As generative AI transitions from a “novelty” to a “productivity tool” across various industries, the importance of a “fundamental hardware” has reached unprecedented heights — AI chips.
As the “heart” of computing power, AI chips not only determine how fast and stable large models can run but also relate to the “lifeline” of China’s digital economy. Recently, the “2025 China AI Chip Market Insight Report” was released, containing numerous key signals: a trillion market scale, accelerated domestic substitution, and liquid cooling technology becoming a necessity… Today, we will break down the core insights of this report in simple terms to see how “attractive” the AI chip sector is and what challenges lie ahead.
1. Current Status: AI Chips Stand at the “Trillion Explosion Point”
First, let’s look at a set of data that has excited the industry:
The scale of China’s AI market is expected to exceed 1 trillion yuan by 2025 (up from only 137.2 billion in 2019), with a 6-year CAGR of 40.28%;
The demand for intelligent computing power is even more staggering: from 59.2 EFLOPs in 2020 to 438 EFLOPs in 2024, with a CAGR exceeding 64%, and it is projected to reach 3036 EFLOPs by 2029 — equivalent to doubling computing power every two years.
This wave of demand is supported by the “AI penetration rate” across various industries: the internet sector leads (91%), followed by finance (78%) and telecommunications (69%), with even traditionally conservative sectors like manufacturing and healthcare rapidly embracing AI.
1. Five Types of Chips “Dividing the World”: The Battle Between General and Specialized
AI chips are not “one-size-fits-all” but are divided into five mainstream types, each with its strengths:
GPU: Currently the “absolute main force”, accounting for 70% of the market in 2024. The ecosystem is mature (e.g., NVIDIA CUDA), and it has strong computing power, suitable for large model training, but it also has obvious drawbacks — high power consumption and cost;
ASIC: The “custom king”, optimized for specific tasks (e.g., image processing, recommendation algorithms), with top performance and energy efficiency, but long development cycles and no flexibility, suitable for stable demand scenarios (e.g., custom solutions for cloud vendors);
FPGA: The “flexible type”, can be reprogrammed repeatedly, with low latency, but less computing power than GPUs, suitable for scenarios where algorithms are still iterating;
NPU/TPU: The “specialized selection”, the former (e.g., AI chips in smartphones) is suitable for edge device inference, while the latter (Google TPU) specializes in tensor computation, with high energy efficiency but poor versatility.
In simple terms: choose GPU/FPGA for flexibility, and ASIC/NPU for extreme efficiency; there is no absolute “best”, only the “most suitable”.
2. Two Major Drivers: Policy Support + Accelerated Substitution
The rise of AI chips cannot be separated from “external push” and “internal pressure”:
Policy support: From the State Council’s “AI + Action” (AI integration into six major fields by 2027) to the Ministry of Finance’s “20% price deduction for domestic chips in government procurement”, and the Ministry of Industry and Information Technology’s “breakthrough in GPU technology”, the national level provides support across the entire chain from R&D to procurement and application;
Accelerated substitution: The U.S. has been continuously upgrading restrictions on high-end chips (e.g., the “de facto ban” on NVIDIA H20 in 2025), forcing domestic chips to “fill the gap”. Data shows that China’s chip self-sufficiency rate has increased from 16% in 2015 to 31% in 2024 — nearly doubling.
3. Competition: NVIDIA Still Leads, Domestic Manufacturers “Emerging Strongly”
When it comes to AI chips, many people’s first reaction is “NVIDIA”. Indeed, with the “moat” of the CUDA ecosystem, NVIDIA accounted for over 80% of domestic AI chip shipments in 2022, especially in the cloud training field, where it is almost a “monopoly”.
However, changes are happening: by 2024, domestic manufacturers’ share has reached 30%, and it is expected to hit 40% by 2025. Behind this are the “hardcore breakthroughs” of companies like Huawei, Cambricon, and Moore Threads.
Three representative domestic companies, each with their “killer features”:
Huawei HiSilicon (Ascend series): Following a “full-scenario computing power” route, covering from cloud to edge to terminal. The newly launched Ascend 910C in 2025 has an FP16 computing power of 800 TFLOPS; even more impressive is the iteration plan — the Ascend 970 in 2028 is expected to reach 8 PFLOPS (10 times the current performance) and will support more flexible computing precision.
Cambricon: The “speeding dark horse”. In 2024, cloud chip revenue reached 1.166 billion (a year-on-year increase of 1187%), and in the first half of 2025, it skyrocketed to 2.87 billion (a year-on-year increase of 4600%)! The core is “integrated cloud-edge-end”, with the Siyuan 370 chip capable of adapting to large model inference, and they have also launched training machines to fill previous gaps.
Moore Threads: The domestic “full-featured GPU” standard bearer. Their self-developed MUSA architecture is compatible with CUDA — this means developers do not need to change much code to migrate applications from NVIDIA. The fourth-generation chip “Pinghu” launched in 2024 supports FP8 precision and can support training for tens of thousands of cards, directly competing in high-end scenarios.
4. Demand: Three New Trends Determine the Direction of the Sector for the Next Three Years
The demand for AI chips is no longer simply about “computing power” but about “efficiency, adaptability, and security”. The three demand insights mentioned in the report can be considered the “golden tracks” for the next three years.
