Episode 3: DeepSeek’s Open-Source R1 Model Shook Nvidia & Democratized AI

You're listening to the Elefic Research podcast. Today, we're diving into a story that reads like Silicon Valley fiction, except it actually happened. How did a relatively unknown Chinese AI lab send shockwaves through the trillion-dollar tech industry, wiping out nearly $600 billion from Nvidia's market value in a single day?

This is the story of DeepSeek and their R1 reasoning model. A tale of constraints becoming catalysts, and how smart engineering trumped brute force computing. Picture this scene. It's January 27th, 2025. Trading floors from New York to Hong Kong are in chaos. Within hours, Reuters reported Nvidia alone shed roughly $593 billion in market value as its shares fell almost 17%, dragging the wider tech sector down. The Philadelphia Semiconductor Index plummeted over 9%. Microsoft, Oracle, and Alphabet all took significant hits. What could possibly cause such carnage?

A tiny AI research lab from Hangzhou, China had just released something that challenged everything the market believed about artificial intelligence. Let's rewind to understand who DeepSeek actually is. Founded in May 2023 by Liang Wen Feng, DeepSeek AI isn't your typical tech startup. Liang isn't a fresh-faced Stanford dropout. He's the founder of High-Flyer, one of China's premier quantitative hedge funds. And here's where it gets interesting. As Liang told Chinese tech publication 36KR, and as reported by Wired in January 2024, "I wouldn't be able to find a commercial reason for founding DeepSeek even if you asked me to. Basic science research has a very low return on investment ratio." Think about that for a moment.

In an era where every AI startup pitches itself as the next unicorn, here's a founder explicitly saying he's not in it for the money. DeepSeek was born from scientific curiosity, not commercial ambition. High-Flyer had been accumulating GPUs for years to analyze financial data. They had the computational firepower. What they didn't have was the typical Silicon Valley playbook. The company's approach to talent was equally unconventional. While Google and OpenAI poached seasoned researchers with million-dollar packages, DeepSeek recruited bright PhD graduates fresh out of Peking University and Tsinghua. As Liang explained in interviews cited by both Wired and Fortune India in early 2025, "Our core technical positions are mostly filled by people who graduated this year or in the past one or two years." The pitch to these young researchers, "Come solve the hardest questions in the world."

But here's where the story takes its first dramatic turn. In October 2022, the United States imposed export controls on advanced AI chips to China. Companies like DeepSeek couldn't buy Nvidia's latest H100 GPUs in the quantities needed for traditional AI development. As Liang acknowledged in a 2024 interview with 36KR, "The problem we are facing has never been funding, but the export control on advanced chips." Most would see this as a death sentence for AI ambitions. DeepSeek saw it as an opportunity. If you can't win the hardware race, change the game entirely. Marina Jong from the University of Technology Sydney, quoted by Wired, explained it perfectly. "DeepSeek has focused on maximizing software-driven resource optimization. They couldn't buy their way to success, so they had to think their way there."

This constraint-driven innovation led to several pivotal technical decisions. First, they embraced the Mixture of Experts architecture, or MOE. Imagine having a team of specialists where only the relevant expert works on each problem, rather than everyone tackling everything. That's MOE in simple terms. DeepSeek's R1 has 671 billion total parameters, but only 37 billion are active at any given time, as reported by UNU in January 2025.

But the real breakthrough came from their approach to training. While most AI labs rely on supervised learning, essentially showing the model millions of examples of correct answers, DeepSeek took a different path. They used reinforcement learning, letting the model discover effective reasoning strategies through trial and error. As Michigan State University PhD student Yihua Zhang told IBM in January 2025, they essentially asked, "Can we just reward the model for correctness and let it discover the best way to think on its own?"

The results were staggering. According to Jefferies analysts cited by Built In, R1 was trained for approximately $5.6 million. Compare that to OpenAI's GPT-4, which reportedly cost over $100 million to train. But the real shock came with usage costs. According to reports from DeepSov.ai in February 2025, R1's API costs were up to 96% less than some OpenAI models. That's not a typo, 96% cheaper.

January 20th, 2025. Launch day. DeepSeek didn't hold a glitzy press conference. They simply released R1 to the world. Model weights, code, technical papers, everything under an MIT license. The post on their API documentation site was almost understated. "Performance on par with OpenAI's 01, fully open-source model and technical report, MIT licensed, distill and commercialize freely."

