Technical Deep Dive
The three companies represent fundamentally different architectural bets on AI value capture.
Zhipu AI has built its foundation on the GLM (General Language Model) architecture, which differs from the GPT-style decoder-only paradigm. GLM employs an autoregressive blank-filling objective that combines bidirectional attention for understanding with unidirectional generation—a hybrid that the team argues yields better performance on Chinese-language tasks and reasoning benchmarks. The company has open-sourced multiple model sizes, including ChatGLM-6B (6.2 billion parameters), which became a darling on GitHub (over 40,000 stars) for enabling local deployment. More recently, Zhipu released GLM-4, a 130B-parameter model that benchmarks competitively against GPT-4 on Chinese NLP tasks. The technical moat lies in their training infrastructure: they developed a custom distributed training framework called 'SwissArmyTransformer' (also open-source) that optimizes memory usage and communication overhead for large-scale model parallelism. However, the open-source strategy, while generating goodwill, creates a monetization tension: enterprises can self-host the free models, reducing the incentive to pay for API access.
Horizon Robotics takes a 'full-stack' approach. Its Journey series of automotive-grade AI chips (Journey 3, Journey 5, Journey 6) integrate a proprietary BPU (Brain Processing Unit) architecture optimized for computer vision and sensor fusion workloads. Unlike NVIDIA's general-purpose GPU approach, Horizon's BPU is a domain-specific architecture that trades flexibility for power efficiency and deterministic latency—critical for automotive safety. The Journey 5 chip delivers 128 TOPS (INT8) at 35W TDP, positioning it between Mobileye's EyeQ5 and NVIDIA's Orin. Horizon also provides a complete software stack, including the 'TogetherOS' middleware and 'OpenExplorer' model zoo, to reduce automakers' integration effort. The technical challenge is that this 'chip + algorithm' model requires simultaneous excellence in hardware design, compiler optimization, and algorithm development—a triple constraint that few teams globally have executed successfully.
Momenta is primarily a software-first company, focusing on a 'flywheel' approach: deploying a base-level ADAS (Advanced Driver Assistance System) to generate real-world driving data, which then trains higher-level autonomous driving functions. Their architecture uses a 'MonoNet' perception network that can handle camera-only or multi-modal sensor inputs, and a 'Planning' module based on imitation learning and reinforcement learning. Momenta's technical edge is its data engine—a closed-loop system that automatically mines corner cases from fleet data, labels them, and retrains models. This creates a virtuous cycle, but it also means Momenta's technical progress is directly proportional to the number of vehicles on the road running its software, making mass production partnerships existential.
| Company | Core Technology | Key Metric | Open-Source Repo (GitHub Stars) | Training/Inference Efficiency |
|---|---|---|---|---|
| Zhipu AI | GLM architecture (130B params) | MMLU: 82.3 (GLM-4), C-Eval: 72.1 | ChatGLM-6B (40k+ stars), SwissArmyTransformer (8k+ stars) | 4.2 TFLOPS/GPU (training), 0.8s per 1k tokens (inference, A100) |
| Horizon Robotics | BPU (Journey 5) | 128 TOPS @ INT8, 35W TDP | OpenExplorer (5k+ stars) | 3.2 TOPS/W (power efficiency) |
| Momenta | MonoNet + Data Engine | L4 disengagement rate: 0.2 per 1,000 km (test fleet) | Not open-source | N/A (proprietary) |
Data Takeaway: Zhipu's open-source strategy gives it the strongest developer mindshare but the weakest monetization lever. Horizon's chip-level efficiency advantage is real but narrow—NVIDIA's next-generation Drive Thor will likely surpass it. Momenta's closed-loop data engine is powerful but opaque, making it hard for investors to independently verify progress.
Key Players & Case Studies
Zhipu AI has secured over $1.5 billion in total funding from investors including Alibaba, Tencent, and state-backed funds. Its key product lines are: (1) the GLM API platform, targeting enterprise customers for content generation, customer service, and code assistance; (2) customized model fine-tuning for government and large state-owned enterprises; (3) the 'Zhipu Pu' platform for building AI-native applications. The company has announced partnerships with several Chinese banks and telecom operators, but contract sizes and renewal rates remain undisclosed. A critical case study is Zhipu's collaboration with a major Chinese bank to deploy a GLM-based assistant for internal document processing—the project took 9 months from contract to deployment, illustrating the long sales cycle inherent in enterprise LLM solutions.
