Technical Deep Dive
The valuation explosion of Zhipu AI and Cambricon is rooted in deep technical moats that few competitors can replicate. Zhipu AI's core advantage lies in its GLM (General Language Model) architecture, which has evolved through multiple generations. Unlike the dense transformer architectures used by many peers, Zhipu has invested heavily in mixture-of-experts (MoE) designs that dynamically activate only relevant sub-networks for each query. This approach dramatically reduces inference cost while maintaining or improving accuracy. In their latest GLM-5 model, Zhipu reportedly achieved a 4x reduction in per-token inference cost compared to GPT-4-class models, while scoring competitively on benchmarks like MMLU (88.2) and HumanEval (82.7). The company has also open-sourced key components, including the ChatGLM series on GitHub (repository 'THUDM/ChatGLM-6B', now with over 45,000 stars), which has fostered a vibrant developer ecosystem and accelerated enterprise adoption.
On the hardware side, Cambricon's technical journey has been arduous but strategically sound. Their MLU370 and upcoming MLU590 chips are designed specifically for AI inference workloads, not general-purpose computing. This specialization allows for higher throughput per watt for transformer models. Cambricon's Bang architecture uses a unique systolic array design optimized for matrix operations common in LLMs. Recent benchmarks from the company show their MLU590 achieving 85% of NVIDIA A100 inference throughput on GPT-3-scale models at 60% of the power consumption. The company has also developed a software stack, Cambricon Neuware, that supports PyTorch and TensorFlow, lowering the barrier for developers migrating from CUDA.
| Model/System | Parameters | MMLU Score | Inference Cost (per 1M tokens) | Power Efficiency (TFLOPs/W) |
|---|---|---|---|---|
| Zhipu GLM-5 (MoE) | ~130B active / 1T total | 88.2 | $0.80 | 2.1 |
| GPT-4o (dense) | ~200B (est.) | 88.7 | $5.00 | 1.0 |
| Llama 3.1 405B (dense) | 405B | 87.3 | $3.50 | 0.9 |
| DeepSeek-V2 (MoE) | ~21B active / 236B total | 78.5 | $0.50 | 2.5 |
Data Takeaway: Zhipu's MoE architecture delivers near-GPT-4o-level accuracy at 84% lower inference cost, a critical advantage for enterprise scaling. This cost efficiency is a primary driver of their revenue growth and valuation.
Key Players & Case Studies
Zhipu AI's rise is inseparable from its strategic partnerships. The company has deployed its models across major Chinese enterprises, including state-owned banks, insurance companies, and telecom operators. A notable case is China Merchants Bank, which integrated Zhipu's GLM for customer service automation, reducing response time by 60% and handling 40% of queries without human intervention. Another is Sinopec, using Zhipu models for safety monitoring and document analysis across its refineries.
Cambricon's trajectory is equally instructive. The company has secured design wins with several major cloud providers, including Alibaba Cloud and Tencent Cloud, for AI inference acceleration. Their chips are also used in edge computing scenarios, such as smart surveillance cameras from Hikvision and autonomous driving systems from Baidu's Apollo platform. The company's revenue mix has shifted from 90% government-funded research projects in 2020 to over 70% commercial sales in 2025, reflecting genuine market traction.
| Company | Product | Market Cap (CNY) | Primary Revenue Driver | Key Customer |
|---|---|---|---|---|
| Zhipu AI | GLM series, ChatGLM | ~1.0 trillion | Enterprise LLM subscriptions & API | China Merchants Bank, Sinopec |
| Cambricon | MLU370/590, Neuware | ~900 billion | AI inference chips & software | Alibaba Cloud, Baidu |
| Baidu | ERNIE Bot | ~400 billion | Consumer & enterprise AI | Multiple SMBs |
| SenseTime | Large vision models | ~250 billion | Computer vision & edge AI | Government, automotive |
Data Takeaway: Zhipu and Cambricon command a combined market cap nearly 3x that of Baidu's AI unit, despite Baidu's longer history. This premium reflects investor belief in pure-play AI companies over diversified tech conglomerates.
