The End of Nvidia? This AI Changes Everything
100x Cheaper and Open-Source?!
This week started in a way I could never have imagined. Obviously, you can expect the market to be shaken by many things, but what doesn’t cross your mind is that it could be shaken by an AI that has been created for just 99% of what giants like OpenAI are spending.
You think, "If there's something better than this, some company insider must have discovered it." But then, a Chinese company with four friends and $5M launches a product that can compete with OpenAI and even improve upon it.
Now, everyone is talking about Nvidia and whether so many graphics cards are really necessary, but I think the main takeaway is not that—it's something no one is considering.
First, I’d like you to understand what DeepSeek is.
DeepSeek, a new Chinese AI model, appears to be a turning point for AI. Ranked #1 in the App Store and dominating media coverage worldwide, with nearly 2,000 published articles today
First of all, and just as a personal reflection… A breakthrough in efficiency this big, if true, is something I see as a positive step for AI.
Here’s the problem: looking at the reaction of certain stocks, maybe the real issue isn’t DeepSeek itself, but rather the huge expectations placed on the “Magnificent Seven.”
They are priced as monopolies because they are monopolies… At least, until now.
Let’s take a closer look at what’s happened:
Training advanced AI models today costs an absolute fortune.
Companies like OpenAI or Anthropic invest over $100 million solely in computing power.
This requires giant data centers with thousands of GPUs, each costing $40,000.
But then DeepSeek came along and said: "What if we do it for $5M*?" And they did.
Their models not only match but surpass GPT-4 and Claude in many tasks. In the AI sector, this has caused a revolution.
How did they do it?
While traditional approaches use numbers with extreme precision, DeepSeek proposed reducing that precision to the minimum necessary while keeping functionality intact. This simple adjustment allowed them to cut memory usage by 75%.
It’s not just about the cost—it’s about the savings for clients. And by “clients,” I don’t mean you or me, but large corporations using AI, like META, AMAZON, and MSFT.
That said… I’m surprised no one is talking about how slow it is.
They created a “multi-token” system.
While traditional models process word by word, DeepSeek reads full sentences. It’s 2x faster and 90% as accurate.
When handling billions of words, this is crucial.
Instead of building one gigantic model that has to handle everything (as if a single person were a doctor, lawyer, and scientist all at once), they opted to create smaller, specialized models that only activate when needed.
Traditional models keep all their parameters active at all times, consuming vast resources. DeepSeek, in contrast, uses a more efficient model: out of its 671 billion parameters, only 37 billion activate simultaneously. It’s like having a great team but only calling in the experts when necessary.
The results are impressive:
Training cost: $100M → $5M
GPUs: 100,000 → 2,000
API costs: -95%
Runs on gaming GPUs, not expensive hardware…
This poses a serious threat to the semiconductor industry, which is expected to reach a trillion-dollar valuation by 2030.
The entire project is open-source. Anyone can review, test, and understand how it works.
Technical documentation is available for anyone who wants to learn.
It’s no longer necessary to spend billions or rely on big corporations to develop competitive models.
Now, even small teams can enter the game with more accessible hardware.
However, the biggest loser here is Nvidia’s business model, which is based on selling extremely expensive GPUs with massive profit margins.
They are priced as a monopoly because they are one—at least, until now.
If this technology allows for the use of cheaper GPUs for advanced tasks, the market could shift dramatically.
DeepSeek seems to have asked itself: “Why keep adding more hardware when we can do things more intelligently?”
Let’s be clear—this efficiency gain is fantastic news for the world, but not for those whose business is built around it.
The implications could be huge:
AI development will become much more affordable.
Increased competition in the market.
The advantages of big tech companies seem far less solid.
Hardware costs will plummet.
OpenAI and Anthropic are likely already working to adopt these innovations. However, efficiency is now the standard, and there’s no turning back to the “just use more GPUs and money” approach.
By the way, perhaps taking advantage of the surge in users, DeepSeek has released another new open-source model today: Janus-Pro-7B. It’s multimodal (it can generate images) and surpasses DALL-E 3 and Stable Diffusion from OpenAI in GenEval and DPG-Bench tests.
But beyond the technical aspects, here are key takeaways that very few people are considering:
Open-source: We are looking at one of the few AIs that is easily replicable. Anyone who wants can copy, modify, or improve the source code—without exception.
When OpenAI was created, it was named that way to make AI available to all of humanity.
I firmly believe that open-source technologies (e.g., Bitcoin) give us much more freedom. Anyone with an internet connection and some ingenuity can build a successful startup. It doesn’t matter if you’re the richest or the poorest person in the world—access to information is nearly the same for everyone.The world is nearing the end of the “work era.”
We’ve been seeing this coming for a long time. Bit by bit, robots and AI will take control of everything, and humans will be left with universal basic income.
Humans are facing a paradigm shift where AI will be a far better professional than they ever could be, and they’ll need to reassess their role in society.
This, although it might not seem like it, is the gateway to paradise, where we can live without working.
When you begin to understand this, you realize that society will become even more divided—between those who embrace technology and those who do not.
This will also widen social inequality, because the most productive machines will become exponentially more productive.
For every investor, this is the time to position yourself in a few select sectors and assets.
These will be the primary winners and the ones that will place you in the top 0.1% of society.
Do not see a new open-source AI as a competitor—see it as an opportunity.
See you soon.
Now, our asymmetric portfolio in detail