Horizon Robotics went public on the Hong Kong Stock Exchange in October 2024, raising approximately $600 million at a valuation of around $5 billion—a significant discount from its peak private valuation of $8 billion. The IPO prospectus revealed that automotive chip revenue grew 67% year-over-year to $250 million, but the company remained deeply unprofitable, with a net loss of $180 million on that revenue. Key customers include BYD, FAW, and SAIC, with Horizon's Journey chips powering ADAS features in over 3 million vehicles as of early 2025. The company's strategy is to move up the value chain from L2 ADAS to L3/L4 systems, but this requires automakers to trust Horizon's full-stack solution over NVIDIA's ecosystem, which has a decade-long head start in developer tools and CUDA compatibility.
Momenta has raised over $1 billion from investors including SAIC Motor, Mercedes-Benz, and General Motors. The company's business model is to license its autonomous driving software stack to automakers on a per-vehicle basis. Momenta has secured production programs with SAIC, GAC, and a European premium automaker (widely believed to be Mercedes-Benz). However, the revenue recognition is back-loaded: Momenta receives upfront engineering fees during development, but the majority of revenue comes from per-vehicle royalties only after mass production begins. As of mid-2025, Momenta's software is deployed in fewer than 500,000 vehicles, meaning the royalty stream is still nascent. The company's valuation has been reported to be under pressure, with some secondary market transactions indicating a 30-40% discount from its peak.
| Company | Total Funding | Latest Valuation | 2024 Revenue (Est.) | Key Customers | Revenue Model |
|---|---|---|---|---|---|
| Zhipu AI | $1.5B | ~$10B (private) | $80-120M | Banks, telecoms, government | API calls, custom projects |
| Horizon Robotics | $2.3B (incl. IPO) | $5B (public) | $250M | BYD, FAW, SAIC | Chip sales, software licenses |
| Momenta | $1.0B | ~$3.5B (private) | $40-60M | SAIC, GAC, Mercedes-Benz | Engineering fees + per-vehicle royalties |
Data Takeaway: Horizon's public market valuation provides a reality check—its $5B market cap is roughly 20x revenue, a multiple that implies high growth expectations but also leaves little room for error. Zhipu and Momenta, still private, face the risk of significant down rounds if they cannot demonstrate similar or better revenue multiples.
Industry Impact & Market Dynamics
The capital retention crisis is reshaping the AI investment landscape. In 2021-2023, the dominant narrative was 'scale at all costs'—investors poured money into AI startups based on team quality and technical ambition, with little regard for near-term revenue. That era is over. The shift is driven by three factors: (1) rising interest rates have increased the cost of capital, making investors more risk-averse; (2) the public market reception of AI IPOs has been mixed, with companies like Horizon trading below their last private round; (3) the emergence of powerful open-source models (Llama, Mistral, Qwen) has commoditized base-level AI capabilities, compressing margins for API-based businesses.
The market is now segmenting into three tiers. Tier 1: companies with clear, defensible revenue streams and growing gross margins (e.g., NVIDIA, ASML). Tier 2: companies with strong technical moats but unproven commercial models (Zhipu, Horizon, Momenta). Tier 3: companies with neither technical differentiation nor revenue (most AI startups). The capital will concentrate in Tier 1 and the strongest Tier 2 players.
For Zhipu, the risk is that the LLM market becomes a commodity business. OpenAI's GPT-4o, Anthropic's Claude 3.5, and Google's Gemini are all competing on price and performance, driving down API costs. Zhipu's differentiation in Chinese-language tasks is real but eroding as global models improve in Chinese. The company's best path forward is to deepen its vertical-specific models for regulated industries (finance, healthcare, government) where data privacy and compliance requirements create switching costs.
