Chinese AI rivalry threatens US margins, not its ecosystem

Chinese AI models are narrowing the performance gap with their US rivals, often apparently at a fraction of the cost. This is raising questions about the returns on US AI infrastructure spending, and the valuations of many leading US technology companies.

At first glance, some Chinese models seemingly deliver almost state-of-the-art capabilities with fewer resources, and at lower cost for their end users. Some are also ‘open weight’, meaning that their underlying parameters can be downloaded and run on local networks, easing data privacy concerns. Why then would corporate users continue to pay a premium for frontier US models? Such concerns intensified as US technology companies continue to commit capital to AI’s infrastructure and investors question whether the sector can generate returns in line with its spending.

However, this interpretation misses a bigger picture. While Chinese open-weight models are likely to put pressure on the pricing power and profit margins of US model providers, we think they do not represent a fundamental threat to broader US-led AI ecosystem. Cheaper and more capable AI models should accelerate adoption of the technology, expanding demand across the infrastructure, semiconductor, cloud and application parts of the value chain, where much of the long-term value creation is likely to occur.

China’s cost advantages are less stark than they look

First of all, the growing perception that China is building the same or better AI for a fraction of the price is misleading. Many claims about the cost advantages of Chinese models in capex terms underestimate the impact of ‘distillation,’ or training a model partly on outputs from a more advanced US system. More importantly, a low price-per-token from a Chinese model does not necessarily mean a low cost-per-result once the AI model is actually being used. A less capable model may need more tokens, computing power, energy and time to finish the same job. Cost-per-completed task and time-per-task are therefore more useful benchmarks. By these metrics, many US models do compare favourably with Chinese rivals.

Nor are US developers standing still. They have slashed prices, and some – including Meta, Nvidia, and OpenAI – are also providing open-weight models, with more on the way.

…many US models do compare favourably with Chinese rivals

Tech diffusion versus frontier models

If cheap or locally-deployed models become standard for many AI applications, would that significantly undermine valuations across the broad US AI ecosystem? We do not think so, for three reasons.

First, lower AI prices should expand AI demand. This is a version of ‘Jevons paradox’: when a technology becomes cheaper, people often use more of it. For AI, a lower cost-per-task should encourage more tasks, longer reasoning and more autonomous agents. The US-led ecosystem could therefore continue to thrive because of its strong position in cloud infrastructure, global networks and proprietary data, none of which depend strongly on the particular AI model being used.

Second, frontier models may still attract a premium. The strongest US models still retain an overall capability lead. Chinese models match them on some benchmarks, but gaps remain on demanding cyber, scientific and long-horizon agentic tests – complex, multi-step tasks that AI systems must complete autonomously over time. For some applications, such as scientific research, deploying the best intelligence available will still generate an advantage, and hence some pricing power for the most advanced model providers. Such capabilities at scale can be plausibly monetised, maybe aggressively, justifying some positive valuations for frontier developers in the US, such as Anthropic and OpenAI, although margins for less cutting-edge applications will be under pressure.

Third, security and sovereignty concerns point to a hybrid market where both US and Chinese models will play a role. The brief US restrictions on foreign access to Anthropic’s Fable 5 and Mythos 5 models highlighted the risks of relying on technology from a single country. Yet ‘sovereign AI’ initiatives outside the US are advancing only slowly because building a competitive AI stack from scratch remains extremely expensive. At the same time, given the lack of alternatives, large foreign users of AI and cloud services may still prefer US to Chinese models because of governance and security concerns.

We expect both systems to coexist

In other words, it is plausible that demand for AI evolves towards the widespread use of cheaper models, including open-weight systems deployed locally for many applications. That would make it harder for frontier providers such as OpenAI or Anthropic to sustain premium pricing, margins and valuations. But does not mean that users will abandon US AI ecosystem and switch exclusively to Chinese alternatives. We expect both systems to coexist, with businesses using local or open-weight models, either from China or the US, for sensitive or high-volume workloads, while relying on hosted frontier models when superior capability or simpler operation outweighs the risks of using a third-party provider.

Read also: Oil market outlook – Chinese demand and post-conflict lessons

The next leg of the US-China race

The US-led AI ecosystem has a significant advantage in advanced accelerators – the specialised chips used for AI – as well as high-bandwidth memory, which feeds those chips data quickly, and the high-speed links that allow many chips to work as a single system. US export controls on cutting-edge chips and chipmaking equipment should help preserve these advantages in the near term. The controls could tighten further if new Chinese innovation increases concerns in Washington.

related content.

investment solutions.

learn more

investment insights

The Intelligent Allocator: The humbling art of stock picking

August 11, 2026

How China works around these restrictions will be one of the most important questions in AI in the years ahead. In the near term, the country may remain constrained by a shortage of advanced computing capacity. It will continue building domestic chipmaking capabilities, including lithography technology, while encouraging closer cooperation among its technology companies, exploring alternative chip architectures.

However, China has a critical advantage in electricity production. The country generated roughly 10.6 petawatt-hours (PWh) in 2025, more than twice the US, and is expanding both generation and long-distance transmission at a faster pace. The most important issue in this context is ‘speed-to-power,’ or how quickly a new data-centre site can secure a large, reliable connection. A Chinese data centre may take between one and three years to build and supply with power, compared with a five-to-ten-year timeline for a major US grid connection and transmission upgrade. Such delays could therefore slow US AI deployment and force expensive on-site electricity generation.

This race will continue to generate economic momentum, with positive impacts on growth in both countries

For now, however, the US retains a sizeable lead in installed AI compute capacity, reflecting its hardware advantage. This currently more than offsets China’s greater electricity supply, even as individual US projects face grid-connection bottlenecks. China’s power advantage therefore remains a source of strategic opportunity, one that could become decisive if the country closes the gap in chips and talent. 

Read also: Energy shock exposes asymmetries in Asia

The AI cycle will continue in both superpowers

The ‘frontier-versus-adoption’ framework is still valid. The US retains the lead in frontier AI capabilities, while China is well placed to accelerate adoption through lower-cost models and deployment at scale. But the race is far from settled. US developers are responding through price cuts and open-weight offerings, while China tries to close the gap with frontier capabilities.

Neither country will accept the lead of the other in either frontier models or AI adoption. Rather than converging on a single winner, we expect the two ecosystems to evolve in parallel, driving continued innovation and investment across the global AI landscape, with some volatility as one temporarily gains at the expense of the other. This race will continue to generate economic momentum, with positive impacts on growth in both countries.

For investors, value creation is likely to accrue to the owners of AI infrastructure, chips, data and distribution networks, rather than exclusively to frontier model developers. This is why we maintain a positive view of hyperscalers and semiconductor companies relative to pure-play model providers.