Summary of Key Points
Chinese AI model companies are transitioning from a phase of following others in competition to a phase of leading the way in cutting-edge innovation. However, this competition is no longer just about the capability of a single model, but about the ability to continuously develop more advanced intelligence. Currently, domestic model companies are engaged in a costly race involving parameters, computing power, and capital: the number of parameters has increased from trillions to 3 trillion, with ByteDance even aiming for 5-10 trillion; computing power requires the construction of data centers with GW (gigawatt) capacity; and the demand for funding has exploded (Moonlit Side raised $7.5 billion this year, while Alibaba raised HK$80 billion through an equity offering). Yet, revenue lags far behind these investments—top Chinese and American model companies have a 50-fold difference in annual revenue, and the domestic market is limited, making it difficult for open-source models to generate profits. Faced with resource constraints, different companies have chosen different strategies to break through: some are betting on large-scale model pre-training, some are optimizing existing models to reduce costs, and others are developing their own hardware. The key to this competition is whether they can continue to create advanced intelligence despite limited funds and low revenue.
I. Parameters, Computing Power, and Capital: An Unstoppable Race of Spending
The approach to competition in the Chinese AI industry has changed. It used to be about the size of model parameters, which increased from trillions last year to 3 trillion with Kimi K3, with ByteDance aiming for 5-10 trillion. But increasing parameters requires substantial computing power—for example, building data centers with GW capacity (1 GW equals 1 million kWh of electricity, which is equivalent to the power consumption of a small county). And computing power itself requires significant investment:
- Moonlit Side raised $7.5 billion this year, and DeepSeek’s first round of financing of RMB 50 billion was not enough; they need to raise another RMB 50 billion.
- Tencent and Alibaba invested RMB 120 billion in AI in one quarter, with Alibaba even raising HK$80 billion through an equity offering specifically for AI.
In essence, the current AI competition is about who can spend the most to gain an advantage in terms of parameters and computing power.
II. Revenue as a Hindrance: The Realities of Commercialization
Spending money is easy, but making a profit is difficult. This is the biggest challenge for Chinese model companies:
- There is a huge revenue gap between China and the US: OpenAI and Anthropic earn over $100 billion annually, while domestic companies like Zhipu, DeepSeek, and Kimi together earn only around $2 billion.
- The domestic market is limited: Alibaba Cloud’s AI-related business generates only $7.3 billion annually, which reflects the upper limit of what domestic companies are willing to invest in AI.
- Open-source models are difficult to monetize: Many open-source models are widely used, but they don’t generate profits. Kimi is trying to solve this by charging licenses—companies that use Kimi-K3 for commercial products have to pay, marking a shift from free use to profit generation.
Moreover, the valuation-to-revenue ratio (valuation/ARR) of Chinese model companies is much higher than that of American companies. If revenue cannot keep up with spending, valuations will plummet, making it even harder to raise funds.
III. Three Different Paths: Strategies for Breakthrough
Without the financial support of giants, domestic model companies must find their own ways forward. The three leading companies have chosen different approaches:
1. Kimi (Moonlit Side): Rebetting on Pre-training
Previously, Kimi was criticized by Anthropic for its “distillation” method of quickly replicating others’ models. Now, it has changed its strategy to increase the scale of pre-training, which is like building a larger building from scratch instead of copying others. Kimi-K3 has performed well in multiple tests, closing the gap with OpenAI in the short term.
2. Zhipu: Optimizing Existing Models to Reduce Costs
Zhipu is not blindly increasing parameters but is focusing on “post-training” existing models, such as enhancing long-text processing and using reinforcement learning for optimization. This approach improves performance while reducing inference costs, making it suitable for private deployments by government and enterprise clients who need to set up their own servers with limited budgets.
3. DeepSeek: Developing Hardware Themselves
DeepSeek is not only increasing parameters but also developing its own chips and optimizing its computing infrastructure. As the bottleneck of computing power becomes more apparent (due to a shortage of GPUs), developing hardware can give them a competitive advantage.
IV. The Essence of Cutting-edge Competition: The Ability to Continuously Develop Intelligence
As the article states, the dividing line in cutting-edge competition is not about how powerful the current models are, but about the ability to continuously create more advanced intelligence. For Chinese model companies, this means:
- They cannot rely on a one-time investment to increase parameters; they need to find ways to continuously improve performance with less money.
- The overseas market is essential: The domestic market is too small, so they must seek profits globally. For example, CoreWeave (an AI cloud invested in by NVIDIA) will deploy Chinese open-source models immediately, and Harvey (a legal AI company) is using Kimi-K3 in its products—these are opportunities.
- They must break away from the mindset of following others: While Chinese companies currently focus on OpenAI, can they develop their own unique paths in the future? DeepSeek’s hardware-plus-model approach and Zhipu’s cost-optimization strategies are examples of such attempts.
This competition is just beginning. Those that survive will be the ones that find a balance between spending and generating profits, and continue to create truly valuable intelligence.
(The entire text is explained in plain language, making it easy for non-financial professionals to understand the core logic.)