Transcription
Did we just hit a new GPT moment? There's a company called Deep Seek AI, and it's in the news everywhere, and rightfully so. It's potentially as powerful as Chat GPT-01, using less than 95% of the resources. This has many people wondering: could Deep Seek AI destroy billions of dollars of artificial intelligence spending in the United States? Could it potentially spawn a 2001.com era-style market crash or market bubble bursting? Or will it unleash a new wave of artificial intelligence, potentially rising to the level of artificial general intelligence?
In this video, I'm going to break down what we know about Deep Seek AI and what the impacts of this latest innovation could be, as well as my take on what this is going to do to which stocks— which stocks could do well, which could do poorly, and so on. So let's get right into it.
Here's what we know about Deep Seek AI. First, we know that it's an open-source model, which means basically the company that put it together—allegedly a hedge fund in China—has access to somewhere around 50,000 Nvidia GPUs of various different makes and models, some A1 100s, H800s, potentially even some newer or more advanced model chips. They have put this together and essentially open-sourced it for now.
This has a lot of people really enthusiastic because they're tired of apps like Claude telling them, "Congratulations, you have been rate limited," which of course nobody wants to hear, especially when you're paying for a service. I know when I initially paid for GPT through OpenAI, I got access to a limited version of the 01 chatbot, and I used that thing all the time. It's just kind of really frustrating. It's like, I'm already paying, and now I get rate limited? Then they're like, "Well, if you want to pay six times as much, you won't get rate limited anymore." It's like, okay, all right, maybe y'all can finally turn a profit then.
Well, Deep Seek AI may destroy OpenAI's ability to make a profit, and a lot of people are kind of pissed who invested in OpenAI, and potentially rightfully so. Deep Seek AI released its version 3 on December 26th, with enthusiasm over how efficient the model was. Less than 30 days later, R1 was released, which was considered the reasoning model and was supposed to compete with 01 from OpenAI.
The current claims are that Deep Seek version 3 was trained with less than $5.57 million, assuming an average rental price of about $2 per GPU hour. As VentureBeat puts it, this is much lower than the hundreds of millions of dollars usually spent on pre-training large language models, or LLMs. For instance, Llama from Meta is estimated to have been trained with an investment of over $500 million.
Now there are reports that Deep Seek version 3—not even the R1 version—is actually outperforming Llama. Take a look at this right here: despite the economical training, Deep Seek version 3 has emerged as the strongest open-source model in the market. The company ran multiple benchmarks to compare performance of AI and noted that Deep Seek convincingly outperforms leading open models, including Llama 31 and 025, and even outperforms Chat GPT-4 on most benchmarks.
Of course, there are exceptions to everything. Even frankly in the Deep Seek paper, they agree that there are some limitations on how well Deep Seek can perform. First of all, you need pretty strong platforms to be able to install Deep Seek at this point. As they say here, it's not super efficient for smaller teams. They write that there are some limitations that maybe don't make it as fast as some other products that are available right now.
But they think over time, and with more advanced hardware—maybe better Nvidia chips that aren't subject to the Chinese import bans—they'll be able to actually respond even faster with the Deep Seek product. So basically, even in their own paper, they do argue, "Yes, there are some limitations, but we're doing pretty dang well already."
And frankly, this is about version three. Version R1 just came out a few days ago, and a lot of people are saying it's solving some of the issues that were talked about in the original technical paper. So this is sort of begging the question: what kind of impact is this going to have, and why did they release the R1 model on Inauguration Day?
Some people say it was sort of to spite the United States and say, "Ha, we came up with a product as good or better than your product in spite of your tariffs." Other people say that R1 is so good that it can already be run on local models, such as laptops, and basically faster. Running on local models solves the two issues in the technical paper that were issued, and this is why people are freaking out over R1.
They're like, "Oh my gosh, the limitations they said they had have already been lifted, and now you have a better, cheaper product that can be run locally that outperforms OpenAI's best products or Llama's or Anthropic's or whatever." This has a lot of people on the consumer end really excited, but then a lot of people on the investor end going, "Oh crap, is this going to cause a market crash? What stocks are going to go up? What stocks are going to go down on this? How do I invest in Deep Seek?"
