Cheap Chinese AI models like Kimi are not a threat to US tech

Chinese AI models are gaining traction by making AI tokens dramatically cheaper. While this may sound counterintuitive, it should benefit US hyperscalers (Google, Microsoft, Amazon and Meta) and other software companies by accelerating adoption despite putting pressure on the model makers.

OpenRouter is a popular platform that allows developers to access hundreds of models through a single interface and select the most suitable ones based on price and performance. Its latest data show Chinese models from Tencent, Xiaomi and DeepSeek gaining share quickly alongside the “US team”.

The script looks familiar. Much like EVs, China successfully captured market share by offering decent quality at significantly lower prices with their AI models.

Treating a cold does not require a neurosurgeon

Will users always choose the most powerful model? In reality, most tasks do not require PhD level intelligence. A model that is capable enough and available at a much lower cost makes more economic sense.

A multi-model approach is increasingly popular. An application can now orchestrate with one AI model for image recognition, another for data retrieval and a third for preparing responses, while leaving deeper reasoning to the most advanced model, which the token costs may be significantly more expensive.

Why are Chinese models so cheap?

There is no single explanation, but rather a combination of technological, market and government forces.

First, there is architectural efficiency. Mixture-of-experts design is a common approach among Chinese models, under which only part of the model is activated for each task. Chinese developers have also focused heavily on memory efficiency and hardware use. US restrictions on access to high-performance chips have likely accelerated this focus.

Second, similar to what happened in China’s EV market, domestic competition among AI model developers is fierce. DeepSeek has declared a price war against competitors and all Chinese models went lower. As in the EV industry, many Chinese companies are prioritising market share over near-term profitability.

Third, AI is a core priority in Beijing’s development plans, supported through subsidies for domestic chips, lower electricity costs, tax incentives, infrastructure funding and more…

Cheaper means more will be consumed

As tokens become cheaper, usage increases (an example of Jevons paradox). Agent based applications perform more steps and consume more tokens. As token costs fall, running AI agents becomes more economically viable. And the demand curve is basically flat once token costs become minimal.

Hyperscalers may become the AI toll roads

This trend should be positive for the cloud businesses of Microsoft Azure, AWS and Google Cloud.

If the multi-model approach becomes the standard way of building AI agents, the value of hyperscalers will be less dependent on whether they own the best model, as Google does, or which model provider they partner with, such as Microsoft with OpenAI and AWS with Anthropic. Instead, it will depend on how effectively these hyperscalers provide computing power and reliable infrastructure for managing workloads.

Western enterprise customers are likely to focus on reliability, security, regulatory compliance, ease of integration and total cost. When selecting models, however, they may still prefer providers such as OpenAI and Anthropic.

Who else benefits?

Chip manufacturers and data centre operators should benefit if lower costs drive higher usage.

Investors have also started to realise that software companies with strong competitive moats are likely to transform, with AI reducing development time and allowing smaller teams to produce more.

Who may lose?

Premium model providers face the greatest pressure, as it will become harder to justify the price difference if Chinese models are “good enough”. Although these models will remain important for advanced reasoning, their pricing power is likely to weaken.

The employment outlook for software engineers may also look increasingly gloomy. The rising number of layoffs across technology companies could tell.

The investment takeaway

While cheaper Chinese models appear to threaten the leadership of US technology, the economic effect could be very different. Lower costs should accelerate AI adoption, create more AI agents and generate more workloads for cloud providers to host. No single model needs to dominate, but more customers to keep consuming computing resources on their infrastructure.

Software businesses with strong application moats are also likely to benefit from lighter operating models with lower headcounts and greater productivity.

For investors seeking exposure to this trend, the ETFS US Technology ETF (ASX: WWWW) provides access to leading hyperscalers, semiconductor companies and software businesses positioned to benefit from growing AI adoption.