Open Source vs. Closed Source: Who Will Win the US-China AI Battle?

Recently, Chinese open-source large language models have exploded onto the scene collectively. From DeepSeek and Kimi K3 to Zhipu GLM and MiniMax, a wave of domestic Chinese models has swept the globe with an "open source + free" strategy, directly challenging the commercial foundations of American closed-source companies like OpenAI and Anthropic. On the other side, NVIDIA CEO Jensen Huang has publicly championed open source, Elon Musk has retweeted in support, and Microsoft's CEO has also cast his vote in favor. This debate over AI technology roadmaps is becoming a new focal point in US-China tech competition.

This article is based on discussions from the Taiwanese political commentary show Headlines Talk, organizing the commercial logic and strategic dynamics behind the open-source and closed-source approaches.

Open Source vs. Closed Source: A Battle of Interests

Before diving deeper, it's worth clarifying the basic concepts of these two approaches. Open-source models refer to AI models whose weights, architecture, and even training code are publicly released, allowing anyone to download, modify, and deploy them. Closed-source models only provide services through API interfaces — users cannot access the underlying weights or code. The fundamental difference lies in who controls the technology: open source hands control to the community and users, while closed source retains control within the developing company. This distinction directly determines the business model — closed-source vendors charge per API call based on usage, while open-source vendors profit indirectly through ecosystem services, cloud computing resource sales, and other means.

In this debate, one noteworthy phenomenon is that even within the American tech community, clear divisions have emerged. OpenAI and Anthropic are reportedly in closed-door discussions with US regulators, hoping to tighten oversight of Chinese open-source models, arguing that "once model capabilities approach or exceed human levels, fully open-sourcing them increases the risk of misuse."

However, NVIDIA's Jensen Huang offered a starkly different perspective. He emphasized on social media: "AI will reshape every industry, power every company, and be built by every nation. Open-source models enhance safety and cybersecurity, accelerate innovation diffusion, and enable sovereignty." His conclusion: "The world needs both frontier closed models and frontier open-source models."

Show guests pointed out incisively that these two camps represent different interest positions. For OpenAI and Anthropic, the biggest pressure is monetization and profitability — closed source means "you pay to use my tools." For NVIDIA, whether open or closed source, everyone still needs to buy GPUs and compute power, so Huang naturally "profits from both sides" and welcomes everyone in the world using AI.

The Market Shock Triggered by Kimi K3

What truly made the American AI community collectively nervous was the launch of Kimi K3 by Moonshot AI. According to the show's analysis, Kimi K3 has 2.8 trillion parameters, making it the world's largest open-weight model, freely downloadable by developers.

More critically, Kimi K3's impact area is no longer chatbots but coding capability — precisely Anthropic's core advantage. Guests noted that Kimi K3's coding ability can already match Claude's top-tier models, directly undermining the established notion that "you must use American closed-source models for programming."

After its debut, Kimi K3's benchmark performance boosted Moonshot AI's daily sales by at least 6x, while competitors' AI stocks triggered sell-offs. The show mentioned that the Philadelphia Semiconductor Index has been on a rollercoaster ride recently, with Kimi K3 being one of the key drivers behind it.

Regarding external claims that Kimi K3 "distilled" other models, guests explicitly rebutted this. Here's an explanation of knowledge distillation: Knowledge Distillation is a model compression technique proposed by Geoffrey Hinton and others in 2015. Its basic principle is to use an already-trained large model (called the "teacher model") to guide the training of a smaller model (called the "student model"). The teacher model's output probability distribution contains rich "dark knowledge" — it not only tells the student what the correct answer is but also conveys the relative relationships between incorrect answers. By mimicking the teacher's output distribution rather than merely learning labels, the student model can achieve performance close to the teacher's with a much smaller footprint. Since distillation's logic is "big teaches small," Kimi K3 itself has 2.8 trillion parameters — virtually the world's largest model — meaning there's no larger model that could serve as its "teacher." This is the core argument guests used to refute the distillation claims.

This technical breakthrough stems from its independently developed Latent Mixture of Experts (Latent MOE) algorithm. Mixture of Experts (MOE) is a sparsely-activated neural network architecture whose core idea is to decompose a model into multiple specialized "expert" sub-networks, activating only a small subset of experts for each inference to handle specific inputs. The advantage: total model parameters can be enormous (like Kimi K3's 2.8 trillion), but the parameters actually involved in each inference are far fewer than the total, maintaining powerful capabilities while controlling computational costs. Kimi K3's "Latent MOE" is a further evolution of MOE, introducing latent space representations in the expert routing mechanism, making expert selection and combination more flexible and efficient — considered the key technical innovation enabling the model to achieve top-tier performance at relatively low inference costs.

The Strategic Value of Open Source: Sovereign AI and Ecosystem Lock-in

China's adoption of an open-source free strategy is far from pure altruism. Guests analyzed that this approach dramatically lowers AI's barrier to entry, allowing any country or company — even those without the ability to develop their own large models or with limited budgets — to modify and deploy based on open-source models.