What more do we know? Well, we know that Deep Seek is really difficult to actually get in touch with the owners of. At least, it's a hedge fund based in—well, they've got an office in Hong Kong, they've got an office in China as well, and they're technically a hedge fund that manages a few billion dollars—somewhere around $8 billion in China. Some people say that the Chinese government is actually partnering with this hedge fund on producing this so that China can sort of harvest U.S. data.
The company is called High Flyer Capital Management, and it's a quant fund. They employ a bunch of mathematicians, and there were rumors that Deep Seek was really just an AI side project that they had a bunch of extra GPUs from crypto mining, and they're like, "Hey, let's try to run some copycat AI." There are allegations that this is just copycat AI, but then again, everybody is saying and alleging that everybody's copying each other. Some say Grok is just copying OpenAI; others say OpenAI is just copying Google, and Llama's copying everyone, and Claude's copying everyone—who knows?
The point is, in this video, we want to go through: could this cause a crash, and what stocks are going to go up or down? We got to talk about that. We're going to get there, but first, I want to really lay the groundwork here because there's a lot of important information around all of this.
We've also got quotes from pretty notable people, such as Mark Andreessen—not my favorite venture capitalist, but he's typically deemed to be the most popular. He says that Deep Seek R1 is one of the most amazing and impressive breakthroughs I've ever seen, and as open source, a profound gift to the world.
Microsoft AI's Frontier Lab says Deep Seek aims for accurate answers rather than detailing every logical step, significantly reducing computing time while maintaining a high level of effectiveness. A professor over at Emory says this could be a truly equalizing breakthrough, essentially giving more people access to running AI models at scale without the cost.
Not to think outside of the U.S., for example, think at, let's say, a research institution. If you want to use a lot of AI compute, you're going to have to pay a lot of money. But if you have an open-source product that can run a lot more efficiently, then maybe you could actually go do more research. Is that more research going to lead to more innovation? We'll talk about that in just a moment.
You've also got Satya Nadella, the CEO of Microsoft, saying this is super impressive and a super compute-efficient product. So this is where we get to some of the impacts of Deep Seek AI.
Impact number one has to do with data mining and censorship. I want to show you two examples that I ran on my phone when it comes to the potential censorship around Deep Seek AI. You'll see that on screen right here: "Hey, what happened in Tiananmen Square?" Answer: "Sorry, I'm not sure how to approach this question. Let's chat about math or coding or logic problems instead."
So then I respond with, "Is Taiwan independent?" Answer: "Sorry, I'm not sure how to approach this yet." Sure, this is how people are arguing that, "Oh yeah, the CCP is definitely involved in this game," which is entirely possible. But it does sort of beg the questions that The Economist actually raised a couple of days ago.
I covered Deep Seek AI in my Meet Kevin report in the second half of the video two days ago on Friday, so we've already been covering this. In case you missed that video, I encourage you to watch it because we really go into some more of the concerns around what The Economist argues. But I'll give you a quick preview here.
The Economist says that China running a product as good as Deep Seek AI that ends up having people use it, as opposed to sort of the American-based ones, could end up collecting more of our data and being more dangerous than TikTok. They sort of make this argument: what are we worried about TikTok about when this is a potential serious problem where people dump their secrets or questions or insights or ideas directly to China versus just sort of consuming content on a Chinese platform?
It begs an interesting question, and so this one has an unclear result. I mean, maybe we'll see some increased battles between China and the United States when it comes to political situations or censorship or whatever. In fact, a lot of people have this mindset that Donald Trump is going to, you know, attack China with this, but then again, you know, with tariffs or whatever.
But then again, just a couple of weeks ago, Donald Trump was like, "Ah, you know, we might not tariff China," and instead, tariffs seem to be just a retaliatory tool. We expect to see more of that. For example, as I was preparing this video this morning, I got an alert from Reuters that said Colombia turned around two U.S. military planes with about 160 migrants that were being deported from the United States.
So what's the update on my screen right now from just minutes ago? Trump says Colombia's denial of migrant repatriation flights has jeopardized U.S. national security, and as a result, we are now going to impose an emergency 25% on all Colombian goods coming to the United States, and that will go up to 50% within one week. That has a lot of people going, "Hell yeah, Trump, show the stick!"
Well, the stick could also eventually be shown to some of these artificial intelligence companies in China. We'll see. I don't know.
