Disrupting the disruptors: China’s challenge to US AI

Over the past several decades, Chinese companies have provided lower-cost alternatives to US and other developed market products in a wide variety of industries ranging from consumer staples to electronics to automotive. Now, as AI models developed by US-based companies disrupt industries, economies and even societal models, China is again posing an increasing challenge by offering comparable AI products at lower costs.

When a Chinese start-up called Moonshot AI launched its Kimi K3 AI model in early July, investors reacted with concern over what the new model could mean for US-based AI providers. Touted as offering superior performance to models from US-based companies such as Anthropic and OpenAI, the Kimi K3 debut came on the heels of China’s Zhipu unveiling its GLM 5.2 AI model as another viable lower-cost alternative to US models.

China’s AI industry may still be underappreciated by investors, but both launches demonstrate that its influence on global AI dynamics is becoming harder to ignore. Meaningful AI innovation is taking place in China, and the entry of cheap and open large language models (LLMs) from Chinese companies has the potential to limit both the pricing power and the revenue potential of US LLM providers. That matters because it could undermine the current financing model of the whole supply chain and erode the value propositions of many of the large US providers.

Too cheap to ignore

Kimi K3 and GLM 5.2 have arrived at a time when the reaction against “token-maxxing”— maximising token usage with a view to accelerating potential AI gains — appears to be driving growing uptake of Chinese models. Earlier last year, markets were surprised by the “DeepSeek moment”, when the lightweight model achieved comparable performance at much lower cost. Users of AI today may increasingly need justification for the large price premium charged by frontier LLMs. If low-cost models can keep delivering broadly similar capability with less capital investment and less advanced hardware, investors will rightfully scrutinize the rationale for the ongoing high spending by US companies.

In my view, the growth of Chinese models is a reminder for investors to distinguish efficiency from excess. Almost every technology revolution, from railroads to broadband to mobile networks, inevitably overbuilds and overpromises beyond economic reality, and only after the hype fades does the excess become evident. The presence of a competitive alternative keeps hyper-spenders benchmarked and provides a counterweight to optimistic projections. This doesn’t mean the prevailing thesis is wrong, but I think it increases the burden of proof for today’s capital allocators.

An “open” question

China’s government and Chinese labs have embraced open-source models as a core pillar of China’s “AI+” initiative and as a competitive strategy. It started off as a forced choice in the face of hardware constraints but has gradually redefined the competition in the AI field and challenged the token-centric business models of the closed-source LLMs.

Open models are easier to diffuse and adopt, and harder to contain. They travel through online communities, local deployments and customised redevelopment. They scale through network effects rather than sales pipelines and embed rapidly across the global ecosystem.

Open models, which allow for user-contained deployment, also better protect data privacy and ease fears that proprietary business secrets might be at risk. For model providers, they help shift the compute burden to the user end. With model access made almost free, open models expand the AI battleground from core intelligence to deployment, distribution, servicing, data, workflows and business outcomes, where the pricing logic has yet to emerge.

Policy versus technology

Perhaps the biggest influence on the medium-term future of AI may not lie with the technologists who build and train the models, but with the policymakers who must decide how far and how fast AI is allowed to proliferate and how best to govern it as it does so. The narrowing gap in performance between US and Chinese LLMs is also putting pressure on regulators and politicians to consider trade-offs between slowing down technological progress for safety and governance or speeding up progress in hopes of winning the “AI race”.

The US government’s short-lived shutdown of Anthropic’s Claude Fable 5 model earlier this summer reflected this dilemma. When a sweeping shutdown and functional downgrade were deemed necessary to contain emerging risks, we may have reached a stage where governance, rather than technology, becomes the primary constraint on progress. The tension between technological advancement and societal tolerance will only grow from here.

Adding geological complications on top, if both the US and China can agree to prioritize safety guardrails first, the current projections by the market may only need to build in additional speedbumps. However, if neither side can afford “unnecessary” regulatory delays versus the other, we could see a reliance on export bans or access control instead, leading potentially to a more disorderly and fragmented ecosystem. In my view, these are no longer hypothetical scenarios. The safety and governance issue is eventually unavoidable for all models, open or closed, US or Chinese, but the US-China competition may skew the incentives and amplify the risks of governance missteps.

For investors, I believe the takeaway is not to become bearish on AI, but instead to become more cautious with projections and assumptions and recognize that national security and geopolitical factors may matter far more for AI than they have for previous waves of technological advance. China could keep being a credible challenger in the AI race long enough for markets to question the current consensus. This suggests the need for a more deliberate approach across equities and credit exposures with an emphasis on diversification across both US and Chinese AI ecosystems.