This aligns perfectly with Jensen Huang's repeatedly emphasized concept of "Sovereign AI": every country can have its own large model. Sovereign AI is a concept that has rapidly gained traction in international tech policy circles over the past two years. Its core proposition is that every nation should possess the ability to build AI infrastructure using its own data, language, and culture, rather than relying entirely on AI services provided by foreign companies. The backdrop is that AI large models are becoming infrastructure comparable to electricity and the internet — if a country's AI capabilities depend entirely on overseas closed-source models, it faces not only data sovereignty risks (sensitive data must be uploaded to foreign servers) but also potential "supply cuts" during geopolitical conflicts. Open-source models provide exactly a low-cost path to sovereign AI — countries can download open-source models, deploy and fine-tune them on local infrastructure, achieving autonomous control over AI capabilities.

Data shows that the country with the most DeepSeek downloads is India — the reason is simple: open source means no need to train a large model from scratch. This also explains why the Middle East, Southeast Asia, and other regions have shown strong interest in Chinese open-source models.

From a corporate cost perspective, the differences between the two models are stark:

• Closed-source models: Charged per input/output token, costs determined by the vendor
• Open-source models: Users can freely choose their compute source (e.g., Alibaba Cloud, Baidu Cloud), with far greater cost controllability

Here's an explanation of token-based billing. A token is the basic unit by which large language models process text — one token corresponds roughly to one English word or one to two Chinese characters. Closed-source model APIs typically charge separately for input and output tokens. For example, GPT-4o is priced at $2.5 per million input tokens and $10 per million output tokens. This means companies incur costs with every model call, and as usage grows, costs increase linearly or even super-linearly. In contrast, open-source models allow companies to deploy on their own servers or third-party cloud platforms, converting per-call fees into fixed compute costs — for high-frequency use cases, the cost advantage is extremely significant.

This also explains why the most-used large models globally are almost all products from mainland China.

The deeper strategic significance lies in ecosystem lock-in: once enterprises adopt open-source models like DeepSeek or Kimi K3, they are to some extent drawn into China's compute ecosystem. This has also driven domestic chip and HBM (High Bandwidth Memory) prices to skyrocket.

The Chip Ecosystem: CUDA's Moat Is Being Filled

The rise of open-source models has triggered chain reactions. According to the show's analysis, starting from DeepSeek V4, eight Chinese domestic chip companies can all connect to these models, no longer dependent on NVIDIA's CUDA ecosystem.

CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model launched by NVIDIA in 2006, allowing developers to leverage GPUs' massive parallel processing capabilities to accelerate general-purpose computing tasks. Over the past decade-plus, virtually all mainstream deep learning frameworks (PyTorch, TensorFlow, etc.) have been deeply bound to CUDA. Millions of AI developers worldwide have built their code on CUDA, creating extremely high migration costs and ecosystem lock-in — this is the so-called "CUDA moat."

Guests pointed out that the CUDA moat is essentially a computing paradigm based on graphics cards repurposed for AI computation. But now, an increasing number of AI chip companies are building AI-specific ASICs directly, bypassing the graphics processing layer entirely. Meanwhile, the maturation of open-source compilation frameworks (such as Triton and MLIR) is also reducing CUDA dependency, with more and more AI workloads able to bypass CUDA and run directly on non-NVIDIA hardware. CUDA's barrier effect has therefore been significantly weakened. The recent surge in China's chip industry is fundamentally driven by massive expansion at the application layer.

On the other side, memory chips — especially HBM (High Bandwidth Memory) — have seen rapidly rising prices. HBM is an advanced packaged memory that vertically stacks multiple DRAM chip layers using Through-Silicon Via (TSV) technology. Compared to traditional GDDR memory, HBM offers several times to over ten times higher bandwidth (the latest HBM3E exceeds 1.2TB/s), with lower power consumption and smaller footprint. For large language model inference, bandwidth is often more critical than raw compute — model parameters need to be quickly read from memory to computing units, and HBM's ultra-high bandwidth directly determines inference speed and efficiency. Currently, the global HBM market is highly concentrated among SK Hynix, Samsung, and Micron, with SK Hynix holding approximately 50% market share.

Guests used the analogy of "compute centers being equivalent to power plants": when the entire world is building compute centers, memory becomes a scarce resource in short supply. As the global AI compute center construction boom progresses, HBM has become one of the most critically scarce semiconductor products, with prices surging continuously through 2024-2025. This also explains why when Apple reportedly considered using Chinese memory chips, Micron would take strong exception — because it could mean losing a crucial major order.

Who Will Win: Open Source or Closed Source?

On the question of "which will ultimately prevail — open source or closed source," show guests generally leaned toward open source. Their reasoning: the closed-source model burns through massive capital while the US faces enormous fiscal pressure. Meanwhile, if the US and China each block the other's AI, it would be "cutting off their own paths" — harmful to both innovation and markets.

Guests also mentioned that OpenAI's recent internal system attack issues have exposed AI safety concerns, while such phenomena have yet to appear in Chinese open-source models. Of course, this represents a single-source perspective that requires more information to verify.

What you might not have noticed is that the essence of this battle is a deep-level contest between business models and technology ecosystems. Open source spreads rapidly through low barriers and low costs, capturing global markets and developer mindshare. Closed source maintains high premiums through technological leadership. In the short term, both will likely coexist for an extended period, as Huang said — "both are needed." But from a trend perspective, open source is rewriting the global AI landscape at a faster pace.

For professionals following the AI industry, this is not merely a choice between technology roadmaps — it's a critical entry point for understanding the future trajectory of global AI competition.

Key Takeaways