Secures 33 Billion RMB in Financing, Yet Stock Price Experiences Initial Dip.

On September 13, Zhipu AI announced the successful completion of a substantial $5 billion financing round, equivalent to over 33 billion RMB. The following day, its stock price experienced a temporary intra-day plunge of over 10%.

Short-term capital exits, but long-term investors stand firm. The same announcement triggered two contrasting market reactions.

33 Billion RMB Financing Secured: AI Industry Embarks on Capital Arms Race

This financing round transcends Zhipu AI as a standalone event.

The source of the capital, its allocation, and market perceptions—answers to these three pivotal questions unveil the current dynamics within the large model industry.

Capital thresholds continue to escalate, marginalizing mid-sized and small players. In the first half of this year, the domestic large model sector (note: 'sector' here refers to the industry segment) secured over 100 billion RMB in financing, with nearly all funds flowing to top-tier firms. Smaller teams face mounting challenges in competing.

Training a single model entails billions in costs, demanding top-tier standards in computing power, data, and talent. Teams unable to secure funding must retreat to niche markets or risk marginalization.

In the foreseeable future, the AI industry will not be dominated solely by giants; mid-sized and small teams still have opportunities in vertical applications and ecosystem collaborations. However, the long-term trend of resource consolidation at the top seems inevitable.

Precision Capital Operations: Dilution Costs Below Industry Average. Zhipu's financing structure showcases strategic acumen—of the $5 billion, $2 billion comes from share placements and $3 billion from convertible bonds. Notably, the convertible bonds are interest-free, with a conversion price set 25% higher than the placement price.

Compared to pure equity dilution, this hybrid approach delays equity dilution—convertible bonds only dilute ownership post-conversion, serving as low-cost long-term debt until then.

Zhipu's management avoided the easiest path to secure funds, aiming to replenish resources while minimizing losses for existing shareholders.

Tech-Centric Focus: Rejecting Short-Term Arbitrage. Zhipu's announcement emphasized that all funds will support next-gen GLM R&D, self-training systems, and computing infrastructure—not for liquidity replenishment or profit enhancement.

AI remains one of the most capital-intensive sectors, with many firms prioritizing revenue scaling and rapid monetization. Zhipu, however, chose a more challenging path: prioritizing foundational technology over short-term returns.

Stock Volatility as Short-Term Noise: Long-Term Capital Remains Optimistic. On September 14, Zhipu's stock dipped due to dilution concerns, but long-term investors focused on a different narrative—Zhipu's MaaS revenue continues to rise, and domestic computing power adaptations accelerate.

This optimism is not isolated: MiniMax secured $2 billion in refinancing in July, DeepSeek advances toward listing, and Moonshot AI explores dual listings in Hong Kong and Shanghai.

Top-tier large model firms are racing against time—securing funds first grants a critical competitive edge.

Is 33 Billion RMB Worth It? Zhipu's Four Core Strengths

Despite securing 33 billion RMB, Zhipu's stock price fell.

The market does not doubt the financing itself but questions a fundamental issue: Can this capital buy competitive advantages?

To answer, we must examine Zhipu's four key assets.

Accelerated Commercialization: MaaS Revenue Dominates. First, subscription models prove viable. Zhipu's H1 revenue hit 954 million RMB, surpassing total 2023 revenue.

More critically, revenue sources shifted—MaaS and API businesses contributed 825 million RMB, with their share jumping from 26.3% to 86.5%. Even more telling, H1 MaaS gross margins turned positive at 24.6%, up from negative margins a year earlier.

If an AI firm's core revenue becomes as predictable as utility bills, valuation logic fundamentally changes.

Rapid Tech Iteration: GLM Large Model Advances. Second, tech updates. From February to August, GLM evolved from v5.0 to v5.3, releasing new versions roughly every two months—a brisk pace among domestic large models.

Crucially, iterations build incrementally: GLM 5.3 and 5.2 share the same technical foundation, with capability improvements driven by reinforcement learning during 'post-training.' Each model enhances the previous one rather than starting from scratch.

Open-Source Ecosystem: Building Developer Scale Barriers. Third, open-source strategy. Zhipu open-sourced over 60 cutting-edge models under MIT licenses, amassing over 100 million global downloads and serving over 5 million developers and enterprise users. Its MaaS platform exceeds 7 million registered users, including over 20,000 enterprise clients.