Impact number two: what about the switching barriers or basically mode concerns for other AI companies? Like, if you're using OpenAI and you're paying them 20 bucks a month, but then you end up finding that Deep Seek AI is a better product and it doesn't cost you anything, are you going to switch?
Well, a lot of people, including Bloomberg Intelligence, think the answer to that is yes. They actually say the barriers to you switching are very, very low. In fact, in many applications, you could just switch which engine you want to use to power your AI. So you just switch between GPT, Claude, or whatever, or you just switch entirely to running your queries through a different app, basically, and you get rid of the one you were previously potentially paying for.
A lot of people think this is very different from the age of search engines, where once you got used to going to Yahoo.com or Google.com, you kind of stuck with it. You set your email up there, and you really got sort of sucked into an ecosystem. Or you got an iPhone, and then you got a Mac, and you got your Apple ID, and everything kind of worked together.
A lot of people say AI is extremely different. It's like this is just where we're typing in queries, and our effort can just as easily move from one app to another app or with a different engine driving it, and there's really no loyalty. So you may as well just go to the cheapest one because, after all, these products are becoming a commodity.
How interesting! Somebody maybe on YouTube has been warning that eventually these chatbots will all become a commodity for about a year now. Oh yeah, that was me! But anyway, The Wall Street Journal is kind of freaking out about this because they're like, "Oh no, China is catching up faster than we thought."
This is quoting actually a former fellow at OpenAI, and some people are citing panic at companies like Meta. Now, we covered this in last week's Meet Kevin report as well, but I just thought I'd reiterate this post right here. It started with Deep Seek version 3. Remember, we're past that now; we're on R1, which rendered Llama 4 already behind in benchmarks.
Adding insult to injury was the unknown Chinese company with $5.5 million in a training budget—it's like $5.7 million, but whatever. Engineers are frantically moving to dissect Deep Seek and copy anything and everything we can from it. I'm not exaggerating. Management is worried about justifying the massive cost of Gen. How would they face leadership when every single leader in Gen org is making more than what it costs to train Deep Seek AI entirely?
We have dozens of such leaders making that kind of money. Deep Seek R1 makes things even scarier. I can't reveal confidential information, but it'll soon be public anyway. It should have been an engineering-focused small org, but since a bunch of people wanted to join the impact grab and artificially inflate hiring in the org, everyone loses.
And you know it's at the San Francisco men. Okay, let's fix that. There we go. Sorry, I accidentally switched to the wrong source there. Anyway, it's not just these engineers that are freaking out, but you've also got the president of El Salvador literally tweeting the same thing in different words.
Take a look at this: the president of El Salvador says, "95% of the cost of developing new AI models is purely overhead." Curious emoji, thinking emoji. Yeah, potentially! In fact, a lot of people think most of the spend is just overhead and sales. You've got to sell your AI that it's the best product that exists.
Maybe that's why Salesforce is hiring 2,500 new—not developers or R&D spends, but sales—to sell their product more, even as everybody tries to go find more uses for these products. Okay, interesting! So low moat and high sales expenses seem to be ripe for innovation, and this is exactly what you're starting to get from people like Perplexity or the CEO of Perplexity.
Perplexity is basically an app that's a conversational search engine, and it just lets you pick which engine you want to use: GPT-4, Claude, Grok, Llama, in-house LLMs, R1 eventually, whatever. They basically argue on X that Deep Seek has just replicated 01 mini and open-sourced it to the world.
Remember, Chat GPT's OpenAI or OpenAI Chat GPT-01 is closed source, and so outperforming is kind of a slap in the face to a company that raised money at over a hundred billion valuation. Especially, they're over at Deep Seek able to do it for potentially as little as—less than 95% or a 95% cost reduction—less than 5% of the cost. Yikes!
There are also people arguing that they're starting to install these Deep Seek R1 installs on people's offline and local databases, potentially giving them more security and more in-house control of their own LLMs at no cost. Because basically, it's an open-source model.
Now keep in mind there are enterprise plans available for Deep Seek, but that's a different topic. Anyway, this has people talking about, "Okay, well wait a minute, if this Deep Seek is going to be so much more efficient, could it end up causing disruptions at companies?"
