The Big Model “Money-Burning” War Escalates: Zhipu Raises Another $5 Billion; Computing Power Becomes the New Battlefield
I. Summary of Key Points
In simple terms, the rules of competition in the big model industry have changed. Previously, the focus was on “who has the smarter model”; now, it’s about “who has more money and can sustain the long-term consumption of computing power.”
Zhipu AI, which recently went public, has quickly launched another massive round of financing, amounting to $5 billion (approximately RMB 35 billion). This money is not used for salaries or marketing but for purchasing chips, building data centers, and developing infrastructure.
At the same time, leading domestic players like DeepSeek and Kimi, as well as international giants like OpenAI and Anthropic, are all investing heavily in computing power infrastructure. This indicates that the big model industry has shifted from a “light-asset” software business to a “heavy-asset” infrastructure business. Only those who can secure cheap, stable, and large-scale computing power will survive in this marathon.
---
II. In-Depth Analysis: Why Have Big Model Companies Suddenly Become So “Heavy?”
To help you better understand the logic behind these moves, we break down the news into the following five aspects:
1. Accelerated Financing: Going Public Is Not the End, but the Beginning of “Blood Transfusion”
Many might think that once a company goes public, it has the funds and can start making money slowly. However, in the big model industry, going public is merely obtaining an “entry ticket,” not a “guaranteed victory.”
- Time Frame: Zhipu went public on the Hong Kong Stock Exchange in January this year, raised $4 billion in July, and now has launched another $5 billion in September. In just eight months, it has raised funds three times from the capital market at an extremely fast pace.
- Where Does the Money Go? The news clearly states that the money is primarily used for “research and development, computing resources, and infrastructure.” This suggests that big model companies are currently experiencing negative or very thin cash flows and need to continuously obtain external funding to sustain their substantial R&D and computing power expenses.
- Implication: If a big model company stops financing, its R&D and computing power services could come to a halt. Therefore, competing for funds and capital is as important as competing for talent.
2. Major Shift in Role: From “Computing Power Buyers” to “Computing Power Owners”
Over the past three years, the business model for big model companies was straightforward: **raise funds → rent GPUs/purchase chips → train models → release products.* Computing power was treated like a cost, purchased as needed.
But now, leading players (such as Zhipu and DeepSeek) are building their own data centers, purchasing domestic chips, and even customizing them.
- Why the Change?
- Previously: Computing power was a consumable resource that was used up when needed.
- Now: Computing power has become a strategic resource. Models are updated rapidly (pre-training, post-training, reinforcement learning, etc.), and Agent applications continuously consume tokens (computing units).
- Example: Previously, you might rent a car when you needed more; now, if renting cars is too expensive and unreliable, you would buy your own, build your own charging stations, and maintain your own infrastructure.
- Advantage: By controlling computing power, companies can more stably manage the training process, potentially significantly reducing training and inference costs in the long run. Those who control chips, electricity, and scheduling software hold the key to the models’ success.
3. Rising Barriers to Entry: From “Intelligence Competition” to “Financial Strength Competition”
During the “Hundred Models Battle” in 2023, the focus was on benchmark test scores and model size. A strong model was enough to secure the next round of funding.
By 2026 (the current time frame mentioned in the news), the competition logic has changed:
- Model capability is no longer the sole barrier: The capabilities of leading companies (Zhipu, DeepSeek, Kimi, MiniMax, etc.) are not significantly different, and they are all in the top tier.
- New Barriers:
- Can you continuously iterate?
- Can you run large-scale clusters stably?
- Can you integrate different brands of domestic chips to work efficiently?
- Can you reduce costs?
- Conclusion: Model capability is important, but computing power infrastructure and financial reserves are even more crucial. Without a solid foundation, even the best models won’t last long. This is the so-called “second barrier” that eliminates companies unable to support cutting-edge model operations.
4. Global Arms Race: Trillions of Dollars Flowing In, Turning AI into a “Infrastructure Project”
This is not a Chinese phenomenon but a global trend:
- Data Shock: The International Settlement Bank predicts that global tech giants will invest over $1 trillion in AI between 2025 and 2026. By 2030, AI investment could grow from the current $500 billion to $4 trillion.
- International Developments:
- Anthropic: Planning an IPO that could raise $100 billion and committing $100 billion to AWS (Amazon’s cloud services).
- OpenAI, Meta, Google: All investing heavily in data centers, custom chips, and energy.
- Significance: AI has evolved from a lightweight internet application to a major infrastructure project, similar to building high-speed railways or power grids. Funds are flowing into hardware at an unprecedented rate. This means that future competition will not only be about algorithms but also about comprehensive national capabilities in energy, chip manufacturing, and data center construction.
5. Zhipu’s “Domestic Strategy”: The Strategic Importance of Self-Sufficiency
The news mentions that Zhipu’s 1GW data center uses all domestically produced AI chips.
- Why the Emphasis on Domestic Chips?
- Supply Chain Security: In the context of geopolitics, there is a risk of supply disruptions for imported high-end chips.
- Cost and Customization: Collaborating with domestic chip companies (such as those mentioned in the news for custom chip development) allows for hardware optimization tailored to big model training, potentially offering better long-term cost-effectiveness and stability.
- Policy Alignment: It aligns with the country’s focus on technological autonomy and makes it easier to obtain policy support and resources.
- Challenges: Domestic chips may still lag in ecosystem maturity and single-chip performance, requiring software solutions (such as heterogeneous computing power scheduling) to compensate. Zhipu’s acquisition of software companies like Zhongke Jiahe aims to bridge this gap.
---
III. Implications for the General Public
1. The Big Model Industry Enters the Second Half of the “Elimination Round”: It used to be a time of diverse innovation; now, only the strongest companies survive. Small companies without substantial capital and computing power will struggle to stand out.
2. Computing Power Is Power: In the future, those with access to cheap and stable computing power will have a significant advantage in the AI era, similar to how those with steam engines dominated the Industrial Revolution.
3. Investment Perspective: If you are interested in tech investments, focus on companies that produce chips, build data centers, and manage power infrastructure. They are the “water suppliers” of this war and offer higher investment certainty.
4. Technological Progress May Slow but Become More Stable: Due to the need for infrastructure, the iteration speed of big models may slow from monthly updates to quarterly/annual updates, but the stability and cost-effectiveness of models will improve, which is good news for users.
In Summary: The big model war has shifted from “who has the smarter model” to “who can burn the most money and sustain the longest.” Zhipu’s $5 billion is not just for models but for the “entry ticket” and “ammunition” for the next round of competition.