Open-source may seem altruistic, but it's strategic—if developers rely on GLM's workflows and toolchains, switching costs become prohibitive.

Domestic Computing Power Alignment: Fourth, adapting to national strategies. Zhipu aligned with 'self-reliant computing power' mandates by enabling GLM models to run on eight domestic chip platforms, including Huawei Ascend and Cambrian, achieving 100,000-chip-scale inference capabilities.

This isn't mere technical compliance but a prerequisite for government and enterprise markets.

Combining domestic large models with local computing power builds trust with government clients. Since 2024, Zhipu has secured multiple government-enterprise contracts from China Unicom, Postal Savings Bank, and Three Gorges Group.

Zhang Peng: The Ace Up Zhipu's Sleeve

Zhang Peng, holding an M.S. in Computer Science from Tsinghua University, founded Zhipu in 2019 and led its 2025 IPO. Every company milestone ties to him.

Visionary Strategy: Locking In B2B Early. Zhang's greatest asset is timing.

Zhipu skipped C-side traffic wars at inception, targeting enterprise services instead. At the time, B2B seemed unglamorous. Now, MaaS anchors Zhipu's stable revenue.

Zhang summarizes the strategy: evolving from selling products to capabilities to outcomes, diving deeper into the value chain at each step.

Extreme Long-Termism: R&D Investment Soars. Zhipu's H1 R&D spending hit 2.13 billion RMB, up 33.6% YoY. The latest $5 billion financing will fund model R&D and computing infrastructure—not for liquidity.

Unlike peers obsessed with parameter counts, Zhang prioritizes training efficiency and inference depth—emphasizing 'self-training' models over brute-force computing.

This approach seems unrewarding short-term but converts efficiency gains into cost advantages long-term.

All-Rounder Leadership: Balancing Three Fronts. Zhang juggles three battles simultaneously:

R&D: Focuses on post-training and self-evolution; Business: Shifts from project-based to subscription models, with API revenue leading; Capital: Listed in Hong Kong in 2024 as the 'world's first large model IPO,' now securing $5 billion more. All three fronts advance cohesively.

Win-Win Philosophy: Redefining Industry Authority. While overseas leaders close ecosystems, Zhang opens doors: 60+ models open-sourced under MIT licenses for anyone to use.

This isn't philanthropy but rule-setting—when global developers build on Zhipu, it becomes an unavoidable infrastructure layer.

Openness breeds ecosystems, which translate into authority. That's Zhang's differentiated path.

Loaded for Battle: Zhipu's Next Moves

Securing 33 billion RMB marks not an endpoint but the next competition phase. Zhipu's path forward is clear.

Pursuing Next-Gen AGI: Exploring 'Self-Evolution' Paths. Next-gen GLM aims to train models in self-designed environments with self-generated data, self-built infrastructure, and autonomous optimization. The logic is simple: let AI teach itself.

Zhipu's 'iteration' definition shifts from parameter counts to a model's ability to evolve in deeper training environments and complex tasks—a key competitive frontier.

Doubling Down on Computing Infrastructure: Targeting Domestic Adaptation. A major financing focus is computing infrastructure, especially domestic capabilities.

Recent progress shows GLM 5.3 tripling 'end-to-end' service performance on domestic chips while halving operator development cycles. Future Chinese chips won't just run GLM—they'll excel at it.

Full-Scenario Adoption: Revenue Poised to Explode. Coding marked Zhipu's first productivity breakthrough, with GLM now independently handling programming tasks—goal comprehension, task decomposition, tool invocation, debugging, and delivery.

Its greater value lies in capability transfer: a code-writing model can automate cybersecurity, data analysis, and more. Coding is merely the first application springboard.

First-Mover Advantages Remain, But the Game Isn't Over. Zhipu leads in MaaS revenue (80%+ of total), domestic computing adaptation, and open-source models (60+). These assets keep it in the game.

Yet rivals like Alibaba Tongyi, ByteDance Doubao, DeepSeek, and Moonshot AI aren't idle. Their tech directions remain fluid, and commercialization just began. Future generational shifts could rewrite rankings.

For Zhipu, the real question is whether it can build commercial and ecological barriers high enough before competitors catch up.

As for AI large models' trajectory, time will tell.