Technical Deep Dive
The valuation reset for Zhipu AI and MiniMax is rooted in a fundamental technical reality: the marginal value of raw model performance is declining. For the past two years, the primary competitive metric was benchmark scores—MMLU, HumanEval, GSM8K. Companies raced to publish state-of-the-art results, and investors rewarded them accordingly. However, the gap between top-tier models is now razor-thin.
| Model | MMLU (5-shot) | HumanEval (Pass@1) | GSM8K (8-shot) | Inference Cost (per 1M tokens) |
|---|---|---|---|---|
| GPT-4o | 88.7 | 90.2 | 96.8 | $5.00 |
| Claude 3.5 Sonnet | 88.3 | 92.0 | 96.4 | $3.00 |
| Zhipu GLM-4 | 87.2 | 85.1 | 93.5 | $1.80 |
| MiniMax-01 | 86.5 | 83.4 | 92.1 | $1.50 |
| DeepSeek-V2 | 88.5 | 90.0 | 95.8 | $0.50 |
Data Takeaway: While Zhipu and MiniMax are competitive, they are not leaders on any single benchmark. The real story is the cost gap: DeepSeek-V2, an open-source model, offers comparable performance at a fraction of the cost. This commoditization means that proprietary models must justify their premium through integration, data, and service—not just intelligence.
The technical moat has shifted from model architecture to data infrastructure. Zhipu AI, spun out of Tsinghua University, has a unique advantage in knowledge graphs (KGs). Their GLM series is not just a transformer; it is deeply integrated with structured knowledge bases that allow for more factual, traceable reasoning. This is particularly valuable in regulated industries like finance and healthcare, where hallucinations are unacceptable. The GitHub repository for their KG toolkit, `zhipuai/kg-construction`, has seen steady growth (currently ~2,300 stars), providing a practical tool for enterprise customers to build custom knowledge bases.
MiniMax, on the other hand, has bet heavily on multimodal generation, particularly video and audio. Their architecture uses a hybrid of sparse and dense attention mechanisms to handle long-context video inputs efficiently. Their open-source text-to-video model, `MiniMax/text-to-video`, has garnered over 4,500 stars on GitHub and is used by a community of indie creators. However, the technical challenge here is twofold: consistency (maintaining character identity across frames) and cost (generating high-resolution video is computationally expensive). The moat is not the model itself but the fine-tuning pipeline and the proprietary dataset of licensed video content they have curated.
Key Technical Takeaway: The battle is no longer about who has the best model. It is about who has the best data flywheel—the ability to collect high-quality, domain-specific data from real user interactions, and use it to continuously improve a specialized model that competitors cannot easily replicate.
Key Players & Case Studies
Zhipu AI: The Enterprise Incumbent
Zhipu AI’s strategy is to embed itself into the IT infrastructure of large state-owned enterprises (SOEs) and financial institutions. Their flagship product, the GLM-4 Enterprise Edition, is sold as a suite that includes a private deployment option, a knowledge graph builder, and a compliance auditing tool. This is a classic enterprise moat: high switching costs, long sales cycles, and deep integration with legacy systems. A notable case is their deployment at the Industrial and Commercial Bank of China (ICBC), where GLM-4 powers an internal risk assessment system that analyzes loan applications against regulatory documents. Replacing this system would require retraining on years of proprietary data and re-certification by regulators—a multi-year, multi-million dollar process.
MiniMax: The Creator Economy Play
MiniMax has taken a different route, targeting the consumer and creator markets. Their Hailuo AI platform allows users to generate short videos from text prompts. The monetization model is a freemium subscription, with a Pro tier at $19.99/month for 4K resolution and longer clips. The moat here is network effects: as more creators use Hailuo, they generate a library of user-created content that serves as both a showcase and a training dataset. However, this moat is weaker than Zhipu’s. The creator market is fickle, and competitors like Kuaishou’s Kling and ByteDance’s Jimeng are offering similar features, often at lower prices.
| Company | Primary Moat | Key Customer Segment | Revenue Model | Estimated 2024 Revenue |
|---|---|---|---|---|
| Zhipu AI | Enterprise integration, knowledge graphs, regulatory compliance | SOEs, banks, government | Per-seat licensing + API usage | $80-120M (est.) |
| MiniMax | Multimodal generation, creator community | Content creators, SMEs | Subscription + API | $30-50M (est.) |
| Baidu (ERNIE) | Search ecosystem, cloud infrastructure | Broad enterprise | Cloud credits + API | $500M+ (est.) |
| DeepSeek | Open-source, low cost | Developers, startups | API (low margin) | $10-20M (est.) |
Data Takeaway: Zhipu AI’s revenue is 2-3x that of MiniMax, and its customer base is stickier. However, both companies are dwarfed by Baidu’s ERNIE, which benefits from its existing cloud and search infrastructure. The A-share market will scrutinize not just revenue growth but gross margins and customer retention rates—areas where Zhipu likely has a clear advantage.
