with the U.S. frontier.

Meanwhile, Chinese AI models are increasingly finding users outside China. Their appeal is not that they are the smartest models in the world. Rather, they are often "good enough" for many everyday tasks while costing a fraction of leading U.S. alternatives, Tirumalai said.

But the emergence of cheaper, more efficient models from China does not necessarily mean that demand for underlying hardware such as semiconductors and data centres will fall. As some analysts have pointed out, when technologies become cheaper, people often use them more, not less.

Moreover, there is evidence to suggest that K3 does not offer the dramatic cost reductions that some assume. Nomura, citing benchmark provider Artificial Analysis, estimated K3's average cost at about $0.94 per task, close to OpenAI's GPT-5.6 Sol.

It is also worth looking back to what happened following the "DeepSeek moment" in January 2025. After the Chinese start-up's R1 model shocked global markets by matching top-tier U.S. systems at a fraction of the cost, there were widespread fears that Silicon Valley might have massively overinvested in AI compute.

But after the shock died down, the world's technology ecosystem doubled down with huge capital expenditure on AI hardware, underscoring the idea that cost-effective models do not necessarily reduce compute demand.

Washington is also seeking to reinforce America's existing advantage in computing resources. U.S. Treasury Secretary Scott Bessent said in an interview released this month that the U.S. could soon account for 80 per cent of the world's computing power, although he did not explain the methodology or provide a timetable.

Tirumalai said total infrastructure demand depended on the computing required for each AI task multiplied by the number of tasks performed. Efficiency should reduce the first figure, but cheaper and more capable models could cause the second to rise more quickly as adoption expands.

He compared the process with mobile communications. Successive generations of networks sharply reduced the cost of transmitting data, but overall consumption surged as video streaming, gaming and social media created new uses.

"The rise of Chinese AI should not be viewed simply as a threat to today's winners," he said. "It may instead reshape where value is created and who captures it."

Lei Meng, China equity strategist at UBS, said technology and AI would remain the main investment themes in 2026, with sectors linked to AI capital expenditure benefiting from spillover beyond narrowly defined AI spending.

The more immediate pressure from K3 may therefore fall on the margins of model providers rather than on demand for computing infrastructure.

Technology investor Gavin Baker argued that an industry dominated by only a few frontier laboratories would allow those companies to preserve high inference margins and gradually expand into infrastructure and software. But greater competition from Chinese and other open-weight models could make it harder for OpenAI and Anthropic to capture such a large share of AI spending.

That could leave more spending for chipmakers, cloud providers, data-centre operators and application developers. Models developed in China would still require processors, memory and electricity wherever overseas users choose to deploy them, according to Baker.

Moonshot's own roll-out offers an early test of that argument. Despite claiming a sharp improvement in development efficiency, the company is still racing to bring more graphics processing units online after demand for K3 overwhelmed its existing capacity.

Read the article at SCMP.