Summary of Key Points
NVIDIA recently invested $19.9 billion (approximately 133.7 billion RMB) in acquiring and partnering with three open-source AI-related companies: $12.9 billion was spent on Hugging Face, the largest open-source AI model community in the world; $7 billion went towards acquiring the model licenses and team from Poolside, an AI programming company; and negotiations are also underway for an investment in Perplexity, an AI search engine. The core objective of these moves is for NVIDIA to shift from being a “chip seller” to a direct participant in the development of open-source large models. Its goal is to build its own open-source ecosystem, strengthen its demand for computing power, and challenge Chinese open-source initiatives such as DeepSeek and Kimi.
The Three “Generous” Investments: Targeting the “Golden Positions” in Open-Source AI
All three transactions share two common characteristics: high cost and open-source nature:
- Hugging Face: The largest open-source AI model community, acting as the “app store” for AI developers, where new models from DeepSeek and Kimi are released promptly. With annual revenue of only $150 million, NVIDIA’s investment of $12.9 billion represents 86 times their annual income, highlighting the company’s emphasis on this critical model hub.
- Poolside: An AI programming company founded by a former GitHub CTO that develops open-source models capable of understanding code. NVIDIA’s investment of $7 billion (more than three times its previous valuation of $2 billion) not only includes model licenses but also incorporates over 100 of its employees, effectively acquiring both the technology and the team.
- Perplexity: An AI-powered search engine (similar to Google’s search service) that has joined NVIDIA’s open-source alliance. The post-investment valuation may exceed $30 billion, a 50% increase from last year.
These three companies hold key positions in the open-source AI ecosystem: Hugging Face serves as a traffic gateway, Poolside is a technical cornerstone in a specific domain (programming), and Perplexity represents a practical application. By controlling these “nodes,” NVIDIA essentially gains control over a significant portion of the open-source landscape.
From “Chip Seller” to “Direct Participant”: NVIDIA’s Anxiety and Ambition
NVIDIA was once the dominant “chip seller” in the AI era, with its GPUs being essential for training and running large models by companies like OpenAI and Anthropic. However, the situation has changed:
1. Changing Customer Behaviors: Major open-source model companies (such as OpenAI and Google) are beginning to develop their own chips, reducing NVIDIA’s market share. For example, Google’s TPU and OpenAI’s own chip development efforts could potentially steal NVIDIA’s business.
2. Restrictions on Chinese Open-Source Models: The U.S. government’s efforts to ban Chinese open-source models like DeepSeek and Kimi, which use NVIDIA chips for training, could further reduce NVIDIA’s demand.
3. The Open-Source Opportunity: Jensen Huang sees the potential of open-source models, which are free and flexible, attracting more developers and businesses, thus driving demand for computing power. Instead of waiting for others to develop open-source models, NVIDIA decided to get involved directly to secure this demand.
In short, NVIDIA is concerned about losing its market share and is taking action to create its own open-source models, enabling more users to leverage its GPUs for their own AI projects.
Replacing CUDA with an Open-Source Ecosystem
NVIDIA’s current strategy is reminiscent of its early move with CUDA:
- What is CUDA?: A free software tool launched in 2006 that transformed GPUs, originally used for gaming, into versatile tools for scientific computing. Scientists and developers had to use NVIDIA GPUs to run CUDA-enabled applications, which boosted GPU sales and ultimately contributed to NVIDIA’s success in the AI market.
- The Current Open-Source Model Ecosystem: NVIDIA develops open-source models (such as the Nemotron series) and promotes them through communities like Hugging Face. The more users these models attract, the more NVIDIA GPUs are needed for training and running them, effectively binding users to its products.
Jensen Huang’s strategy is to build a large open-source ecosystem that ensures a stable demand for NVIDIA chips, even if major customers develop their own chips.
Current Situation: Can NVIDIA’s Open-Source Models Compete with Chinese Players?
Despite its substantial investment, NVIDIA’s open-source models are not yet at the top level:
- Independent evaluations show that NVIDIA’s Nemotron 3 Ultra scores only 38 points, compared to 60 for China’s Kimi and 53 for DeepSeek; even Korean companies have stronger models.
- NVIDIA’s open-source model development started later, with its first fully autonomous model only being released at the end of 2025, while Chinese open-source initiatives have been in progress for years.
However, it’s important to remember that CUDA was also initially underestimated but later became a crucial asset for NVIDIA. With its financial resources, chips, and ecosystem, NVIDIA has the potential to overtake its competitors.
Conclusion
NVIDIA’s recent moves are aimed at consolidating its dominance in AI through a combination of financial investment and ecosystem development. By acquiring key open-source companies, it aims to create an ecosystem that makes its chips indispensable. Although its current model performance is not at the forefront, its ambition and resources give it a significant advantage. The future competition in the AI market will not only focus on chips but also on ecosystems, and NVIDIA has already started playing a strategic game.