Technical Deep Dive
The crash of Zhipu and MiniMax's Hong Kong shares is not merely a market sentiment event; it reflects a structural mismatch between the capital-intensive nature of large language model development and the liquidity preferences of public equity markets. To understand why, we must examine the underlying economics of training and inference.
Both companies operate on a transformer-based architecture, with Zhipu's GLM series and MiniMax's MiniMax-01 model reportedly exceeding one trillion parameters. Training such models requires clusters of thousands of NVIDIA H100 or domestic equivalent accelerators. At current cloud rental rates of approximately $2.5 per GPU-hour for H100s, a single training run can cost $10–$20 million. Inference costs are equally punishing: serving a 100B+ parameter model to millions of users can consume $0.01–$0.03 per API call, far exceeding the $0.001–$0.005 per token charged to customers.
A key technical differentiator is the attention mechanism. Zhipu has open-sourced its GLM-130B and ChatGLM series, which use a bidirectional attention variant optimized for Chinese-language tasks. MiniMax, meanwhile, has invested heavily in a hybrid architecture combining sparse and dense attention to reduce inference latency. However, neither has publicly demonstrated a breakthrough in quantization or pruning that would dramatically lower cost-per-token.
| Model | Parameters (est.) | Training Cost (est.) | Inference Cost per 1M tokens | MMLU Score (Chinese subset) | Open Source?
|---|---|---|---|---|---|
| Zhipu GLM-4 | ~1T | $15M | $3.50 | 86.2 | Partial (ChatGLM-6B) |
| MiniMax-01 | ~1T | $18M | $4.00 | 85.8 | No |
| DeepSeek-V2 | ~236B | $5M | $0.50 | 88.5 | Yes |
| Qwen2.5-72B | 72B | $2M | $0.20 | 87.3 | Yes |
Data Takeaway: The table reveals a stark efficiency gap. DeepSeek and Qwen, with smaller but highly optimized models, achieve comparable or better Chinese-language benchmarks at a fraction of the training and inference cost. Zhipu and MiniMax are caught in a 'bigger is better' arms race that is not translating into proportional revenue.
Key Players & Case Studies
The dual-listing saga involves three key groups: the AI companies themselves, the Hong Kong exchange (HKEX), and the A-share market (Shanghai/Shenzhen).
Zhipu AI (Beijing) was founded by Tang Jie, a Tsinghua professor, and has raised over $2 billion from investors including Alibaba, Tencent, and state-backed funds. Its flagship product, ChatGLM, competes directly with Baidu's ERNIE Bot and ByteDance's Doubao. Despite strong brand recognition, its enterprise API business has grown slower than expected, with annualized recurring revenue (ARR) estimated at only $80–$100 million — a fraction of its valuation.
MiniMax (Shanghai) was founded by Yan Junjie, a former SenseTime executive. It has raised approximately $1.5 billion, with backing from Hillhouse Capital, IDG, and Tencent. Its product, Hailuo AI, focuses on both text and video generation. MiniMax's video generation model, released in early 2025, received mixed reviews compared to Kuaishou's Kling and OpenAI's Sora. Its enterprise revenue is estimated at $50–$70 million ARR.
| Company | Total Funding | Est. Valuation (pre-crash) | Est. ARR (Enterprise) | Primary Product | Key Investor
|---|---|---|---|---|---|
| Zhipu AI | $2.1B | $12B | $80–100M | ChatGLM, GLM-4 API | Alibaba, Tencent |
| MiniMax | $1.5B | $8B | $50–70M | Hailuo AI (text+video) | Hillhouse, Tencent |
| Baidu (ERNIE Bot) | Public | $35B (market cap) | $300–400M (est.) | ERNIE Bot | Public |
| ByteDance (Doubao) | Private | $268B | $500M+ (est.) | Doubao | Private |
Data Takeaway: Both Zhipu and MiniMax are significantly smaller in revenue than their state-backed or Big Tech competitors. Their high valuations relative to revenue (120x–160x ARR) were justified by growth expectations that are now being questioned.
Industry Impact & Market Dynamics
The Hong Kong stock crash sends a powerful signal to the entire Chinese AI ecosystem. It suggests that the window for 'narrative-driven' IPOs on HKEX is closing. Hong Kong investors, burned by the post-IPO declines of companies like SenseTime and Megvii, are now demanding clear paths to profitability.
This is reshaping the competitive landscape. Companies that can demonstrate strong B2B revenue — such as iFlytek, which focuses on education and government contracts — are being rewarded. Those relying on consumer subscriptions or API calls are being penalized.
| Metric | HKEX AI Stocks (2025) | A-Share AI Concept Stocks (2025) |
|---|---|---|
| Average P/S Ratio | 8x | 25x |
| Average P/E Ratio | 45x (loss-making excluded) | 60x (includes loss-making) |
| Median Revenue Growth YoY | 35% | 50% |
| % of Companies Profitable | 30% | 15% |
Data Takeaway: A-shares offer a significantly higher valuation multiple for AI concept stocks, even for loss-making companies. This explains the appeal of a dual listing. However, the A-share market is also more volatile and subject to regulatory whims. The 'new quality productive forces' narrative can shift quickly.
Risks, Limitations & Open Questions
1. Profitability Trap: Neither company has a clear timeline to profitability. Zhipu's gross margins on API calls are estimated at 40–50%, but after accounting for R&D and sales costs, net margins are deeply negative (estimated at -80% to -120%).
2. Commoditization: The Chinese LLM market is becoming a race to the bottom on price. ByteDance's Doubao and Baidu's ERNIE Bot are offering free tiers, squeezing smaller players.
3. Regulatory Risk: A-share listings require approval from the China Securities Regulatory Commission (CSRC), which has recently tightened rules for tech companies with 'variable interest entity' (VIE) structures. Both Zhipu and MiniMax use VIEs, adding execution risk.
4. Talent Retention: The stock crash has wiped out significant employee equity value, raising the risk of key engineer departures to competitors or overseas.
5. Open Source Threat: DeepSeek and Qwen are open-sourcing high-performing models, making it harder for Zhipu and MiniMax to charge premium prices for their APIs.
AINews Verdict & Predictions
Verdict: The dual-listing strategy is a high-risk gamble that reveals a fundamental weakness in the business models of Zhipu and MiniMax. They are caught between two incompatible investor bases: Hong Kong demands profits, A-shares demand stories. Currently, they can credibly deliver neither.
Predictions:
1. Within 6 months: Both companies will release 'killer' AI agent products — likely integrated with WeChat or enterprise SaaS platforms — to try to demonstrate user engagement and revenue acceleration. If these fail to gain traction, share prices will fall another 30–40%.
2. Within 12 months: At least one of the two will abandon the Hong Kong primary listing entirely and go private via a management buyout, seeking a full A-share IPO at a higher valuation.
3. Within 18 months: The Chinese government will introduce a 'strategic AI listing fast-track' for A-shares, specifically designed to rescue struggling unicorns like Zhipu and MiniMax, but with strings attached regarding state ownership and data localization.
4. Wildcard: If ByteDance or Alibaba makes a strategic acquisition offer for either company, the dual-listing plan will be abandoned. The most likely target is MiniMax, due to its video generation capabilities.
What to watch: The next quarterly earnings calls. If Zhipu and MiniMax can show sequential revenue growth above 20% and a narrowing of net losses, the narrative may shift. If not, the 'trust crisis' will deepen into a full-blown solvency crisis.