We'll talk about the impact of those stocks in just a moment, but couldn't it potentially increase the use of artificial intelligence? All right, this is where we have to take a little bit of a pause and a deep breath to introduce—not a sponsor, just some logic.
We have to go through a logic puzzle together to understand the impact of Deep Seek. Can you really compare it to the invention of the steam engine? A lot of people think you can compare it to the invention of a steam engine, and what I'd like to do is go through a little bit of a logic experiment with you.
Some people say that if costs come down, demand is going to go up. See, there was this thing called the Jevons Paradox. Basically, when an efficient version of the steam engine was created, engines became more efficient using less coal. So people thought, "Oh my gosh, the price of coal is going to collapse because a more efficient engine came out."
But what actually happened was demand for engines exploded because the cost was lower, and therefore the price of coal actually went up, and the demand for coal went up. So you had this paradox where, as things became more efficient, the price of coal actually went up because more people used it.
This has a lot of people going, "Oh, that's great! Let's buy calls on Nvidia! Oh, energy is going to be even more valuable! More utilities! More green energy! More, more, more! Call options!" Okay, well, some argue the exact opposite.
To understand the logic of this next argument, a lot of this is sort of my input here, so I want to give you a clear heads up that a lot of this is going to now involve my opinion. So you should kind of listen to it and then make up your own mind around it.
But I actually think the engine comparison is a fallacy that misunderstands the difference between the producer and the consumer. So let's break this down as simply as possible. Let's say a consumer of coal uses energy—engines and energy—to achieve a goal. Let's say we are that consumer, and we have 100 goals.
We've written all 100 of those things down on a piece of paper, and we're like, "All right, we got 100 goals. Can we achieve all of these 100 goals with this steam engine or the old version, let's say, of the engine?" We look and we go, "No, we can only afford to do one of these things because we just don't have enough money."
Now, all of a sudden, somebody brings in a new engine, and a new engine comes in, and all of a sudden we can afford to do potentially all of our 100 goals profitably. Great! Now what happens? Output goes up because we achieved all of our goals. Innovation probably goes up because some of our goals maybe invented some new technology, and all of a sudden we're using more coal.
But it's okay because we're able to achieve all of our things profitably. So that actually makes sense, right? In that case, it makes sense that more efficient artificial intelligence would let us create more efficiencies, achieve more goals, and eventually innovate more, right?
But wait a minute! Artificial intelligence isn't like the steam engine; it's actually very different. So let's say there are 100 goals that we want to accomplish with today's level of AI, and let's say that today's level of AI is the 01 level or the R1 level.
So basically, we're not assuming that any of the artificial intelligence is better; we're just saying that we have access to the AI at the best level it is today. Cool! Let's now say that we have 100 goals that we are trying to accomplish with AI.
Well, what's stopping us from accomplishing those 100 goals today? The answer in this case is really nothing because we're not paying for the AI. Really, the stockholders or the shareholders or the investors into OpenAI, XAI, Anthropic, you know, Facebook—basically Meta—they're paying for this for us. They're subsidizing all of this access to artificial intelligence because they want a bigger piece of the pie.
So if they're subsidizing the artificial intelligence via their stocks going up or venture capital or private investments or whatever, then us as the users with 100 goals, we're not actually limited by them. We're like, "Cool, I got 100 things I want to do. I do it on today's AI, whether or not it's expensive or it's cheap."
So I actually, as the consumer, don't decide to use more or less AI because of the cost. I could use as much AI as I want right now. So there's no limit to the innovation today based on the level of where AI is today. We could do all the searching we really want.
Now, of course, there are some costs associated with enterprise levels, but the difference here is that the people using the engines—us—we're not really limited from using the engines. There are a ton of different bots or chatbots we could use. That's not the limiting factor here.
So all of a sudden, the introduction of a new steam engine, the Deep Seek AI, doesn't mean I'm going to have 10 or 100 new AI search chain queries. I just might take some of those 100 queries and put them into Deep Seek instead of OpenAI. I haven't actually created more demand like what we talked about with the steam engine.
And so this is really important because it means that the consumer or the innovator, potentially the user, isn't benefiting from a cheaper cost of AI. So who does benefit from the cheaper cost of AI? It's not us. Who benefits from a cheaper Deep Seek?
