DeepSeek Proves AI Models Are Worthless – Data and Adoption Are Gold

This week’s spotlight is undoubtedly on DeepSeek, and for good reason—it touches on multiple angles.

Some view it through a geopolitical lens, seeing China’s ability to develop something on par with the U.S. despite strict chip restrictions as a significant milestone.

Others downplay its impact, arguing that DeepSeek is understating the number of GPUs it uses and may have trained its model illegally with ChatGPT.

Some see it as a challenge to the U.S.'s high-cost AI investment model, which has relied on massive infrastructure spending.

Others contend that, regardless of model efficiency, hardware remains essential for compute power. Ironically, greater efficiency could accelerate adoption, further driving demand.

Whatever the case, DeepSeek strikes multiple chords at once. By opening up new possibilities and increasing uncertainty, it has captured widespread attention, leaving many wondering what’s next. The reality is, no one knows for sure.

I want to take the discussion further—beyond the immediate debates—to explore the second- and third-order effects that may not yet be widely considered.

AI Models Will Become Commodities

AI models are increasingly heading toward commoditization. In the near future, building a model will no longer be a challenge—anyone will be able to adopt and customize existing models for their own use.

We've already seen a proliferation of models following ChatGPT’s success—Perplexity, Anthropic, Claude, Ernie, Qianwen, and many more. DeepSeek is just another addition to this growing list. The reality is, creating a model isn't the hardest part. The real challenge lies in breaking through the noise and achieving widespread adoption and monetizing it.

DeepSeek has demonstrated that building models is becoming even easier and more cost-effective, thanks to a technique known as distillation. Simply put, instead of training a model from scratch using vast amounts of data—an expensive and time-consuming process—one can use an already trained AI to "teach" a new model.

This is where OpenAI and Microsoft have raised concerns, alleging that DeepSeek might have used ChatGPT to train its model. OpenAI has spent billions scraping the internet to train its model, despite facing ongoing copyright disputes. For DeepSeek, replicating this approach would be both inefficient and constrained by limited GPU availability and financial resources. Instead, it leveraged distillation to train its model.

If distillation becomes widespread, building AI models will become significantly easier. Developers can now train new models using multiple existing AIs, each excelling in different areas. This is precisely what DeepSeek did—it distilled knowledge from Meta’s open-source Llama model and Alibaba’s Qianwen. Furthermore, it incorporated a Mixture of Experts (MoE) approach, activating specific AI components only when required, thereby reducing compute costs.

Think about it this way: I could build my own model by combining open-source models and refining them through distillation, using various trained AIs. If I wanted to, I could even distill DeepSeek into my own model. This effectively makes AI models a commodity—similar to how anyone can find free recipes online and mix elements to create their own unique dish.

With AI models becoming commoditized, many closed, paid models will struggle to compete. However, ChatGPT holds a distinct advantage—it’s not just an AI model but a fully developed consumer product.

Consider Microsoft Office as an analogy. Free alternatives like Google Docs, OpenOffice, and LibreOffice exist, yet Microsoft Office remains dominant. Why? Because of its network effect—users stay within its ecosystem due to familiarity, compatibility, and widespread adoption.

ChatGPT has achieved a similar effect, becoming the most widely used LLM in the world with an estimated 300 million weekly active users. In contrast, most other models have yet to reach this scale or evolve into full-fledged consumer products. As a result, many of them will be commoditized and primarily serve as one of the many models available rather than standalone AI products.

Data is the Gold

AI models rely on data to learn, and data is what will ultimately separate the winners from the losers. More importantly, data is where the real revenue lies. We are already seeing this play out, with OpenAI striking major licensing deals—and more are expected to follow: