Huawei v. Nvidia - The Rise of China AI

Oct 10

Story Begins…

At the end of last year things were not going well for China. The trade war, the slumping real estate sector, and declining consumer spending and consumer sentiment all indicates a recession. On the technology front, it appeared far behind US semiconductor and AI companies, and it’s almost impossible to catchup because China was missing so many parts of the semiconductor ecosystem, from ASML’s EUV to advanced packaging, to HBM, etc.

But the Deepseek moment was really a moment of awakening for China. People realized that they don’t need Gigabytes of Nvidia computing power to run a LLM as good as OpenAI. And China is not far behind in AI research. In fact, if there were any country in the world that could challenge US leadership on AI, it is China. International investors were also giving China second looks.

The trade war was brutal but China’s hard stance paid off and eventually reached a standoff, not a one sided defeat. I think this is a result acceptable to both countries. Investors overjoyed.

Then comes the rise of NV-Chain, the optical modules and AI PCB receiving massive orders from hyperscalers. Generating billions of profits and allows China to share the fruits of the AI supercycle. The innovation index surged 50%+ within two months following the news and good earnings. Then domestic chips came along as Huawei released it’s Ascend 910C superpod, first time since 2019 when it was targeted in the trade war.

Overall, I think we can be certain about two things: 1) China’s technology capabilities were underestimated and are catching up fast 2) China and US will cowrite the rules of AI and share the productivity improvements brought by AI. China is leaning heavily to semiconductors and AI, much like the US companies, so investors in China must have exposure to AI & Semiconductors, otherwise they cannot outperform the index.

Horizons Capital have closely followed the developments in China’s semiconductor industry, and we have written research on semi equipment company (CFMEE), testing company (V-Test), optical module company (Everbright), and upcoming ones on connectors and more. Many companies have achieved technology breakthroughs and we are seeing a domestic ecosystem forming, coming up to challenge incumbent players.

Huawei’s Strategy

In a recent People’s Daily interview, Huawei founder Ren Zhengfei outlined four guiding principles to overcome China’s chip disadvantages under US export controls: “use mathematics to replace physics,” “use post-Moore to replace Moore,” “use cluster computing to replace a single chip,” and “use electric power to compensate for computing power.”

These phrases encapsulate how Huawei and China plan to get around hardware limits. In plain terms, they mean optimizing software and algorithms (math), exploring beyond-Moore’s-law technologies (post-Moore), networking many chips into one system (clusters), and leveraging abundant electricity and data-center scale. Each approach is already visible in Huawei’s new SuperPoD AI clusters and China’s broader chip strategy

1. “Use mathematics to replace physics”

This means optimizing algorithms and software to make up for weaker hardware. Same way Nvidia works closely with game developers to optimize their GPU’s performance on those games, Huawei is working closely with DeepSeek and other domestic LLMs to better cater to their needs. Here Huawei has the homegame advantage and they can better service domestic hyperscalers/LLM clients.

Instead of relying solely on raw chip performance (physics), Huawei focuses on sophisticated mathematics – better compilers, machine-learning algorithms, neural network optimizations and compiler tricks – to boost efficiency. In practice this looks like software frameworks and AI models that squeeze more performance from each chip. For example, Huawei’s Ascend AI chips saw huge gains through software and architecture improvements: a recent paper reported Huawei’s CloudMatrix 384 cluster outperformed Nvidia’s H800 GPUs on a large language model, thanks to a custom “unified bus” interconnect and optimized code. In short, Chinese engineers are using clever software mathematics to offset the fact that their individual chips aren’t as fast.

Investor Note: This strategy implies heavy investment in AI software, compilers, and algorithmic research. Chinese companies and startups (like Huawei, Cambricon, etc.) are likely pouring resources into optimization tools and AI frameworks. The current Ascend chip supports CUDA but also integrates its own software for AI. Much like the HarmonyOS Huawei created to surpass Android operating system.

2. “Use post-Moore to replace Moore”

This phrase means pursuing alternatives once Moore’s Law (transistor scaling) stalls. Since making ever-smaller, faster chips is hitting physical limits, and China does not have access to EUV necessary to make advanced nodes below 7nm. This requires a new solution to circumvent this short coming. Examples include 3D chip-stacking, photonic/optical computing, neuromorphic chips, or specialized AI accelerators. Huawei’s strategy includes building processors (the Ascend series) on available process nodes (e.g. 7 nm) and planning next-gen chips (Ascend 950/960) for 2026–27. Meanwhile, Chinese foundries (SMIC, CXMT) are heavily funded to push their processes and memory fabs.

Investor Note: Watch for Chinese firms in advanced packing technologies and SMIC-the only fab for advanced node semi. Also CPO, light optics are not constrained by Moore’s Law and both companies are working on CPO, US is the leader in this field but China is not far behind.