Industry Impact & Market Dynamics
The valuation reset for Zhipu and MiniMax is a microcosm of a broader shift in the global AI market. The era of 'infinite funding for infinite scaling' is over. Investors are now demanding a clear path to profitability, and the A-share market is particularly unforgiving. In 2023, Chinese AI startups raised over $5 billion in funding, with Zhipu and MiniMax accounting for roughly $1.5 billion combined. In 2024, that number is expected to drop by 40%, as investors pivot to later-stage, revenue-generating companies.
The impact on the competitive landscape will be profound. Companies that cannot demonstrate a durable moat will be forced to consolidate or shut down. We are already seeing this: several smaller LLM startups, such as Baichuan and 01.AI, have pivoted to vertical applications or are seeking acquisition. The survivors will be those that have built a defensible position in a specific vertical (e.g., Zhipu in finance) or have a unique technical capability that is hard to replicate (e.g., MiniMax in video generation, though this is more fragile).
Another key dynamic is the rise of open-source models. DeepSeek-V2, released in May 2024, achieved GPT-4-level performance at 1/10th the inference cost. This puts enormous pressure on proprietary API pricing. For Zhipu and MiniMax, the only way to compete is to offer a value proposition that goes beyond the model itself—integration, customization, data privacy, and compliance. This is exactly the playbook that Zhipu is executing, and it is why they are better positioned for an A-share listing than MiniMax.
Market Data Takeaway: The Chinese AI market is projected to grow from $20 billion in 2024 to $60 billion by 2028 (CAGR of 31%). However, the growth will be concentrated in enterprise AI (48% CAGR) rather than consumer AI (22% CAGR). This favors Zhipu’s enterprise-first strategy over MiniMax’s consumer focus.
Risks, Limitations & Open Questions
For Zhipu AI:
- Over-reliance on SOEs: While sticky, SOE sales cycles are long and dependent on government policy. A shift in regulatory priorities could slow revenue growth.
- Talent Retention: As a spin-off from Tsinghua, Zhipu has a strong research team, but the best researchers are often lured by higher salaries at Alibaba, Tencent, or overseas. Maintaining technical leadership is expensive.
- Open-Source Threat: If a model like DeepSeek-V2 becomes the de facto standard for enterprise deployments, Zhipu’s proprietary advantage could erode.
For MiniMax:
- Commoditization of Video Generation: The gap between MiniMax’s output and that of free, open-source alternatives (e.g., Stable Video Diffusion) is narrowing rapidly. The premium for 'slightly better' video is hard to sustain.
- High Burn Rate: Video generation is computationally expensive. At current API prices, MiniMax may be losing money on every inference. A price war with ByteDance or Kuaishou could be devastating.
- Regulatory Risk: The Chinese government has strict rules on AI-generated content, especially video. Any compliance misstep could shutter the platform.
Open Question: Can either company achieve a gross margin above 60% within two years? This is the threshold that A-share investors typically demand for tech companies. Zhipu, with its high-margin software licensing, has a better chance than MiniMax, which is essentially selling compute cycles.
AINews Verdict & Predictions
Verdict: The 'July Siege' is a healthy correction. The market is right to demand moats over stories. Zhipu AI is better positioned for a successful A-share IPO, with a valuation in the $5-7 billion range (down from its last private valuation of $10 billion). MiniMax will face a harder road, likely needing to down-round to $2-3 billion and pivot to a more defensible niche, such as enterprise video production tools for marketing teams.
Predictions:
1. By Q4 2025, Zhipu AI will file its IPO prospectus, revealing a gross margin of 55-60% and a path to profitability by 2026. The stock will be a moderate performer, trading at a 30-40% discount to Baidu’s AI business on a revenue multiple basis.
2. MiniMax will not IPO in 2025. Instead, it will raise a bridge round at a flat or down valuation, and use the capital to acquire a smaller competitor in the enterprise video space (e.g., a company like Vidu) to build a moat.
3. The broader implication: The A-share market will become a graveyard for AI companies that cannot demonstrate a clear, defensible business model. The next wave of AI IPOs will be from vertical SaaS companies that use LLMs as a feature, not a product.
What to watch next: The key metric is not benchmark scores but 'Net Dollar Retention' (NDR) for enterprise customers. If Zhipu can show NDR above 120%, it will signal that their moat is real. For MiniMax, watch 'Monthly Active Creators' and 'Revenue per Creator'—if these stagnate, the story is over.