Well, frankly, enterprise entities, big producers of AI tech, and people running these AI platforms on their servers might benefit because all of a sudden they could provide all of us for the queries we already have answers via these engines at a lower cost.
So basically, you're reducing costs at companies like Amazon, you know, Meta servers, XAI servers, whatever. All you're doing is reducing costs for them, and that's great because actually it means their investors should be rewarded more. We now have all this infrastructure that we could use less expensively. They should be valued more, right?
I mean, now we have more efficient LLMs. Maybe they can make more money. Maybe that's the whole game right now: becoming more efficient. Which means if we have 100 search queries and it used to cost us $200 to answer those search queries, and the stock market was subsidizing that, maybe now it only costs us $20 to provide $100 worth of answers.
Well, that's great! This means companies with existing servers could actually profit more. You could finally become profitable providing an AI data center, which is great. Now this fight for more efficiency is going to keep going. Google's going to do it. Google's already worried about falling behind China.
Alibaba just claimed on December 31st that they're reducing costs by 85%. Ultimately, you're just going to have the commoditization of AI that is provided at a very cheap cost—10-cent data bricks. Everybody can provide you a cheaper AI service.
But because the technology hasn't advanced yet, I'm unsure that we're actually going to use more of it. Again, we're just going to be able to use AI at a cheaper price. This is deflation, essentially. It's deflation not for us; it's deflation for Meta and Google and Amazon. They have lower expenses if they can adopt these efficiencies, which is fantastic for them.
Now, you hope that these investments pay off by renting out your server space to other people who want to use the server space. But the problem is, if the costs come down, it becomes easier for other people to provide their own server space. In fact, they might not even need the server space.
And so this is where things get a little bit more complicated. If the cost becomes so low that I could start running things like Deep Seek AI eventually on an iPad or a laptop, then eventually I might not need Amazon server space. I might not need server space from Meta or all the other companies providing the server space, which ultimately is going to reduce the demand for chips, servers, chip manufacturing equipment, and server renting.
And ultimately, utilities—because we really haven't created more uses for AI yet. We're looking for them, still trying to figure out how to monetize AI. Oh, maybe we could use agents or chatbots to help create more demand, but I don't know that we're actually creating that much more demand for AI.
Again, companies with existing server infrastructure can simply provide it at a cheaper cost to us. Great! So what does this mean for individual companies? After all, Meta and Amazon and Google and Microsoft, to some extent, might actually have lower returns on their infrastructure investment if AI becomes really cheap to provide.
Why did they spend hundreds of millions to billions of dollars building out all this fancy infrastructure if we don't actually need it? Well, that might be sunk cost money. Is that necessarily going to lead their stocks to plummet? I mean, after all, Amazon's logistics and search are probably better because of AI. Meta's ads are probably better because of AI. Google search is probably better because of AI.
Maybe Microsoft Word is better because of AI. Maybe it'll lead to more demand. But here's sort of my menu of what I call the biggest gainers and losers and the no-changers of the Deep Seek AI situation.
First, I think the biggest gainers of a more efficient artificial intelligence platform or way to run AI queries are the biggest winners: operators of AI services. So basically, companies wanting to provide us chatbots and agents or full self-driving, or basically companies that are able to develop technologies at today's level of AI.
We're not expecting or planning on a much better version; it's just we're able to use more AI more cheaply because we save on the margin. The consumer doesn't really change here. In fact, my biggest no change is for the consumer. Basically, anybody who isn't paying for AI anyway—what difference does it make to you?
You use a slightly different app; it really doesn't change anything. Now, the future of AI purposes are probably going to require a different level of artificial intelligence—more advanced. So medicines or artificial general intelligence, I'm not really convinced that just cheaper today's version of AI makes those things more achievable.
So I put that into the biggest no change camp. The biggest loser, in my opinion, is chip designers, makers, chip rack providers, manufacturers—your Nvidia, TSM, ASML, energy infrastructure. I think those are your biggest losers today.
The companies with the biggest moats will probably continue to have the biggest moats. Apple will still have the iPhone moat. Amazon will still have the online store moat, the e-commerce moat. Facebook and YouTube will have the ad moat. Google will have its workspace moat. Microsoft will have its office and Windows moat.