Horizon's challenge is the NVIDIA ecosystem. NVIDIA's Drive platform is the default choice for automakers developing L3+ systems, and its CUDA ecosystem makes it difficult for automakers to switch to Horizon's BPU. Horizon's counter-argument is that its chips are more power-efficient and cost-effective for L2/L2+ ADAS, which represents the bulk of volume today. The data supports this: Horizon's Journey 5 costs approximately $100-150 per chip, compared to $200-400 for NVIDIA's Orin. For mass-market vehicles, this cost advantage is significant.
Momenta's fate is tied to the autonomous driving regulatory timeline. China has accelerated L3 testing permits, but mass-market L4 deployment remains years away. Momenta's per-vehicle royalty model means its revenue is a function of vehicle volumes times average selling price per vehicle. If L4 systems remain niche (luxury vehicles only), the royalty stream will be too small to justify the company's valuation.
| Market Segment | 2024 Market Size (China) | 2028 Projected Size | CAGR | Key Growth Driver |
|---|---|---|---|---|
| Enterprise LLM Services | $1.2B | $8.5B | 48% | Government digitalization, finance compliance |
| Automotive AI Chips (ADAS) | $1.8B | $5.2B | 24% | L2+ penetration from 30% to 70% of new cars |
| Autonomous Driving Software | $0.6B | $3.5B | 42% | L3 regulatory approval, robotaxi expansion |
Data Takeaway: The LLM services market is growing fastest, but it is also the most competitive with the lowest barriers to entry. The automotive AI chip market is growing more slowly but has higher margins and stronger customer lock-in. Autonomous driving software has the highest potential upside but the greatest regulatory and technical risk.
Risks, Limitations & Open Questions
Zhipu AI faces three critical risks. First, the commoditization of LLMs: if open-source models continue to improve at the current pace, enterprises may choose to self-host rather than pay for API access, compressing Zhipu's revenue potential. Second, the 'China discount': Chinese AI companies face restrictions on accessing cutting-edge hardware (NVIDIA H100/B200), which could widen the performance gap with US-based models. Third, the lack of a consumer product: unlike Baidu's Ernie Bot or ByteDance's Doubao, Zhipu has no direct-to-consumer channel, limiting its ability to generate user data and brand recognition.
Horizon Robotics must navigate the geopolitics of chip supply chains. While Horizon designs its chips in China, manufacturing is done at TSMC (Taiwan), creating potential supply chain vulnerabilities. Additionally, the company faces competition from domestic rivals like Black Sesame Technologies and SemiDrive, as well as international giants like Mobileye and NVIDIA. The open question is whether automakers will standardize on one chip platform or maintain multi-sourcing strategies, which would limit Horizon's ability to capture full-stack margins.
Momenta is exposed to the 'valley of death' in autonomous driving. The company must continue investing heavily in R&D and test fleets while waiting for mass production revenue to materialize. If automakers delay L4 deployment or choose to develop software in-house, Momenta's revenue projections could collapse. The company's reliance on a few key automaker partners creates concentration risk—losing one major contract could be existential.
AINews Verdict & Predictions
Prediction 1: Zhipu AI will be the first to face a down round. Despite its technical credibility, Zhipu's revenue base is too small and too slow-growing to justify a $10B valuation. The company will likely raise a 'down round' or 'flat round' in the next 12 months, forcing a valuation reset. This is not a death sentence—it will simply align expectations with reality.
Prediction 2: Horizon Robotics will survive and potentially thrive, but as a chip company, not a full-stack AI company. The market will force Horizon to focus on its core competency: efficient, cost-effective ADAS chips. The software stack will become a differentiator but not the primary revenue driver. Horizon's public market discipline will force it to prioritize profitability over growth, which may slow its expansion but improve its survival odds.
Prediction 3: Momenta will be acquired within 24 months. The company's technology is valuable, but its business model requires scale that it cannot achieve independently. A strategic acquirer—most likely SAIC Motor or another automaker—will absorb Momenta to internalize its autonomous driving capabilities. The acquisition price will be a fraction of its peak valuation, but it will provide an exit for investors and a path to deployment for the technology.
The broader takeaway: The AI industry is entering a 'survival of the fittest' phase where capital efficiency, not capital raised, is the new metric of success. Companies that can demonstrate a clear path to $100M+ in revenue with improving unit economics will attract capital. Those that cannot will be acquired or shut down. This is not pessimism—it is the natural maturation of a transformative technology.