Low-cost Chinese AI models forge ahead, even in the US, raising the risks of a US AI bubble

Nvidia’s latest earnings report reassured some. But Chinese AI models are fast gaining a following around the world, underlining concerns over an ‘AI bubble’ centered on high-investment, high-cost US models.

Expert comment

Published 20 November 2025

Updated 21 November 2025 — 4 minute READ

Image — Nvidia CEO Jensen Huang listens as US President Donald Trump delivers remarks at the US-Saudi Investment Forum at the Kennedy Center on 19 November 2025 in Washington, DC. (Photo by BRENDAN SMIALOWSKI/AFP via Getty Images)

James Kynge

Senior Research Fellow for China and the World, Asia-Pacific Programme

This week Jensen Huang, CEO of Nvidia, took issue with talk of an ‘AI bubble’, as Nvidia announced bumper earnings growth in the latest quarter on strong sales of AI chips. ‘There has been a lot of talk about an AI bubble. From our vantage point we see something very different,’ he told analysts on Wednesday.

Yet Huang’s previous remark on 6 November, that China will ‘win’ in AI , should be taken seriously in the West – in part for what it says about bubble fears. 

Huang is not the only prominent US figure to have commented recently on China’s remarkable progress in artificial intelligence.

Brian Chesky, CEO of Airbnb and a friend of OpenAI founder Sam Altman, said that his company ‘relies heavily’ on Alibaba’s Qwen because the Chinese model is ‘very good’ and ‘also fast and cheap’. In contrast, he said of OpenAI’s ChatGPT: ‘I don’t think it was quite ready’ for Airbnb’s needs.

Remarkably, Martin Casado, a partner at US venture capital firm 16z, says that most of the AI entrepreneurs he sees in the US are using AI models made in China. ‘I’d say 80 per cent chance [they are] using a Chinese open-source model,’ he was quoted as saying. 

Chamath Palihapitiya of Social Capital, a US investment firm, said he has chosen Kimi K2, an open-source AI model developed by the Chinese company Moonshot AI, to perform much of his company’s work. This is because Kimi K2 ‘was really way more performant and frankly just a ton cheaper than OpenAI and Anthropic’, he said.

The AI race between the US and China is still in its early stages. It is too early to call a definitive winner. But the assessments of key industry figures that low-cost, high performance Chinese AI models are gaining ground in the US should add to concerns over an AI bubble’ – one centred around high-cost US models. 

China has huge cost advantages 

It is clear that Chinese AI companies enjoy key structural advantages over US counterparts. The main one is cost. This includes both the cost of developing new AI models and updates, as well as the cost of operating Large Language Models (LLM) for customers.

In terms of development costs, China is quite simply in a different universe. By some estimates, Moonshot AI’s Kimi K2 cost just $4.6 million to train, a fraction of the billions that OpenAI is spending on research and development. 

Similarly, the costs of running the leading Chinese LLMs – DeepSeek, Qwen-Plus and Kimi K2 Thinking – come in at a sharp discount to operating both GPT 5 from OpenAI and Claude Sonnet 4.5 from Anthropic. For reference, the Claude Sonnet 4.5 costs $15 per million output tokens, while the Kimi K2 Thinking costs $2.5 .

Another aspect of the leading Chinese LLMs’ appeal is that they are open-source. This allows users to deploy models on their own private infrastructure and ensure that sensitive data remains in-house. 

Several US LLMs are closed source, including ChatGPT. However, Llama from Meta, Phi from Microsoft and Gemma from Google are open-source. Indeed, even OpenAi launched the open-source GPT-oss in August.

The US has access to much more powerful chips

The US AI companies, however, enjoy an advantage in their unrestricted access to the latest AI chips, the Blackwell series made by Nvidia. The US has banned these chips from being sold to China, with US president Donald Trump saying this month that only US customers should have access.

US export controls are far from watertight. Some banned chips are smuggled into China. Others are obtained by Chinese companies in third countries. 

As things stand, the most advanced Nvidia chip that China can access under Washington’s export control regime is the H20, a high performing semiconductor which is nevertheless a full generation behind the latest Blackwell B200 and B100 chips. This gulf is thought to give users of the latest Blackwell chips up to 15 times more compute power over those using the H20.

However, the full picture is more complicated. For one thing, US export controls are far from watertight. Some banned chips are smuggled into China. Others are obtained by Chinese companies in third countries. Still others – such as the Nvidia H800 – are available from stockpiles amassed by Chinese companies before the US export controls came in.

For example, the Kimi K2 Thinking was trained on Nvidia H800 chips that had been stockpiled from before an export ban took force in late 2023. In spite of its inferior chips, Kimi K2 outperforms ChatGPT-5 and Anthropic Claude Sonnet 4.5 on several standard evaluations. 

The reason that some Chinese AI models can achieve superior performance in some evaluations while using inferior chips derives from architectural and training innovations – more impressive Chinese accomplishments

The US has much greater data centre capacity

AI models are trained in data centres, which are expensive to build and consume a lot of electricity to run. The US has much more data centre capacity than China and therefore more capacity to train new AI models or their updates. 

China…has gone on a data centre building spree.

China, however, is moving to catch up and boost its competitiveness. 

The country has gone on a data centre building spree, and subsidizes electricity, cutting energy bills by up to half for some of the country’s largest data centres. Additionally, some new data centres are being built in desert areas, fed by power from solar panels – the world’s cheapest source of electricity.

Chinese AI models are gaining global acceptance

There is no doubt that several Chinese LLMs are gaining a global following. Alibaba’s Qwen has surpassed 600 million downloads since launching in 2023, with more than 170,000 derivative models developed, according to the People’s Daily, an official Chinese newspaper. 

Much of this activity is in the developing world, as users with fewer financial resources seize on Qwen’s open-source service and plentiful language functions to develop their own AI systems.

China is also a formidable force in AI-related intellectual property, with 1.6 million AI-related patent applications made as of April this year, or 38.6 per cent of the global total – the highest share of any country.

How the West should respond

Sam Altman was correct when he said earlier this year that OpenAI had been on the ‘wrong side of history’ in its initial strategy to make ChatGPT closed source. 

US…strategy should involve efforts to discover many more revenue-generating applications for LLMs.

If US and other Western AI model developers intend to boost their global footprint against Chinese competition, they should go all-in on making open-source models available at lower operating costs to users. 

US AI models should also seek to slash costs not only to compete, but to allay concerns of an emerging AI bubble in the US. Part of this strategy should involve efforts to discover many more revenue-generating applications for LLMs, just as China is doing. 

Should US AI companies’ remaining competitive advantages be eroded, while their models struggle to compete with Chinese alternatives even in the US, the enormous AI investments feeding US stock market growth may indeed begin to look like risky bets. In that case, the bubble Huang dismissed this week may begin to look very real.