China is moving two pieces simultaneously in its technological race: semiconductors and artificial intelligence. Huawei has just unveiled a roadmap to approach an equivalent density of 1.4 nm in high-end chips by 2031, not through a simple geometric reduction of transistors, but with an architectural approach called Tau Scaling Law. Almost in parallel, Xiaomi is investing billions in open AI models to secure its future in mobile phones, electric vehicles, and connected devices.
These two pieces of news are more connected than they seem. Huawei aims to compensate for US restrictions on advanced lithography with design, system interconnection, and optimization. Xiaomi, on the other hand, wants to prevent its hardware empire from relying on external models for the AI functions that will define the next generation of products. The underlying message is clear: China isn’t just trying to copy a single layer of the tech industry. It’s building alternatives in chips, models, software, and devices.
Huawei seeks an alternative to lithography blockades
Huawei presented at the IEEE International Symposium on Circuits and Systems in Shanghai the so-called Tau (τ) Scaling Law, introduced by He Tingbo, head of the company’s semiconductor business. According to Huawei, this principle proposes replacing traditional geometric scaling with an optimization based on signal propagation time, utilizing technologies like LogicFolding to reduce delays, improve density, and maintain performance evolution in chips.
The key lies in the nuance. Huawei does not claim that China will produce real 1.4 nm chips tomorrow comparable to TSMC or Samsung. Instead, it proposes an equivalent transistor density of 1.4 nm by 2031 through architecture, design, and integration—an approach to boost performance without relying solely on manufacturing smaller transistors. Reuters reminds us that China still lags in advanced fabrication, around 7 nm nodes, while TSMC targets 1.4 nm by 2028.
This approach has strategic logic. Since 2019, Huawei has been subject to US restrictions limiting access to key technologies, including chips, software, and advanced manufacturing tools. The company responded by strengthening HiSilicon, its internal chip design division, and leveraging domestic suppliers like SMIC. Huawei’s return with 5G phones based on chips made in China demonstrated that the blockade hadn’t entirely halted its engineering capacity.
The new roadmap extends beyond smartphones. Huawei plans to apply LogicFolding to future Kirin chips and Ascend AI chips, as well as large clusters composed of hundreds or thousands of chips for data centers. The company also states that, under the Tau Scaling approach, it has designed and mass-produced 381 chips over the past six years, for mobile, telecom, and AI applications.
| Element | What Huawei Proposes | Why It Matters |
|---|---|---|
| Tau Scaling Law | Optimize signal propagation time, not just geometrical reduction | Provides a post-Moore path amid physical limits and sanctions |
| LogicFolding | Architecture to reduce delays and resistive/capacitive loads | Potentially enhances performance without relying on cutting-edge lithography nodes |
| Target 2031 | Equivalent density of 1.4 nm in high-end chips | Trying to catch up with TSMC and Samsung through system design |
| Kirin | First chips expected with LogicFolding | Impact on smartphones and consumer devices | Ascend | Future application to AI accelerators | Significant for China’s alternative to NVIDIA |
| AI Clusters | Optimization from chip to system | Performance no longer depends solely on transistor technology |
Xiaomi wants AI to live inside its hardware
While Huawei works to narrow the semiconductor gap, Xiaomi is accelerating in AI models. The company has committed at least 60 billion yuan, approximately $8.7 billion, to AI over the next three years, announced CEO Lei Jun. Reuters linked this investment to the launch of MiMo-V2-Pro and the market shift from chatbots to agents capable of executing complex tasks with less human supervision.
The next step has been MiMo-V2.5-Pro. Xiaomi presents it as its most capable model to date, with improvements in agentic capabilities, complex software engineering, and long-horizon tasks. According to Xiaomi’s official page, it’s a Mixture-of-Experts model with 1.02 trillion total parameters, 42 billion active ones, hybrid attention architecture, and a context window of up to 1 million tokens.
The technical detail is important, but strategy matters more. Xiaomi isn’t developing AI as an isolated software company. It’s doing so as a manufacturer of phones, connected devices, appliances, and electric vehicles. For such a company, AI isn’t just a cloud product: it could become the layer that coordinates the phone, car, smart home, wearables, and associated services.
| Xiaomi MiMo-V2.5-Pro | Key Data |
|---|---|
| Model Type | Mixture-of-Experts |
| Total Parameters | 1.02 trillion |
| Active Parameters | 42 billion |
| Context Window | Up to 1 million tokens |
| Focus | Agents, complex software, long-horizon tasks |
| Xiaomi’s AI Investment | 60 billion yuan over three years |
| Strategic Sector | Smartphones, EVs, connected hardware |
This progress doesn’t automatically place Xiaomi ahead of OpenAI, Anthropic, Google, or DeepMind. But it does show that hardware manufacturers can become serious players in open models if they have scale, data, devices, distribution, and capital. According to SCMP, MiMo-V2.5-Pro has been recognized by Artificial Analysis as one of the most prominent open-source models in agentic capabilities, and Xiaomi has achieved a notable position in a short time despite having only published open systems for a year.
