Technical Deep Dive
The four IPOs represent distinct technical frontiers in AI, each with its own architectural challenges and engineering trade-offs.
Xizhi Technology (Optical Computing): Xizhi is pioneering photonic computing to break the von Neumann bottleneck. Traditional electronic chips suffer from data movement latency and power dissipation—moving data between memory and processor can consume 100x more energy than the computation itself. Xizhi's approach uses silicon photonics to perform matrix multiplications directly in the optical domain. Their core architecture, detailed in their IPO prospectus, uses Mach-Zehnder interferometers (MZIs) arranged in a mesh network to implement neural network layers. Light passes through these interferometers, and the interference patterns encode the weights and activations. The key advantage: optical signals can carry multiple wavelengths simultaneously (wavelength-division multiplexing), enabling parallel computation at the speed of light. Early benchmarks from their internal testing show energy efficiency of 10 peta-operations per second per watt (POPS/W), compared to 1-2 POPS/W for leading electronic accelerators like NVIDIA's H100. However, the technology is still pre-revenue, with their first commercial product—a co-packaged optics module for data center interconnects—not expected until Q1 2027. The GitHub repository 'Xizhi-Photonic-Network' (currently 2,300 stars) provides open-source simulation tools for their MZI mesh architecture, though the actual fabrication processes remain proprietary.
Zhipu AI (Large Language Models): Zhipu's GLM series is the flagship. The GLM-130B model, released in 2023, was one of the first open-source models to rival GPT-3 in scale. Their latest, GLM-5 (not yet publicly detailed), reportedly uses a hybrid MoE (Mixture of Experts) architecture with 1.2 trillion total parameters and 200 billion activated per token. The training infrastructure is built on a custom cluster of 10,000+ Huawei Ascend 910B chips, a critical detail given US export restrictions on NVIDIA's high-end GPUs. Zhipu has published several papers on their training techniques, including 'GLM: General Language Model Pretraining with Autoregressive Blank Infilling' and a recent preprint on 'Efficient MoE Training with Dynamic Expert Routing.' Their open-source repository 'GLM-130B' has accumulated over 45,000 stars on GitHub and is widely used by the Chinese developer community. On the C-Eval benchmark (a comprehensive Chinese language understanding benchmark), GLM-4 achieves 82.3%, compared to GPT-4's 84.1% and Claude 3.5's 83.5%. The gap is narrowing, but the cost of inference for GLM-4 is approximately $0.80 per million tokens, versus $3.00 for GPT-4o, giving Zhipu a significant price advantage in the Chinese market.
Kling AI (Video Generation): Kling's model is a diffusion-transformer hybrid that generates 1080p video at 30fps for up to 2 minutes. The architecture uses a 3D U-Net with temporal attention layers, trained on a proprietary dataset of 500 million video clips. Their key innovation is a 'motion consistency module' that reduces temporal flickering—a common failure mode in video generation. Kling's inference pipeline runs on a cluster of 2,000 NVIDIA A100 GPUs, with a generation latency of 90 seconds for a 30-second clip. The GitHub repository 'Kling-Video-Diffusion' (8,700 stars) provides a simplified inference script, though the full model weights are not publicly available.
World Model Company (unnamed): This company focuses on 'world models' for embodied AI—systems that can simulate physics, object permanence, and causal relationships. Their architecture is based on a recurrent state-space model (RSSM) similar to DeepMind's DreamerV3, but scaled to 10 billion parameters. The model is trained on 100 million hours of egocentric video data from robots and drones. The key metric is 'planning accuracy' on the Habitat 3.0 benchmark, where they achieve 78.4% success rate on long-horizon tasks (e.g., 'navigate to kitchen, open fridge, retrieve bottle'), compared to 72.1% for the previous state-of-the-art.
Data Table: Benchmark Performance Comparison
| Company | Technology | Key Benchmark | Score | Competitor Score | Cost/Unit |
|---|---|---|---|---|---|
| Xizhi | Optical Computing | POPS/W (Energy Efficiency) | 10 | 1.5 (NVIDIA H100) | N/A (Pre-revenue) |
| Zhipu AI | LLM (GLM-4) | C-Eval | 82.3% | 84.1% (GPT-4o) | $0.80/1M tokens |
| Kling AI | Video Generation | FVD (Fréchet Video Distance) | 45.2 | 52.8 (OpenAI Sora) | $0.12/1s video |
| World Model Co. | World Model | Habitat 3.0 Success Rate | 78.4% | 72.1% (DreamerV3) | N/A (Research) |
Data Takeaway: Zhipu AI is the closest to commercial parity, with competitive benchmark scores at a fraction of the cost. Xizhi and the world model company are still in research phases, while Kling leads in video quality but faces high inference costs.
Key Players & Case Studies
Xizhi Technology: Founded by Dr. Li Wei, a former Bell Labs researcher, Xizhi has raised $420 million across three rounds from investors including Sequoia China and Hillhouse Capital. Their strategy is to first enter the data center interconnect market (a $15 billion TAM) with optical transceivers before moving to full optical computing. The risk: Intel and NVIDIA are also investing heavily in photonics.