The reaction was anything but understated. Within hours, Silicon Valley was buzzing. The DeepSeek chatbot app shot to the top of iOS app charts in both the US and China, according to Recode China AI. On social media, developers were sharing examples of R1's reasoning capabilities with a mix of excitement and disbelief. Marc Andreessen, the legendary venture capitalist, didn't mince words. He called it, "One of the most amazing and impressive breakthroughs I've ever seen," adding that this was "AI's Sputnik moment." The comparison to the Soviet satellite that shocked America in 1957 wasn't hyperbole. This was a wake-up call that the AI race had fundamentally changed.

But the real drama unfolded a week later in the financial markets. The core fear was simple but profound. If a relatively small Chinese lab could match or exceed the performance of models that cost hundreds of millions to develop, using a fraction of the computing power, what did that mean for companies banking on ever-increasing demand for AI hardware? The market's answer was brutal. Nvidia, which had ridden the AI wave to become one of the world's most valuable companies, saw nearly $600 billion evaporate in a single trading session. The broader semiconductor index fell over 9%. It wasn't just about one company, it was about an entire investment thesis being questioned.

Inside the major AI labs, the response was swift. According to reporting from The Information cited by Recode China AI, Meta was concerned that R1 outperformed their upcoming Llama 4 model. OpenAI, ByteDance, and other tech giants were scrambling to understand how DeepSeek had achieved these results. There were even whispers of potential research collaborations.

So how did they do it? The answer lies in turning constraints into catalysts. Unable to compete on raw compute power, DeepSeek had to be smarter about every decision. Their young team, unburdened by conventional wisdom, questioned fundamental assumptions about how AI models should be built. As Wendy Chang from the Mercator Institute for China Studies told Wired, "They employed a battery of engineering tricks that established players had overlooked or dismissed."

The culture Liang Wen Feng created was crucial. In a May 2023 interview translated by Recode China AI, he said, "Innovation requires as little intervention and management as possible, giving everyone the space to freely express themselves and the opportunity to make mistakes." For product leaders and innovation strategists, the lessons are profound. First, embrace constraints as innovation catalysts. DeepSeek couldn't throw money at the problem, so they had to think differently. Second, efficiency can be a powerful differentiator. In a world obsessed with bigger models and more parameters, DeepSeek showed that smarter often beats larger.

The open-source strategy was masterful. By giving away their technology, they built a global community overnight. Within days, developers worldwide were building on R1, validating its capabilities, and spreading the word. It was growth hacking at the frontier of AI. Perhaps most importantly, they proved that disruption doesn't require massive teams or unlimited resources. With focused talent, clear vision, and smart engineering, even established giants can be challenged.

Looking ahead, the implications are staggering. For the AI industry, R1 represents a democratization moment. Advanced reasoning AI is no longer the exclusive domain of companies with $100 million budgets. Small startups, academic researchers, even individual developers can now access and build upon frontier capabilities. For incumbents, it's a clarion call. The modes built on proprietary data and massive compute are shrinking. As reported by TechTarget and ORF Online, companies relying on high-cost, closed models will need to fundamentally rethink their value propositions.

There's a geopolitical dimension too. DeepSeek's success demonstrates that US export controls, rather than stifling Chinese AI development, may have accelerated innovation in unexpected directions. As noted by multiple analysts, this could reshape how policymakers think about technology competition. For DeepSeek itself, R1 is just the beginning. They've already released updates, with rumors of an even more advanced R2 model in development. They've gone from unknown research lab to a name that makes Silicon Valley executives lose sleep.

But perhaps the most important lesson is about the nature of innovation itself. DeepSeek didn't try to out-muscle the competition. They out-thought them. In an industry that often conflates progress with scale, they showed that intelligence, both artificial and human, can find more elegant paths forward. The DeepSeek moment isn't just about one model or one company. It's a reminder that in technology, as in nature, constraints drive evolution. The dinosaurs had size, but the mammals had adaptability. Sometimes being forced to be clever is the greatest advantage of all.

As we watch this story continue to unfold, one thing is clear. The AI landscape will never be quite the same. The age of assuming that only massive corporations with unlimited resources can push the frontier is over. The age of efficient, accessible, and open AI has begun. And it started with a hedge fund founder who insisted he wasn't in it for the money, a team of fresh graduates, and the audacious belief that smarter beats bigger every time.

Thanks for listening. This is Elefic Research, made by builders for builders. Talk soon.