| 
Ever since Chinese AI startup DeepSeek unveiled its new R1 model – which rivals the best U.S. bots like ChatGPT – AI stocks have been on a wild roller coaster ride. The news was explosive, sparking fears that companies will pull back on their extreme AI spending. That led chipmaker Nvidia (NVDA) to lose $600 billion in market value within just a few hours on Monday – the biggest single-day drop in market history. Unsurprisingly, there’s been a lot of talk about DeepSeek, Nvidia, and the rest of the AI universe over the past few days. But the main question most are left wondering is: How? How did a virtually unknown Chinese startup disrupt the entire global AI industry, reportedly developing the world’s most advanced model for just around $5 million? Yet, we think the more important question may be: What does it mean? What does DeepSeek’s breakthrough mean for the industry as a whole? Will it mark the end of the big bull run in AI stocks? Or is the recent selloff just an overreaction, making this a very compelling buying opportunity? In this issue, we will attempt to answer these questions, to the best of our ability. And in so doing, we hope to point you in the direction of some future winning stock picks. Doing More With Less So… how did DeepSeek do it? Based on our research, the startup’s success came via innovative engineering born out of geopolitical necessity. That is, over the past two years, the U.S.’ approach to building next-gen AI models has been to “throw money at it.” Build more data centers. Buy more GPUs. Hire more engineers to build, train, and advance more models on top of all those GPUs. But due to ongoing geopolitical tensions, Chinese AI companies have had to employ a different strategy. Ever since the AI Boom began, the U.S. has consistently enforced export bans on AI chips to China, thereby limiting the number of chips Chinese firms can buy. They haven’t been able to employ the “more, more, more” approach that Microsoft (MSFT), Alphabet (GOOGL), Amazon (AMZN) and others have in the past few years. Instead, Chinese developers were forced to embrace a “do more with less” mentality. That led DeepSeek to focus on an innovative blend of engineering techniques to create a super-efficient AI model. Now, I won’t pretend to understand these techniques at a granular level. While I am familiar with different AI modeling composition, I am not a world-class developer myself. However, I have studied this topic enough to have a general understanding of what made DeepSeek’s model tick. And spoiler alert: it’s pretty neat. An Innovative Architecture Makes DeepSeek Tick At the heart of DeepSeek’s breakthrough is something called a Mixture-of-Experts (MoE) architecture. In short, most AI models today are created to be omnipotent. They try to be doctors, lawyers, and engineers all rolled into one – experts on a near-infinite number of subjects. When you ask a general model like ChatGPT a question, its entire architecture “wakes up” to answer because all its expert knowledge is rolled into one model. But DeepSeek employs a MoE architecture. In a sense, it’s created a room of multiple experts, where each is separate and distinct. It is a model composed of multiple topic-specific sub-models. Therefore, when you ask DeepSeek a question, the only part of the model that “wakes up” is the expert sub-model relevant to your question. Thanks to this modular approach, DeepSeek can save an immense amount of computing power on each query because only part of the model is roused per query. According to DeepSeek’s own numbers, its V3 model is trained on nearly 700 billion parameters. But only about 20 billion of those – or less than 5% – are activated simultaneously at any given time. This drastic reduction in activated parameters is partially what has allowed DeepSeek to create an AI as good as leading models for ~95% lower costs. To be sure, DeepSeek isn’t the only company in the world employing MoE architecture. But through a variety of novel engineering techniques, it appears to be the firm that was able to perfect and scale it. And that is how a virtually unknown Chinese AI startup disrupted the entire global AI industry. But what does that mean for others in the space? Is this the start of the AI Boom’s own dot-com crash? On the contrary, we actually view DeepSeek’s efficiency breakthrough as great news for the industry – and great news for AI stocks, too. Gaining Ground on AGI I fundamentally believe that the best solutions require a combination of innovation and capital. Solutions powered by unlimited innovation, but without ample money, will run into resource challenges. Those powered by unlimited money, but without ample innovation, will run into development challenges. But with unlimited money and unlimited innovation, initiatives will advance far and fast. That’s why I’m highly confident that, thanks to the DeepSeek efficiency breakthrough, the stage is now set for the world’s most powerful companies – like Microsoft, Amazon, Alphabet, Meta (META), etc. – to go all-out with ‘unlimited’ money and ‘unlimited’ innovation, putting them on an accelerated path to creating superintelligent AI that transforms the world forever. DeepSeek is entirely open source. Its MoE architecture and other underlying designs are public. Big Tech companies can and will copy those techniques and incorporate them into their own model building. Then, on top of that, they’ll invest billions of dollars to take those models to the next level – putting us one step closer to artificial general intelligence (AGI). So… as it relates to all the fears about DeepSeek’s breakthrough meaning the end of Big Tech or the AI industry… we think they couldn’t be further from the truth. The Final Word on the DeepSeek Breakthrough’s Lasting Impact The truth, in our view, is that the DeepSeek breakthrough will meaningfully accelerate AI model advancement and meaningfully boost the odds that Big Tech companies achieve AGI much sooner than previously thought. Does that mean you should be buying the dip in AI stocks after this week’s gut-wrenching selloff? Absolutely. We wouldn’t get too bullish on AI hardware and semiconductor stocks like Nvidia. Some will likely lose their pricing power as the hardware in this industry becomes increasingly commoditized, leading to lower margins and slower profit growth. But we would get exceptionally bullish on AI software and services stocks like Meta and Apple (AAPL) or companies like Spotify (SPOT), Intuit (INTU), ServiceNow (NOW) and Atlassian (TEAM) – those creating and deploying AI models and solutions. For those firms, this DeepSeek breakthrough looks like a big-time opportunity. That’s why we’re telling our subscribers to look for plays in the “AI Application Layer,” with the companies developing cutting-edge AI applications. One of the more interesting companies in that sphere right now is xAI – Elon Musk’s AI startup. Over the past few years alone, it has grown from obscurity to sport a multi-billion-dollar valuation. We think it could be a major beneficiary of the DeepSeek breakthrough. And while it isn’t yet publicly traded, we’ve found a compelling ‘backdoor’ way to play it. Learn more about how to invest in xAI right now. Sincerely, | 
No comments:
Post a Comment