THE AI HEIST OF THE CENTURY : The Year Your PC Became a Luxury (Part II)

This article follows the first part : THE AI HEIST OF THE CENTURY : The Year Your PC Became a Luxury (Part I).

WHAT IS ALL THAT RAM ACTUALLY FOR ?

In the field of video generation, this power is used for :

• Temporal Consistency : This is what prevents the cat from turning into a dog between two frames. The AI must keep every previous frame in RAM to remain consistent.
• Physical Simulation : Calculating mud splashing under the tires of rally cars requires a colossal buffer memory to simulate fluids.
• Deepfakes and Alteration : This is the dark side. Generating ultra-realistic human bodies is what requires the most "layers" of calculation.

But this power is not just used to create cat videos or deepfakes. If RAM (and especially VRAM/HBM) has become the rarest fuel on the planet, it is because it is at the heart of much more "invisible" but crucial tasks.

For example, AIs need massive resources to continue learning and developing. We had direct proof of this on January 15, 2026 : for its 25th anniversary, Wikipedia formalized its role as the primary "food" for industry giants.

AI is literally in the process of assimilating Wikipedia. The foundation announced that Meta, Mistral AI, Amazon, and Perplexity were officially joining its Wikimedia Enterprise program. Why ? Because to train models capable of answering everything, they must ingest the 65 million articles on the site. As a result, these firms use real-time APIs so that their models "drink" human knowledge as it is written by volunteers.

But to store, classify, and process these 15 billion monthly views without crashing, a massive infrastructure is required. A "win-win" market has therefore been concluded, as Wikipedia sees its audience eroding : indeed, more and more internet users are using conversational AIs as search engines, thus bypassing the encyclopedia. The Wikimedia Foundation estimates it has already lost 8 % of its "human traffic" in one year. By monetizing access to its data to train these same AIs, Wikipedia is no longer suffering from competition ; it is financing it. This is a strategic agreement that transforms an existential threat into a vital resource to ensure the sustainability of free knowledge.

But Where is All That Silicon Going ?

IN THE "RAM" OF CHATBOTS (KV CACHE)

When you chat with ChatGPT or Grok, the AI must remember everything you said ten sentences ago.

How does it work ? To avoid having to re-read your entire text for every new word it generates, the AI stores the data from your conversation in a "KV Cache".

IN REAL-TIME INFERENCE (THE OPEN BRAIN)

For an AI to answer you instantly, the "model" (the billions of neural connections) must be entirely loaded into memory.

The problem is that a model like Llama 3 (405B) weighs hundreds of gigabytes. It cannot be stored on a classic hard drive to respond, as it would be too slow.

IN WORLD SIMULATION (WEATHER, MEDICINE, FINANCE)

AI is also used to predict the future, and that is a RAM pit :

• Weather : Simulating the movement of every air particle to predict a storm requires a "buffer memory" capable of handling billions of variables at once.
• Medicine : Protein folding (to create new drugs) requires testing millions of spatial combinations. Each test occupies space in RAM to be compared to the others.

IN THE "BANDWIDTH"

It is not just quantity that matters, but speed.

The HBM3e memory that Nvidia and SK Hynix are fighting over (with its 11 billion dollar investment) is capable of transferring terabytes per second.

CAMPUS NOTE

It is a technological vertigo. We now understand that our PCs are not fighting against software, but against reality simulators that never sleep. Every Giga of RAM is a brick of this new world that the giants are building without us.

And if you have any questions or remarks on AI and RAM, feel free to talk about it in the comments right below !