Speaker A
If you bought any hardware in the last year that can run AI locally, you probably made the best deal of your life. If you are waiting, you're probably getting screwed. Now, I know the narrative. The price of the RAM is going to come down. There is a new card coming out soon. A new chip has been developed. Let's wait for it. But the truth here is different. [clears throat] And I paid approximately $3,500 six months ago. Now it is $1,000 more expensive. That is half of the story. Now I can run models that are much better than when I bought it. Now this machine can make better images, better video, better text, can do AI work. So the value of this hardware is not going down, and now what I can achieve with this hardware is much better. Just a week ago, the Deep Seek Flash version 4 came out. This is the latest version, extremely powerful, and can be run on two of these. You can connect them through this really high-speed connector. You can run a model that has a 1 million context window that in some scenarios is comparable with the Frontier model. It's not as powerful as them, but it's not far either. And as a token generation that is as fast as those Frontier models, then now you can own it, run it in your AI lab, build around it, have unlimited token generation, so no subscription, no you reach the limits, can work for you 24/7 with privacy. That is a lot. So two of the biggest problems that we have with cloud AI is one, of course, privacy. You share all your data, you share your client's data, your IP, the recipe of success of your business to those AI corporations that I'm going to use it to train their own model. And then there is another problem. Those AI labs keep tweaking, adjusting their model. So those models, even though they keep having the same name, in reality, they're not performing the same every single day. Now, when you have your own hardware, you set it up, you optimize for it, and it's going to be stable, constant. It's not going to change unless you change it. That level of stability sometimes can make a huge difference. Think about it in these terms. This is the threshold that you need for your AI to work at the level that you're happy with. So the moment that you go over the threshold, you can do the work that you want. You can serve your customers. You can run your business happily. Anything above is fine. But if you go underneath it, now you start to have problems. What those corporations are doing is trying to adjust the threshold to make them more profitable. Sometimes it is about optimization. So they want to spend less electricity, less tokens, and therefore they lower the intelligence of those models. They lower the time of reasoning, and sometimes they want you to just move to the more expensive tool. So they're optimizing for their business, not for yours. But today I want to help you decide if you should buy or wait. So let's look at some data because the data-driven decision is probably something that makes sense, even though have in mind that most likely you will decide from your emotion. So let's try to control those emotions, bring some data in the decision process. Why the price of RAM at the moment is really hard to predict. There are two factors. One, the price is really high. So we had over 300% increase in price in the last year. Two, we see that in the US they've been building so many data centers that now they don't have electricity even for the ones that they already built. So now there is a seven years queue to have access to that kind of amount of electricity because new power stations need to be built, new turbines, new solutions, the grid needs to be expanded, and that will take years. So the feeling now is if they're not going to buy anymore, the stocks, of course, they're going to adjust. But also the price of RAM should come down. But that is the US reality, which is extremely important because they are the ones that are spending the most at the moment on AI. But there is the rest of the world, billions of people that need to be served. If something is becoming clear to everybody, it is the importance of sovereignty with your AI. So at the nation level, they need to have their own data centers. Before we go ahead, let's do a little bit of housekeeping. This channel is also connected to a Discord server. What I'm building is a community of like-minded people that use AI to be augmented. If you're interested to collaborate with others, learn from other peers, find the link in the description for the Discord community. Also, for any video that I do, there is a SupStack article that you can use to read more and go deeper. And last thing, help me grow this channel. So, like and subscribe. And let's carry on. They can't everybody rely on the US ones. And we know why. They made their model not accessible for people that are outside the US. They make their model extremely expensive. For some countries, it's just too much. They learn on your data. They spy basically on you. That is also not acceptable. And you cannot be, for such important technology, dependent on another country that gives you access or not. So we saw already Europe changing on a lot of elements. Now within the governments, they're trying to use only Linux instead of Windows. They're trying to move away from WhatsApp, Microsoft Office, Zoom, and a lot of other software that are used for communication and work and be dependent on a foreign country. This is the new direction for Europe. So of course, they already started investing in AI data centers, and this is a signal now the demand is not going to be as fast as the one from the US, but there are many nations in Europe that will want to have their own AI data center. And look at France, they already invested, they have Mistral, they invested over 830 million on new data centers. It's small money compared to what is happening in the US, but that is the start, that shows the trend, and the same thing will happen with other countries. Now there is also to consider China. China had a problem until not long ago. They couldn't have access to hardware because they couldn't buy Nvidia's latest hardware for AI. So that stopped them from accelerating in that direction. But now they have the hardware to create their own data centers. Now they can start optimizing their model to run on their hardware, and that is changing everything. So China, what they did instead of pushing for the maximum intelligence, it pushed for something that was optimized. So it could run on smaller hardware. It could have the same intelligence at a fraction of the price. Now we see models from China that are extremely cheap. But also we see another important trend that the China business model works on open-weight models. That means that any data center can get those models and offer them to the clients. That means they're going to be cheaper, as I said, almost as powerful, and for 90% of the work, they're going to be good enough. And they keep optimizing those models. So the goal for them is to create an intelligence that can be affordable for the rest of the world because not everybody can afford to work with Anthropic and while OpenAI at the moment it seems like cheap, in reality, they are spending $1.60 for every dollar that you give to them. At the moment, the price that we see with OpenAI, they're not the real price. They will have to triple those prices if they want to stay in profit. So going back to the RAM price, why from the US perspective, they're going to slow down and buy less. And so you could think, okay, the price will go down. In reality, the rest of the world needs to catch up. So they're going to start building. The demand is going to get there. As you can see, you're not using AI less and less. You use AI more and more. So there is more and more demand. Now, let's look at wait for the next thing. So if you see the Mac Studio M5 Ultra will come out by the end of the year because it's being delayed. The price is unknown. Expect really expensive. Also, there is the RTX Spark that is coming out.
