An alternative model: China’s low-cost AI strategy

A different path to being an AI leader

In January 2025, China commanded attention in the artificial intelligence (AI) race with the announcement of DeepSeek’s R1 model, rekindling enthusiasm for Chinese technology stocks and China's AI ambitions. Yet, since then, Chinese markets have largely taken a backseat in the AI investment story as American players spent heavily on new models and data centers to power them. Meanwhile, the supply chain exposed to these investment flows has created winners in select Asian markets, particularly in Taiwan and Korea (see Figure 1). 

On the surface, it appears the market has gauged that China is not the obvious destination for AI-related investment capital. Its most advanced semiconductor foundries are two to three generations behind the frontier, while US-led export controls have restricted access to the most advanced chips. In addition, Chinese AI firms spend a fraction of what US hyperscalers deploy (see Figure 2). By most headline public benchmarks, many Chinese models also trail the leading US models—even if the gap has narrowed materially.¹ In short, the market has tended to view China as lagging where investors have so far placed the highest premium in the AI trade: the semiconductor supply chain, large-scale frontier data centers, and model leadership.

We argue here that China’s position in the AI race should be rethought. Domestically, China is working towards a differentiated AI strategy focused on minimizing the cost of deployment across the economy through a combination of open-source and highly efficient models. We think this is driven in large part by US-led export restrictions designed to slow China’s progress in AI. We anticipate that these low-cost models will help facilitate rapid diffusion and adoption of the technology. Meanwhile, China appears to be expanding its role in building and outfitting data centres. That includes progress in its own domestic semiconductor manufacturing capabilities that will likely benefit from its own data center build-out as well as global supply chain bottlenecks.

In this part one of a two-part series, we examine China’s AI strategy and the investment narrative it supports. We note that markets may be underappreciating China’s role in AI hardware supply chains beyond semiconductors, such as in power and networking equipment, and we will discuss that in more detail in part two of our series.

Limited access to semiconductors has shaped China’s AI strategy

We see the hardware backdrop and US-led export controls as an important context for China’s strategy. Years of US-led restrictions on China’s access to technology have pushed Chinese policymakers to focus on growing a domestic, high-tech semiconductor sector. Progress has been gradual and required innovation that has narrowed but not closed the gap with rival manufacturing centres.

As early as 2018, the first Trump administration curbed exports of semiconductors and related manufacturing equipment to China, with a far more restrictive crackdown beginning in 2022 in response to the growing capabilities of AI systems. These actions have been designed to crimp progress in China’s domestic semiconductor industry, particularly around high-end memory and logic chips. The result could in fact be the opposite as China is more highly incentivized to innovate and develop domestic capabilities.

The Trump administration loosened restrictions in December 2025, but these actions were only a nominal restoration of access, in our view: Only select, older-generation chips were permitted for export—and subject to a 25% revenue surcharge payable to the US government. In effect, China is being offered previous-generation horsepower while its mainland domestic manufacturing capabilities remain cordoned off to advanced equipment used in Taiwan and Korea.

Notably, despite approved exports, we have observed very few exports of Nvidia’s China compliant chips into the domestic market.² Localization has taken precedence with data centres reportedly required to source at least 50% of chips from indigenous suppliers.³ A surge in black market prices of foreign leading-edge chips suggests they are still helping to fill gaps, with domestic players still progressing.⁴

Doing more with less: Hardware

Developing home-grown leading-edge chips to reduce dependence on US technology therefore remains a policy priority. Yet access to manufacturing equipment is a challenge: The export ban on state-of-the-art extreme ultraviolet (EUV) lithography equipment from ASML restricts Chinese foundries to the 7-nanometer (nm) process node, leaving them two to three generations behind the industry leaders TSMC and Samsung.

SMIC, the only Chinese foundry currently capable of fabricating at 7nm, has made limited progress in closing the architecture gap by using a multi-patterning workaround using older-generation deep ultraviolet (DUV) equipment. Instead of shrinking the size of transistors alone, this process involves printing an additional pattern of transistors on a chip. The challenge in this technique, however, is that the error rate and manufacturing time both increases, making this a costly approach that’s more challenging to commercialize.

Separately, China is attempting to close the technology gaps domestically, including early-stage efforts to develop indigenous extreme ultraviolet (EUV) machines, but we think commercial breakthroughs below 7nm are unlikely in the near term, even as some reports suggest progress may be closer than initially believed.⁵

With access to EUV lithography equipment blocked, China has focused on a new front: competing on architecture instead of transistor size alone. Huawei's latest announcement of "logic folding" is a clear expression of this, in our view. Rather than spreading circuits across a flat surface, the design stacks them vertically, shortening the distance signals must travel and squeezing more density and performance out of older-generation equipment. Huawei claims a 55% lift in transistor density and 41% gain in power efficiency when measured against their own conventional single layer (2D) chip designs on the same node and expects its high-end chips will have transistor density equivalent to 1.4nm processes by 2031.⁶

Vertical stacking itself is not new (TSMC, Intel and Samsung have been refining it for years), but unlike conventional 3D stacking which bonds separate chips together, Huawei folds the circuitry within a single chip, accomplishing greater integration and therefore computing speeds. Huawei has packaged the approach under what it calls the “Tau (τ) Scaling Law”, a deliberate counter to Moore's Law that swaps transistor size for signal propagation time (or speed of information) as the metric of progress. Unproven at scale and with hurdles like heat dissipation ahead, the framing still matters, in our view. A 7nm or 5nm chip could in theory rival a leading-edge node through architecture rather than transistor size.

Together with multi-patterning techniques, we suspect the rise of logic folding techniques presents an upside risk for China’s homegrown high-tech semiconductor industry ambitions. Still, China has so far been unable to accomplish the same scale in manufacturing output as rivals at TSMC and Samsung even while working in older-generation architecture. This suggests to us that the Chinese AI strategy must also do more with less at the model layer – and may help to explain why China’s AI strategy has evolved differently.

Doing more with less: Software

Limited access to top-end compute has pushed Chinese AI labs toward smaller, more efficient models and deployment strategies optimized for lower-cost inference on older or domestically produced hardware. Reflecting this, capital expenditures are significantly less at Chinese hyperscalers relative to US counterparts, with a yawning gap in terms of data centre scale and capacity.⁷ We see two core themes in the Chinese AI software strategy: efficiency and openness.

Investment risks

The value of investments and any income will fluctuate (this may partly be the result of exchange rate fluctuations) and investors may not get back the full amount invested. Past performance is not a guide to future returns.

Artificial intelligence (AI) technology companies are sensitive to specific risks such as small markets, business cycle changes, economic growth, technological progress, obsolescence, and regulation. These companies may have limited products, markets, resources, or personnel, making their securities more volatile, especially for smaller start-ups. Rapid technological changes can adversely affect their results. AI companies often rely on patents, copyrights, trademarks, and trade secrets to protect their technology, but there's no guarantee these protections will be sufficient. Significant research and development (R&D) spending doesn’t ensure product or service success.

Investments in Chinese AI companies are subject to risks associated with government intervention, evolving AI and data-security regulations, geopolitical tensions, export controls, potential delisting or audit-related issues, and increased regulatory scrutiny of AI-related business activities and securities markets. These factors may adversely affect operations, valuations, liquidity, and investment returns.