Zhipu AI: Led by CEO Zhang Peng, Zhipu has raised over $1.2 billion, with a post-IPO valuation of $8 billion. Their business model is a mix of API access (charging enterprises per token), model fine-tuning services, and a premium subscription for their ChatGLM consumer app, which has 20 million monthly active users. The key competitive threat is Baidu's Ernie Bot and Alibaba's Tongyi Qianwen, both of which have larger user bases and deeper integration with existing ecosystems.
Kling AI: A spin-off from Kuaishou, Kling has raised $300 million. Their go-to-market strategy targets content creators and advertising agencies, offering a SaaS platform for generating short-form video ads. The unit economics are challenging: each 30-second clip costs $3.60 in compute, while they charge $2.00—a deliberate loss-leader strategy to build market share.
World Model Company: Backed by Xiaomi and SenseTime, this company is targeting robotics and autonomous driving. Their first product is a simulation environment for training robot policies, priced at $50,000 per year per license. The addressable market is small but growing.
Data Table: Funding & Valuation Comparison
| Company | Total Funding | Post-IPO Valuation | Revenue (2025) | Employees |
|---|---|---|---|---|
| Xizhi Technology | $420M | $1.8B | $0 | 450 |
| Zhipu AI | $1.2B | $8B | $180M | 1,200 |
| Kling AI | $300M | $1.2B | $15M | 300 |
| World Model Co. | $250M | $900M | $2M | 200 |
Data Takeaway: Zhipu AI is the only company with meaningful revenue, yet its valuation-to-revenue multiple of 44x is extreme compared to established tech companies (Microsoft trades at ~10x revenue). This is the narrow-gauge effect in action—scarcity inflates multiples.
Industry Impact & Market Dynamics
The 18C chapter IPOs have created a new asset class: 'AI pure plays' in Hong Kong. Before these listings, investors seeking exposure to Chinese AI had to buy into larger tech conglomerates (Baidu, Alibaba, Tencent) where AI was a fraction of the business. Now, they can directly bet on specific AI sub-sectors. This has driven a rotation of capital: the Hang Seng Tech Index has seen a 15% increase in trading volume since the first IPO, with a significant portion attributed to AI-focused funds.
The narrow-gauge mechanism also acts as a signaling device. A successful IPO with a high first-day pop validates the technology narrative, attracting more venture capital to the sector. Since January, Chinese AI startups have raised $2.3 billion, a 40% increase year-over-year. However, the flip side is that a failed IPO or a post-listing crash could have a chilling effect. The lock-up expirations (starting in July 2026 for the first IPO) will be a critical test. If cornerstone investors dump their shares, the narrow float will amplify the sell-off.
Another dynamic is the 'race to list.' Other Chinese AI companies, including MiniMax, 01.AI, and Baichuan, are reportedly accelerating their own IPO plans to capitalize on the current window. This could lead to a glut of AI listings, diluting the scarcity premium.
Risks, Limitations & Open Questions
1. Commercialization Risk: Three of the four companies have negligible or negative revenue. The transition from research to product is notoriously difficult in AI. Zhipu faces intense competition from well-funded incumbents. Kling's loss-leader pricing is unsustainable without a path to lower costs or higher prices.
2. Technology Risk: Xizhi's optical computing may never achieve the scale needed for general-purpose AI acceleration. The world model company's technology is still unproven in real-world robotics deployments.
3. Regulatory Risk: The Chinese government's tightening control over AI model training and data usage could impact Zhipu and Kling. Export controls on advanced chips remain a threat to all four companies.
4. Market Structure Risk: The narrow-gauge mechanism creates extreme volatility. A 10% drop in the broader market could trigger a 30-50% decline in these stocks due to the thin float. Retail investors, who have piled into these names, are most exposed.
5. Valuation Risk: At current prices, these companies are pricing in years of future growth. Any disappointment—a missed product launch, a competitor's superior model, or a macroeconomic slowdown—could lead to a severe correction.
AINews Verdict & Predictions
Prediction 1: Zhipu AI will be the first to achieve sustainable profitability, but its valuation will correct by 30-40% within 12 months. The company has the strongest fundamentals, but the current multiple is unsustainable. As more AI models become available (including from open-source alternatives), pricing pressure will compress margins.
Prediction 2: Xizhi Technology will be acquired within 18 months. Optical computing is a long-term bet that requires deep pockets. A larger semiconductor company (e.g., TSMC or ASML) or a hyperscaler (e.g., Google or Microsoft) will likely acquire Xizhi for its IP and talent, offering a premium to current shareholders.
Prediction 3: Kling AI will pivot to a B2B licensing model or fail. The consumer video generation market is commoditizing rapidly. Kling's best path is to license its technology to media companies and advertising agencies, but this requires a sales force they currently lack.
Prediction 4: The world model company will become a key supplier to the Chinese robotics industry, but its IPO valuation will prove to be too high. The addressable market for world models is real but niche. Expect a 50% decline from peak valuation as investors realize the revenue trajectory is measured in millions, not billions.
Final Verdict: The 18C chapter has successfully created a pricing mechanism for China's AI narrative. But the narrow-gauge rails that enabled these spectacular IPOs will also amplify the pain when the narrative falters. The next 12 months will separate the genuine breakthroughs from the hype. Investors should watch cash burn rates, customer acquisition costs, and product roadmaps—not just stock price momentum.