Just imagine it—before, enterprises would be hesitant because of privacy concerns or cost. But now, they can run their AI locally and get all the benefits without the huge expense. That opens up a lot of possibilities.
The world of AI is on the brink of a massive transformation, and it’s not just because of breakthroughs in technology. It’s because of cost—or rather, how DeepSeek has dramatically reduced it. With the potential to cut the price of AI development by orders of magnitude, DeepSeek could open the floodgates for a new generation of AI-powered startups.
Jordi Torras, AI & Machine Learning Strategist at Torras.ai, believes that DeepSeek’s cost innovation will change the game in 2025. While the technology itself is not new, the affordability it offers will offer unparalleled opportunity for new companies to develop their own AI.
Game changer: Until now, the barrier to entry for AI was sky-high—around $100 million—effectively locking all but the biggest corporations out of the space. Even Microsoft chose to license technology from OpenAI rather than build its own.
"Now, with just $5 million, you can build something that competes with GPT-4, which took around $100 million to train," Torras said. "That changes everything. Suddenly, regular startups can create their own models and improve their own systems."
Torras has a unique perspective on AI development as a consultant for AI strategy and developer of AI programs for business schools. He previously spent over 15 years leading Inbenta, a pioneer in AI and natural language processing, before selling the company in 2022.
Transformer architecture: "What these guys have built is not really new in terms of architecture, capabilities, or intelligence," Torras explained. "It’s the same thing as GPT-4, Gemini, or Claude—just at a fraction of the cost. And that changes everything because today cost is a big part of AI development."
The foundation of modern generative AI models is the transformer architecture, a neural network design optimized for natural language processing. It’s publicly accessible to all major players like Google, OpenAI, and Anthropic.
Instead of hundreds of millions of dollars, we’re talking about millions of dollars. That’s two orders of magnitude.
Orders of magnitude: The assumption has always been that building a competitive AI model required hundreds of millions of dollars in hardware—mostly GPUs, enriching Nvidia’s bottom line.
"Now what DeepSeek has proved is that just by using ingenuity, math, and a good understanding of how transformers work, they’ve been able to build a very similar architecture with similar quality—at a fraction of the cost," Torras said. "Instead of hundreds of millions of dollars, we’re talking about millions of dollars. That’s two orders of magnitude."
Cost cutting: DeepSeek’s potential goes beyond startups. Enterprises concerned about data privacy and security may soon be able to run AI locally, within their own infrastructure, without the hefty price tag or the need to send sensitive information to external providers like OpenAI.
"Just imagine it—before, enterprises would be hesitant because of privacy concerns or cost. But now, they can run their AI locally and get all the benefits without the huge expense. That opens up a lot of possibilities," Torras explained.
Business class for cents: Torras likened DeepSeek’s cost disruption to a hypothetical airline that offers a business-class flight from Barcelona to Atlanta for just 50 cents.
"That would change the airline industry, right?" he said. "It’s the same with DeepSeek. It’s not about the quality being something crazy new; it’s about the cost. And that cost is what will disrupt the AI industry."