1. Cloud Vendors Collectively “Customize ASIC”: Market to Reach $55.4 Billion by 2028
Traditional GPUs are “versatile but not precise”; for example, using them to run recommendation algorithms often leads to wasted computing power. Now, cloud vendors like Google, Meta, and Alibaba are starting to customize their own ASIC chips:
Google’s TPU has iterated to the sixth generation, specifically optimized for large model training;
Alibaba’s Pingtouge launched AI inference chips in 2019, planning to invest 380 billion in cloud hardware over the next three years;
Meta’s MTIA v2 chip has half the computing power of the A100 but higher utilization, specifically designed for running advertising recommendation models.
Data predicts that the global ASIC market will only be $6.6 billion in 2023, but it will surge to $55.4 billion by 2028 — an 8-fold increase, representing a definite “incremental blue ocean”.
2. Chip “Heating” Creates Liquid Cooling Necessity: Perfluoropolyether Market Nears $30 Billion
As AI chips become more powerful, their power consumption has also “exploded”: NVIDIA’s GB200 single chip exceeds 1000W, with a single cabinet power consumption exceeding 132KW, and the Rubin series in 2026 is expected to exceed 600KW — traditional air cooling simply cannot handle it.
At this point, liquid cooling technology has become a “lifesaver”, especially “immersion liquid cooling” (submerging chips in a special liquid). The choice of liquid is crucial, with “perfluoropolyether/perfluoramine” being the top performer: it is insulating, non-flammable, and can be used for over 10 years, with a potential market space of nearly 30 billion yuan — larger than the markets for silicone oil and hydrofluoroether, which is why related companies’ stock prices have surged recently.
3. Ecosystem Breakthrough: 5 Million CUDA Developers, How Can Domestic Chips Break Through?
NVIDIA’s “moat” is not the chip hardware but the CUDA ecosystem — over 5 million developers worldwide are using it, with a rich library of algorithms and toolchains. Switching to other chips incurs high costs just for code modification.
For domestic chips to break through, they must take the path of “coordinated innovation”:
“Chip – Model Coordination”: The Step 3 large model from Juyuan Star considers the characteristics of Huawei Ascend during design, achieving inference efficiency three times that of DeepSeek-R1;
“Cross-Manufacturer Alliance”: Ten manufacturers, including Huawei, Muxi, and Birun, have formed the “Model-Chip Ecosystem Alliance” to unify adaptation standards and avoid each developing their own systems;
“Framework – Hardware Closed Loop”: The Wuyuan Qiong platform connects domestic models, systems, and chips, specifically optimizing the DeepSeek model to thin down the barriers of CUDA.
5. Outlook: SWOT Analysis + Three Major Trends, Where Are the Opportunities for Domestic Chips?
Finally, let’s use the SWOT analysis method to objectively assess the “landscape” of China’s AI chips:
Strengths (S): The largest digital economy market in the world, full policy support, and a complete industrial chain foundation;
Weaknesses (W): Dependence on imports for EDA tools and high-end equipment, weak software ecosystem, and performance gaps in high-end chips;
Opportunities (O): Generative AI drives demand for computing power, domestic substitution opens up markets, and autonomous driving and the metaverse require specialized chips;
Threats (T): Technological blockades continue, chip R&D requires significant investment and long cycles, and there is a risk of being outpaced by iterations.
In the next three years, three major trends will determine success or failure.
From “computing power competition” to “ecosystem competition”: having strong chip performance is not enough; there must be supporting software and developers — whoever builds the “domestic version of CUDA” first will gain the advantage;
Heterogeneous computing + Chiplet will become mainstream: the performance of a single chip has reached its peak, and in the future, CPU, GPU, and ASIC will be integrated together (heterogeneous computing), or Chiplet technology will be used to “piece together chips” to balance performance and cost;
Green and low-carbon will become hard indicators: data center energy consumption is now under strict control, and AI chips must not only have strong computing power but also a “high energy efficiency ratio” — low-power designs will penetrate from edge devices (like smartphones) to the cloud.
Industry Summary: Four Core Conclusions for AI Chips in 2025
Clear market increment: By 2029, the scale of China’s AI chips will reach 1.3 trillion, with a CAGR exceeding 56%, and inference chips will account for 73%, becoming the absolute “main track”;
Domestic substitution enters the “acceleration phase”: market share will rise from less than 20% in 2022 to 40% by 2025, but there is still a 1-2 generation gap in high-end training chips, requiring breakthroughs in ecosystem and energy efficiency;
Three sub-directional areas are most worthy of attention: cloud vendors customizing ASIC (8-fold growth in 5 years), immersion liquid cooling (perfluoropolyether market nearing 30 billion), and domestic ecosystem collaboration (chip – model – framework closed loop);
The ecosystem is the “last hurdle”: the 5 million CUDA developers were not built in a day; domestic chips need to accumulate developers through open-source, alliances, and compatibility — this is more challenging than hardware development but equally important.
The AI chip sector is like a “marathon sprint”: it requires not only current performance but also long-term ecological planning; it must withstand external pressures while seizing internal demands. By 2025, the trillion market door has already opened; whether domestic manufacturers can leverage this opportunity to “break through” remains to be seen.