Two weeks ago, there was a big buzz about Chinese open-weights AI models. For the first time, a Chinese open-weights model, DeepSeek-V3, matched the performance of frontier American models. This success quickly made waves in the Chinese open-source AI community. Following this, NVIDIA CEO Jensen Huang posted on social media for the first time, stating that open-weights models are the future and will play a critical role in maintaining America's AI leadership. Then, Anthropic issued warnings in Washington, cautioning that powerful, cheap Chinese models could pose severe safety risks, painting a dystopian future. The Trump administration is clearly divided and unable to decide whether to completely restrict, limit, or allow open-weights models. To top it all off, ASML's stock crashed. Reports surfaced that Netherlands-based ASML, a key player in chip manufacturing, had shipped a large number of lithography machines to China, showing once again that China's AI hardware foundation is finding ways to bypass export controls. But why are they giving these models away? Who benefits from this? Why is Nvidia a staunch supporter of open weights, while Anthropic and OpenAI are strongly against it? When did Chinese AI labs turn into charitable organizations? What are they really doing? Let's discuss this in today's episode of AI Business.

First, let's get the definitions straight. Many people use 'open source' and 'open weights' interchangeably. So, what is the difference between open source, open weights, and closed models? Imagine a bakery. The bakery makes and sells scones. You go to this bakery and buy a scone. When you take it home, you cannot replicate it. You cannot bake the exact same scone because you don't have the recipe, and you cannot change anything about the scone you bought. But you can eat it as is. You have the finished scone as a product, and you can consume it by adding toppings if you like. This is what a closed model is. Closed models are sold as a service (SaaS). The company keeps the model entirely behind an API. They control the version, the safety parameters, they can change the behavior, and they can change the price. A closed model follows the classic Software-as-a-Service (SaaS) model. As a user, you pay for access to this model. Frontier AI models are typically sold as closed models because they are extremely expensive to build. When the model developer sells it as a finished product, they recoup their costs this way. That's why they host the model behind an API and charge per token. No competitor can directly copy a closed model. Now imagine you go to the same bakery and they tell you: 'Hey, we are starting to sell raw bagels. You can bake them with salmon, with eggs, with cheese, or take them home and customize them however you want.' But again, they don't give you the recipe. They give you the customization capability. This is an open-weights model. In AI models, weights control the model's behavior and how it generates responses. This is the core and most critical part of an LLM. It defines the model's characteristics. Companies that release open-weights models—like Meta, DeepSeek, Alibaba, Moonshot, and Mistral—give you something you can run locally and modify. They release the trained weights. Once you have these weights, you can use them or integrate them into your own products without needing anyone's permission. Now for the last one. When you walk into the bakery, the baker stands at the door and hands you printed recipes for every bagel, scone, and pastry they sell, including all variations, customization tips, and all secrets. They give it to you completely for free. This is true open source, and it is very rare. There are very few examples of true open-source models in frontier AI. This is because they are completely different from traditional commercial models; they are mostly research artifacts or scientific disclosures. They have no commercial nature because the training data, training code, and how they were built and trained are documented step-by-step for the public, allowing anyone to replicate and retrain these models from scratch. It should also be noted that today, there are no true open-source models that match the quality of frontier models. Therefore, when you hear the buzz about Chinese LLMs or Chinese 'open-source' models, keep in mind that what they are releasing are actually open weights. They are open-weights models, not true open source. This distinction is very important.

But why does this matter so much? Because this distinction affects specific groups of users. This brings us to the next point: who does this actually affect? Who is open weights really for? First of all, this does not affect ordinary LLM or AI users. If you are not in the tech business or a developer, what Chinese labs release won't change your daily life. The audience that cares about this consists of businesses—especially those in healthcare, finance, or law, who worry day and night about keeping their data within their own infrastructure. With open-weights models, enterprises can run the entire model on-premise. Instead of sending sensitive data to OpenAI's or Anthropic's servers, they keep all sensitive information locally. The second group that cares about this is startups and small tech businesses. Many of them cannot afford the high API pricing of closed-model companies. Think of LLM wrappers, niche AI businesses, indie startups, indie music apps, or anyone building a real tech business with a real product. If you have a real product and switch to an open-weights model, you will see a massive difference in your profit margins. The third group is the infrastructure layer—cloud providers and chip manufacturers. Now, this might surprise you. Why them? Because any business using an open-weights model needs the infrastructure to run it. They need cloud servers, GPUs, memory, and bandwidth to run these model weights. Therefore, companies that sell infrastructure also promote open-weights models. NVIDIA CEO Jensen Huang aggressively promotes open-weights models because NVIDIA doesn't care who has the best model. They don't care about the competition in the model market. Whether it is open source, open weights, or closed, all businesses using these models need chips to run them. Thus, NVIDIA, along with other chip makers and cloud providers, directly benefits from the growth of open-weights models.