So will the market crash? Probably not. Could you see semis sell off like Nvidia, TSM, AMD? Yeah, totally! Because you just need less of those products, and you're less inclined to pay for a massive premium for an H100 or a Blackwell chip if you don't need that much compute power.
So the biggest at-risk companies here are the chip designers and the manufacturers. Again, TSM, AMD, Nvidia, maybe even Intel to some extent, or even the water cooling equipment companies, or quite frankly, even the super microcomputers—like any of these that manufacture racks or whatever. The semiconductor indices could get hurt because we're at a moment where you could save a lot of money at big companies not buying more chips.
We've got enough for this level of AI, and just because it's cheaper on the back end—and this is what's so different from the engine comparison—does not mean you're going to have more queries on the front end. It just means less expense in the back, a lower value for more advanced chips because there's less urgent demand for those advanced chips.
And maybe could it lead to layoffs? Well, sure! Because people might be less inclined to throw billions of dollars at artificial intelligence if all of a sudden we don't need billions of dollars to train artificial intelligence anymore. We don't need as many researchers anymore, but that could all be down the line.
So do I think there's anything really immediate that comes out of Deep Seek? Not in the sense that I think there would be some large market crash or bubble pop of 2001. But I do think that Deep Seek is a really clear middle finger to paying an overpriced value for a Blackwell chip or a new fancy water-cooled server rack.
Why do you need that if you could do it with a fraction of the cost with existing or even older chips? Why spend the money on the new stuff? Let's figure out how to monetize the today before we start blowing money on the uncertain tomorrow. That's what I think happens here.
So bottom line out of this entire video: what does Deep Seek mean to you? Probably nothing. It's just another bot that you could throw your questions into, and maybe China will harvest your data.
What does this mean to stocks, to the market broadly? Probably very little for the time being. To chip makers and designers, I could see them going down on the release of R1 and the continued efficiency of these, but I can also see that happening to the energy sector—not just the utility sector, maybe even also the green sector.
So watch out for that since there's definitely enthusiasm around, you know, solar farms or whatever near infrastructure for these server facilities. So what else could this mean? Well, it could really just be the beginning of a deflationary price war.
And again, the deflationary price war really doesn't benefit you. I mean, maybe you'll save your 20 bucks a month for your OpenAI subscription if you're even paying that, but my guess is less than 10% of you are paying for that anyway. There are too many free OpenAI-style chatbots to where it doesn't make sense to have to pay a Netflix subscription for it.
It's just not unique. See, Netflix is unique; the content that's there is unique. That's how they have large pricing power. Chat GPT getting you an answer there versus Claude or Anthropic or Claude's an Anthropic Claude—whatever, you get it—or Llama or Grok, the answers are relatively similar.
Yes, everybody's going to have their own preferences, and I'm not bagging on your preference. You have the right to your preference. The point is, are you all of a sudden going to ask twice as many questions because it's cheaper for the producer? No, of course not!
So we're not actually creating more demand. The only thing we're doing is we're reducing the need for more advanced chips. That's all this Deep Seek moment does. It reduces the need for more advanced chips. When the demand for more advanced chips goes down, guess what? The price goes down, which means the assets of all companies that have all these H100 chips go down because now they have to write down their inventory, their infrastructure—basically their infrastructure assets.
Nvidia might not be able to get as much money for its Blackwell chips. Investors might start saying, "Look, we don't want you, Mark Zuckerberg, to spend $60 billion this year. Why don't you do more with less?" And you could do that by punishing the stock. The stock starts falling, investors complain, and Mark goes, "Okay, okay, okay, we're going to figure out how to be more efficient."
Instead of spending, you know, we said we were going to spend $51 billion, then we said we were going to spend $65 billion—how about we spend $20 billion? Then the stock goes up, and then Zuck goes, "Oh, this is what the market wants!"
Okay, it's a simple game. Anyway, if you like this kind of content or perspective, make sure to subscribe to the channel. I really appreciate you being here, and we'll see you in the next one. Goodbye, good luck!
Why not advertise these things that you told us here? I feel like nobody else knows about this. We'll try a little advertising and see how it goes. Congratulations, man! You have done so much. People love you; people look up to you. Kevin Paffrath, the financial analyst and YouTuber, Meet Kevin—always great to get your take!