The new Chinese pattern: hardware, chips, and models under a unified strategy
Huawei and Xiaomi aren’t doing exactly the same. Huawei remains deeply oriented toward infrastructure, telecommunications, semiconductors, and enterprise cloud. Xiaomi is a consumer giant that has expanded into automotive and ecosystems of devices. But both share a strategic intuition: the next tech phase isn’t won solely with an app or an isolated model.
Agentic AI will require chips, memory, networks, data centers, efficient models, orchestration software, and presence within devices. Companies controlling multiple layers will have more leverage to reduce costs, protect data, integrate functions, and differentiate their products. Apple has demonstrated this value through vertical integration of hardware, OS, and services. China is attempting to replicate this logic, but with a more distributed approach among national champions.
In semiconductors, US sanctions have pushed China to seek alternative routes. It doesn’t have full access to ASML’s most advanced EUV machinery and still lags in cutting-edge manufacturing. That’s why Huawei’s approach is significant: if it can’t yet win by lithography node, it aims to win through architecture, interconnection, packaging, software, and system scale.
In AI, the dynamics are similar. Chinese models are increasingly competing on cost, efficiency, and openness. DeepSeek initiated a price competition; Moonshot, Alibaba, Zhipu, MiniMax, and now Xiaomi are competing with high-performance, lower-cost models. Reuters reports that Chinese models are, on average, cheaper per token than many US offerings, partly due to inference cost advantages and technological restrictions that necessitate improvements.
Europe and the US should look beyond headlines
A simple takeaway might be to see Huawei and Xiaomi’s announcements as China’s tech propaganda or an immediate threat to TSMC, NVIDIA, OpenAI, or Anthropic leadership. That would be an oversimplification. Huawei still needs to prove with independent data the real improvements Tau Scaling offers in commercial products. Xiaomi must demonstrate that MiMo can be effectively integrated into devices, cars, and services. Benchmarks aren’t a substitute for sustained adoption.
But it’s also unwise to underestimate the movement. China is learning to compete under restrictions. When it can’t access the best hardware, it optimizes software. When it can’t buy the most advanced lithography, it explores architecture and packaging. When the model market becomes more expensive, it reduces prices and diversifies offerings. Such pressure could impact margins and strategies of current leaders.
For the US, the challenge is that sanctions don’t always curb innovation; sometimes, they redirect it. For Europe, the lesson is different: without an integrated chip, AI, cloud, data, and industry strategy, the continent risks becoming just a user of foreign technologies. Technological sovereignty isn’t declared; it’s built through sustained investment and connected layers.
Huawei and Xiaomi exemplify two versions of this building process. One starts at silicon and moves toward AI clusters. The other begins with mass devices and extends AI downward into physical products. In both cases, the ambition is no longer just to compete in a category—it’s to control the entire chain that transforms data, chips, and models into industrial advantage.
The AI race is no longer limited to labs and models. It now spans wafers, architectures, phones, cars, data centers, and token costs. China understands this and is acting accordingly.
Frequently Asked Questions
Will Huawei manufacture real 1.4 nm chips by 2031?
Huawei speaks of an equivalent density of 1.4 nm achieved through architecture and system optimization, not necessarily by fabricating with a conventional 1.4 nm lithography node comparable to TSMC or Samsung.
What is Tau Scaling Law?
It’s the principle introduced by Huawei to guide the evolution of chips beyond traditional geometric scaling, focusing on reducing signal delays and improving system performance.
Why is Xiaomi investing heavily in artificial intelligence?
Because its business increasingly depends on integrating AI into smartphones, electric vehicles, connected devices, and services. Controlling its own models can reduce dependence on external providers and enhance differentiation.
Does MiMo-V2.5-Pro compete with major Western models?
Primarily in the open-model, agentic, and long-horizon task domain. Its results are promising, but it still needs to prove real-world adoption, integration into products, and sustained performance outside benchmarks.
via: SCMP