Now, if you are listening to this, you might wonder: why do they release these weights? What is the point? If DeepSeek gives away its model weights for free, why would anyone pay DeepSeek? If anyone can download the weights, how do they make money? Here is how. Think of the weights as seeds. Think of the model weights as if you are thinking of seeds. When DeepSeek gives you access to the weights, they are printing instructions on how to plant the crop. Anyone in the world can take these instructions and plant this seed themselves. But owning a seed and owning a harvest are two completely different things. Owning a seed doesn't give you land. It doesn't give you water, fertilizer, sunlight, or the daily labor needed to grow this seed. The same is true for AI models. Running an AI model reliably, at scale, and cost-effectively requires GPU chips, electricity, data centers, systems engineers, and immense technical expertise to get the maximum throughput per dollar. Therefore, to do anything useful with these seeds, you have to build an entire farm. DeepSeek gives you the seeds for free, but owning the seed doesn't mean you can build the farm. That's why they release the weights. Weights are intellectual property, but DeepSeek doesn't make money by selling intellectual property. They make money by computing at scale. Large-scale compute is their product. Giving away intellectual property is not a constraint, because intellectual property has become a sunk cost once training is complete. What actually generates revenue after training is the industrial capability to serve this intellectual property to millions of API requests per second at a cost lower than anyone else. Any lab that sells API access to their model is running an industrial business. To put it simply, the output is generated text, and the unit cost (Unit Cost) is how much compute is required to generate that output, while the margin (Margin) is the difference between what they charge and the cost to produce it. Thus, the economics boils down to one question: who can generate a unit of output at the lowest cost? When we look at American frontier labs, think of them like a massive corporate headquarters. When a customer walks in with a request, no matter how simple or complex it is, they route it to every single employee in the building. Every agent, from the intern to the VP of customer success, has to look at the query, think about it, and only then does one of them respond. On one hand, you might think this is incredibly wasteful. But this is exactly how AI models like OpenAI's GPT-4 family or Anthropic's Claude operate. Every single query activates the entire model, every single weight, no matter how simple the query is. What they do in China instead is use an intelligent router. The router first looks at the incoming query and sends it directly to just two specialists in the call center who can answer it, while the other floors remain idle. If the other floors remain idle, their existence doesn't cost anything. The architecture used by Chinese LLMs is called Mixture of Experts (MoE). The math behind it is very simple. Instead of paying to activate the entire model for every request, you only pay for the small part of the model that is relevant to the question you asked. DeepSeek has a lower cost by design, which is why Chinese LLMs can afford to give away open-weights models. This makes strategic sense for their business. Training DeepSeek's V3 model ultimately cost around $5.6 million. GPT-4 cost well over $100 million. This cost difference is enabled by architectural decisions, because using a Mixture of Experts reduces the compute per query. That is exactly why DeepSeek can charge just a fraction of what GPT charges for equivalent performance. Look at the numbers. On DeepSeek's API, the operating cost for a certain volume might be $80. The equivalent operating cost on GPT-4 would be $450. And on Anthropic's Claude Opus, it would be $900. Again, this is possible because DeepSeek's actual cost to generate a million tokens is significantly lower than what OpenAI or Anthropic pay for their generation.

But look, if DeepSeek saved so much money using this architecture, why are Western labs so expensive? Because labs made different choices. The MoE architecture used in Chinese open-weights models is primarily a routing system, and this routing system exhibits non-determinism, meaning if you ask the exact same question twice, you might get slightly different answers. Anthropic's Claude operates on a different architecture, which is known to be very precise. For a company like Anthropic, whose entire brand is built on safety and alignment, they have to choose a much safer, more deterministic architecture, even at a higher cost. They consciously make this trade-off because Anthropic optimizes for consistency and safety. OpenAI, interestingly, sits somewhere close to DeepSeek's logic. They also use Mixture of Experts models, but unlike DeepSeek, they dedicate significant compute to safety-related tasks, trying to balance safety and cost efficiency at the same time. The key takeaway here is that architectural choices reflect what each lab prioritizes. DeepSeek optimizes for cost. It might not be the most deterministic or precise, but it is extremely cheap (and weights are free). Anthropic optimizes for alignment and consistency, which is why it requires so much capital. OpenAI tries to achieve both, but as you know from our videos, OpenAI is struggling financially under this pressure. But again, this doesn't answer my original question: why give the weights away for free? Of course, your product's margins might be smaller compared to Anthropic's, but it works for the majority of LLM wrappers. Why give the weights away for free? This, ladies and gentlemen, brings me to my final point, because there is no other way to survive. Remember the seeds and the farm analogy? DeepSeek's moat was never in the seed or the model itself. It was in the farm. In other words, it comes down to their ability to run the lowest-cost compute infrastructure. If a competitor cannot match your low operational costs, even if they have the entire recipe, even if someone steals it and leaks it, having the recipe won't protect their pricing power. This is why Jensen Huang supports open weights. Of course, NVIDIA makes money regardless of which model is used—whether it is open weights, open source, or closed. But his key point is that open weights is the future, and even the most closed frontier labs cannot stop their models from leaking into the open ecosystem. We have done entire episodes on corporate espionage, and OpenAI, Anthropic, Gemini, and all other frontier labs are highly vulnerable to corporate espionage. This is a very real threat. Even if no secrets leak directly, or if they guard them securely, competitors can simply query the closed models' APIs repeatedly and use the outputs as training data for their own models. This process is called Model Distillation, and it effectively clones the capabilities of the original model without ever seeing its weights. It turns out weights might not be the ultimate weapon. Ultimately, if everyone has access to the seeds, it is time to build a farm that cannot be beaten. So, let's return to where we started. As Jensen Huang said, open weights is the future. Closed AI advocates and Anthropic claim there are severe safety risks because it threatens OpenAI's and Anthropic's pricing power. Therefore, when they warn Washington about dystopian threats, it is not just about safety. While safety is a legitimate concern, make no mistake: behind the safety narrative, the big players are playing a game of massive business and geopolitics. We hope you found this useful, and we'll see you in the next episode. Goodbye.