Industry Impact & Market Dynamics
The valuation surge of Zhipu and Cambricon is reshaping the competitive landscape in several profound ways. First, it validates the 'AI-native' company model—firms built from the ground up around AI, rather than legacy tech companies adding AI as a feature. This is attracting massive capital: in 2025 alone, Chinese AI startups raised over $15 billion, with Zhipu accounting for $3 billion of that. Second, it is driving a 'chip independence' narrative in China, as Cambricon's success encourages other domestic chip designers to pivot toward AI-specific architectures. Third, it is accelerating enterprise AI adoption: companies that were hesitant to commit to AI are now seeing clear ROI examples from Zhipu's deployments.
| Metric | 2023 | 2024 | 2025 (est.) |
|---|---|---|---|
| Chinese AI market size (USD) | $45B | $68B | $95B |
| Enterprise LLM adoption rate | 12% | 28% | 45% |
| AI chip domestic market share | 8% | 15% | 25% |
| Zhipu AI annual revenue (CNY) | ¥1.2B | ¥4.5B | ¥12B |
| Cambricon annual revenue (CNY) | ¥0.8B | ¥2.1B | ¥5.5B |
Data Takeaway: The market is growing at 40%+ CAGR, and both Zhipu and Cambricon are outpacing that growth, indicating they are gaining market share. Their valuations imply revenue multiples of 80-160x, which is high but not unprecedented for platform companies at this stage of adoption.
Risks, Limitations & Open Questions
Despite the euphoria, significant risks remain. Zhipu's valuation implies expectations of sustained hypergrowth, but competition is intensifying. ByteDance, Alibaba, and Tencent are all investing heavily in their own LLMs, and price wars are already emerging. Zhipu's inference cost advantage may erode as competitors adopt similar MoE techniques. Moreover, regulatory uncertainty looms: China's AI regulations are still evolving, and any tightening could limit Zhipu's ability to monetize certain use cases, particularly in sensitive sectors like healthcare and finance.
Cambricon faces its own set of challenges. The company is still heavily dependent on TSMC for advanced chip manufacturing, a geopolitical vulnerability. Export controls on semiconductor equipment could disrupt their supply chain. Additionally, NVIDIA's dominance in AI training chips means Cambricon is largely confined to the inference market, which has lower margins. The company's gross margins have actually declined from 65% in 2022 to 52% in 2025 as they compete on price.
An open question is whether these valuations are sustainable. The current market cap of Zhipu and Cambricon combined exceeds the entire Chinese AI software market size in 2024. This suggests that investors are pricing in not just current leadership but a winner-takes-most outcome. If the market fragments—which is likely given the number of well-funded competitors—these valuations could correct sharply.
AINews Verdict & Predictions
Verdict: The AI wealth era is real, but it is not a rising tide that lifts all boats. Zhipu and Cambricon have earned their valuations through genuine technical differentiation and commercial execution. However, the market is pricing in perfection, and the path from here is narrower than the path from there.
Predictions:
1. Within 12 months, Zhipu AI will face its first major competitive test as ByteDance launches a free-tier LLM that matches GLM-5 on accuracy. This will compress Zhipu's margins and force a strategic pivot toward vertical-specific models (e.g., legal, medical).
2. Within 18 months, Cambricon will announce a partnership with a major Chinese automaker for in-vehicle AI chips, diversifying beyond cloud inference and boosting revenue by 30%.
3. Within 24 months, at least one of these two companies will experience a 30%+ correction from its peak, as the market realizes that the total addressable market for AI in China, while large, is not infinite and competition will cap margins.
4. The most important metric to watch is not revenue growth but gross margin stability. If Zhipu can maintain gross margins above 60% while scaling, the trillion-yuan valuation will look cheap in hindsight. If margins dip below 40%, a correction is inevitable.
What to watch next: The next battleground will be AI agents—autonomous systems that can execute multi-step tasks. Zhipu has already previewed an agent framework, and Cambricon is developing a chip optimized for agent workloads. The company that wins in agents will likely be the first to reach a $2 trillion valuation.