DeepSeek’s recent announcement sent NVIDIA’s stock tumbling, shedding $600 billion in market cap. Wall Street woke up to the fact that (1) AI is becoming cheaper to train, and (2) AI models can train themselves, further lowering the cost. Both insights are not new, but this cost erosion creates an economic reality in which AI becomes a commodity. The impact on business leaders and governments will be profound:
- Models are difficult to protect from other companies or countries.
- Foundation Models have a first-mover disadvantage.
- AI technology is rapidly becoming more accessible everywhere at lower costs.
The winners of all this innovation will be us—the users—but state actors and enterprises will need to rethink their approach.
What is DeepSeek?
Since it all started with DeepSeek, let’s discuss what it is. DeepSeek-R1-Zero was the first large-scale AI model trained purely with reinforcement learning (RL), eliminating the need for human feedback. This means an AI can now train itself and improve its reasoning autonomously. Sounds amazing, but it’s something we have been doing quite often already: Traditional GAN architecture, AlphaZero from Google, and many other structures use the concept of AI training itself. You can learn more about RL in my eCornell online certificate course “Designing and Building AI Solutions.”
The impressive part of the DeepSeek announcement was the cost of training. DeepSeek-V3’s last training run cost just $5.576 million. Compare this to the $100 million that GPT-4 supposedly cost. That is 18 times smaller.
DeepSeek Presents The Fallacies Of AI Moats
AI does not easily create moats. Early in the LLM hype, I wrote a review of the different moat fallacies of AI. In “Competitive Advantage of LLMs” I wrote:
“So it’s safe to say that a model moat does not exist for OpenAI. This is not surprising. Such a moat also did not exist in previous Machine Learning waves.”
No Model Moat: It’s Math – Math Is Not Protected
That’s what we see here: DeepSeek outpaced OpenAI. The reason is simple: transformer models are math. Math is not protected. It’s a neural network. Very simply put, neural networks are essentially stacked linear regressions within activation functions. Sounds funky? It isn’t. We all know already one linear regressions within an activation function. It’s called “logistic regression”. Neural Networks are no magic. But because they are no magic, one cannot protect them so easily.
No Data Moat: It’s Public Data – Public Data Is Not Protected
For a transformer model to work, it needs to be trained. As much as we train our human brain with experience, we train transformers with data. But general knowledge — like “the earth is round” — is widely available. My first startup faced this reality: we built analytics on public data and sold it to governments and NGOs. Initially, we commanded six-figure subscription fees. But once competitors caught up, using the same machine learning algorithms, , similar services became available for $100 — a 1,000x price drop. The same erosion is happening now in model training. True some data might still give the model an edge. How to protect data is an important aspect of AI business models. Roughly 1/5 of the eCornell online certificate course “Designing and Building AI Solutions” is focused on this topic.
A Real Moat: Engineering Excellence
Smart engineering work pushing the frontier of what is possible is a moat. DeepSeek’s team improved how to effectively train and manage resources such as data and infrastructure. Because of this engineering excellence, they were cheaper than OpenAI.
Without going too much into detail, here are the main novel approaches from DeepSeek — all aimed at training on smaller and more minimal hardware:
- Smart Resource Utilization: DeepSeek optimized inter-chip communication and made algorithmic and hardware tweaks to train on lower-level machines.
- Mixture of Experts: Instead of training the whole network in one go, DeepSeek selectively trained relevant areas of the model.
- Multi-Head Latent Attention: Compresses memory during inference, making it feasible to handle large datasets with minimal hardware.
Economic Implications of DeepSeek And Beyond
AI models are becoming commoditized. Open-source challengers like DeepSeek demonstrate that high-quality AI can be built inexpensively. Effectively there is no longer time barrier to entry. Thus investment returns can not be amortized over long period of time like.
First-Mover Disadvantage
Moreover, we will see that research into frontier models presents an economic disadvantage because AI models can train other AI models. One of the early examples of this was Stanford’s Alpaca model. It was fine-tuned for less than $100, utilizing existing GPT-3 that had cost over $4.6 million. Thus, Alpaca was 46,000 times cheaper.
Investment Returns Under Pressure
This will create the question for investors of how to recover an investment. Let’s think about Donald Trump’s proud Stargate $500 billion announcement. If we use the same 18x discount that we just saw after six months, then someone will be able to do the same $500 billion value creation with just a $27 billion investment. If we apply Alpaca’s approach, the $500 billion investment is now only worth $11 million.
AI: The New Lighthouse
AI infrastructure might become a common good like a “lighthouse”. In the end, AI research is research that will benefit everyone. The lighthouse costs money, but it benefits everyone.
DeepSeek Was Possible Because Of The US Chip Ban
The U.S. imposed chip export bans to limit China’s AI capabilities, operating under the assumption that without advanced hardware, AI development would stall. This logic mirrors nuclear proliferation controls—restrict the technology, and you restrict capability. However, AI is software, and it is created differently compared to the creation of a nuclear facility, making traditional protectionist measures ineffective. Or — in DeepSeek’s scenario — even counterproductive. DeepSeek, banned from using top hardware, ended up building really resourceful approach and became better then top US AI companies.
Democratization of AI
DeepSeek’s emergence confirms a trend we have seen for a while. AI models are becoming a commodity, and AI models will be open source. If the cost for AI continues to drop, more processes, tools, and services will use AI. This democratization of AI will be excellent for users and businesses alike. The real challenge now is not technological but economic and regulatory: how do we govern an industry where barriers to entry are disappearing?