3. “Use cluster computing to replace a single chip”

China’s answer to not having a world-class single chip is to link many chips together into a “SuperPod” or cluster. Huawei’s new SuperPoD systems are real-world embodiments of this idea. At Huawei Connect 2025, CEO Eric Xu unveiled the Atlas 950 and 960 SuperPoDs, each combining thousands of Ascend NPUs (neural processing units) into one logical machine. For example, the Atlas 950 SuperPoD packs 8,192 Ascend NPUs, and the Atlas 960 has 15,488 NPUs – far beyond any single GPU NVL72. These SuperPoDs act as a single AI supercomputer node. Going further, Huawei announced SuperClusters by linking many SuperPoDs: the Atlas 950 SuperCluster spans over 500,000 NPUs and the Atlas 960 over 1,000,000 NPUs.

These massive clusters use proprietary interconnects (Huawei’s “UnifiedBus”) to deliver high bandwidth across machines. This system-level design illustrates Ren’s point: a rack of chips working together can match or exceed the performance of a faster single chip. Chinese press and analysts note that Huawei’s “CloudMatrix 384” rack (384 Ascend NPUs) already rivals Nvidia’s top systems. In effect, cluster engineering and packaging (chiplets, 3D stacking, networking) are being used to “make up for single-chip limitations”

Investor Note: look at networking (optical links, switches) like Everbright and Huafeng Technology. Huawei’s success in cluster design underlines opportunities in AI infrastructure – not just chips – for investors. Huawei has deep know-how in transceivers based on their experience from telecom, which Nvidia does not have. So the idea of a Supernode, connected internally by switches & CPC wires, fiber optics and optical modules in between the racks is truely a remarkable feat, and that’s where I see Huawei is leading.

4. “Use electric power to compensate for computing power”

China is leveraging its abundant energy and data-center scale to offset chip shortcomings. Ren noted that AI development in China benefits from “sufficient electricity and a developed information network”, and ChinaTalk highlights “AI hinges on ample electricity” (via its grid and renewables). In plain terms, China will throw more power (and chips) at problems. For example, analysts found Huawei’s supernodes consume hundreds of kilowatts – several times more than comparable Western systems – which is feasible because China built vast new power capacity (coal, hydro, nuclear) in recent years.

Practically, this means China can afford huge AI clusters like the thousands or tens of thousands supernode clusters. Data centers in resource-rich provinces (“Eastern Data, Western Computation”) use cheap power to train big AI models on hundreds of thousands of chips. While US suffers from an outdated infrastructure, expensive electricity, and lack of supply of power that needed to be secured before building mega datacenters.

Investor Note: Look at electrical equipment companies, those that made transformers, the high voltage power lines, the power generators and BSS for datacenters. We have covered Shemar Electric and Yunlu.

Strategic Implications for China’s AI and Chips

Ren’s four phrases sum up how China is sidestepping semiconductor bottlenecks. Instead of breakthroughs in etching 3 nm chips, China’s playbook is to innovate at the software and system level, diversify materials, and build massive infrastructure. I would call this another Deepseek breakthrough, although the reaction this time is more muted.

When asked about where does China stand in its competition in semiconductors/AI, Nvdia founder Jenson said “they are not far behind. They have less advanced nodes, and their software is lot less mature compared to CUDA, but they have the infrastructure, and over 1 billion in users, accounting for 30% of the world’s market for AI applications.”

Overcoming Sanctions: Even without advanced lithography, China can achieve “state-of-the-art” compute by other means. Huawei openly touts that its Ascend AI clusters now rival Western hardware in many AI tasks. This suggests China’s tech firms can remain competitive in AI despite US export controls, by doubling down on engineering and scale.
Ecosystem Growth: Ren’s emphasis on basic research and many operating systems means China is investing long-term. For investors, this signals expansion in software, cloud services, and diversified chip ventures – not just mainstream semis. Areas like AI middleware, compound-semiconductor fabs, HPC server makers, and renewable-backed data centers could see growth. Having big players like Huawei can support many smaller suppliers, adjacent semi companies, and eventually enterprises that can train their models and have access to computing power that is comparable to businesses in US.
Investor Note: We think the most influencial company in AI applications is Alibaba. It has its own ASIC development, it has the Qwen LLM, and it has AliCloud, the biggest cloud provider in China. AI Applications are still being developed, but Alibaba is best positioned to benefit.

Summary for Investors

In essence, Ren’s four axioms reflect a pragmatic pivot: “If we can’t beat them chip-for-chip, we’ll outcompute them through math, scale and sheer power.” The success of Huawei’s SuperPoDs underscores this. Investors should watch how these strategies play out – they indicate both China’s potential to sustain AI growth and the sectors likely to benefit under sustained technological duress.

We are positioning our portfolio to benefit from these trends.