Summary of Key Points
Moonlit Side has released the world's largest open-source model, K3 (with 2.8 trillion parameters), and has gained widespread attention in Silicon Valley due to its technical capabilities (even received a like from Elon Musk; Apple's AI chief called it an “open-source victory”). However, the company has had to temporarily suspend new user subscriptions on the consumer (C) side due to insufficient computing power. At the same time, there are reports that they plan to launch a Pre-IPO in August with a target valuation of $50 billion, a significant shift from their previous stance of “not being in a hurry to go public with billions in cash” seven months ago. The reasons behind this include soaring computing costs, tightening capital markets, and difficulties in their business model. Nevertheless, K3 has enabled Chinese AI to stand at the global forefront for the first time based on its technical strength, shedding the label of “inexpensive” and beginning to compete for value pricing rights.
Detailed Analysis
1. K3 Becomes a Global Sensation but Faces Computing Power Constraints
The success of K3 comes with substantial computational costs:
- Model complexity: Loading 2.8 trillion parameters into video memory requires nearly 3TB of space, and the “memory storage” (KV Cache) for each conversation increases exponentially, leading to additional costs with each request.
- High deployment barriers: The company recommends using at least 64 H800 chips (each of which costs over 150,000 yuan in China), resulting in hardware expenses in the tens of millions. Moreover, a single high-performance node can only serve a small number of users; to support tens of millions of daily active consumer users, the investment would need to increase significantly.
- Computing power shortage: The demand for AI computing power in China has increased by 417% year-on-year, while supply has only risen by 128%. There are limitations on the import of high-end chips, and the domestic chip ecosystem is not yet well-developed.
Therefore, Moonlit Side’s decision to limit consumer user access is not just a matter of attitude; it’s due to practical constraints. The more consumer users there are, the greater the financial loss (with heavy users incurring costs of over 59 yuan per month). They must prioritize computing power for their profitable business (B) side customers to avoid continuous losses.
2. From “Not Hurrying to Go Public with Billions” to Pursuing an IPO
Seven months ago, Yang Zhilin stated that they were not in a hurry to go public with billions in cash. Now, the urgency to launch an IPO is driven by practical reasons:
- High computing power costs: Maintaining K3’s leading position requires continuous investment in hardware and expanding clusters, which exceeds the amount of capital available.
- Lessons from predecessors: The stock prices of companies like Zhipu and MiniMax plummeted after their IPOs (by 70% and 85%, respectively), indicating that the market’s valuation logic for AI companies has changed. Investors now focus more on fundamentals (profitability prospects) rather than scarcity.
- Seizing the last opportunity: With K3 still gaining global attention, going public within six months might be a good time to join the AI industry’s growth trend. However, the target valuation of $50 billion is higher than that of their predecessors, and it remains to be seen whether investors will buy in, given past experiences of valuation collapses.
3. Challenges in the Business Model
Moonlit Side faces difficulties in both their consumer and business segments:
- Consumer side: For traditional internet businesses, more users reduce marginal costs, but for large models, additional users increase operational costs, with some heavy users even costing the company money, eliminating any scale benefits.
- Business side: The B side accounts for 70% of revenue, but K3’s pricing (100 yuan per million tokens) is among the highest in China. Customers need to prove that the high cost is justified (e.g., through higher task success rates or reduced manual intervention). Additionally, API customers are likely to switch to other models with just a few code changes, making it difficult to establish long-term relationships.
Moonlit Side aims to build an ecosystem by integrating K3 into corporate workflows through tools like Kimi Work/Code, but time is a critical factor at the moment.
4. The Road to Open Source: A Difficult Journey, But Chinese AI Wins Pricing Rights
Choosing to open-source K3 presents both challenges and opportunities:
- Challenges: Since the model is available for free, “parameter size” is no longer a competitive advantage. The company must rely on a developer ecosystem, corporate workflows, and a closed business loop to retain users, which is a more challenging path than maintaining a closed-source model.
- Breakthrough: K3 has enabled Chinese AI to gain recognition globally based on its technical strength, narrowing the gap with top closed-source models like GPT and Claude. This marks an important step for China’s AI industry in transitioning from a follower to a participant in the global market.
Conclusion
K3 has positioned Moonlit Side at the forefront of global AI, but whether they can maintain this position depends on several factors: whether they can overcome computing power limitations, whether B-side revenue can cover costs, and whether the capital market will support their IPO. Regardless of the outcome of the IPO, Moonlit Side has accomplished something significant by demonstrating that Chinese AI can compete on the basis of technical excellence. This is indeed worthy of respect.
(Note: This article does not constitute investment advice, and any risks are borne by the reader.)