AI Video : The Spectacle Devouring Our PCs
We told you about it some time ago : RAM has become the new black gold. But while you thought it was "just" a story of servers or databases, the reality is much more visual... and much more greedy.
Today, new video generation models like Veo 3 (Google), Sora (OpenAI), or Kling no longer just "calculate" ; they are partially devouring global RAM production to bring your digital fantasies to life.
Why is AI "Stealing" Our Memory Sticks ?
To generate a video like those cats racing rally cars or that AI-generated music clip you see below, the machine must anticipate every pixel, every reflection, and every physical movement in real-time.
Unlike a classic video game where a portion of the assets is pre-calculated, here, everything is created ex nihilo. The result : while your Warcraft 3 ran on 512 MB of RAM back in the day, a single training instance for Veo 3 requires terabytes of HBM3e (ultra-high performance memory).
Among other things, this is where your future RAM sticks went...
Who is Buying What ? Numbers That Make Your Head Spin
If you can no longer find Crucial sticks at a decent price, it's because tech giants have pulled out checkbooks of an indecent size. Here are the ones raking it all in and emptying our shelves (Sources 2025-2026) :
Microsoft & Nvidia : The Alliance at the Peak
For these two juggernauts, the absolute priority has shifted : the home PC has become secondary compared to next-generation memory. By signing infrastructure contracts worth tens of billions of dollars, they ensure that factories produce almost exclusively for their data centers. Every H100 or Blackwell B200 chip they deploy is surrounded by dozens of gigabytes of ultra-high-performance HBM memory... a technology we will likely never see in our home machines.
xAI (Elon Musk) : The Excess of "Colossus"
Elon Musk is not doing things by halves for his Grok-3 AI. His firm xAI commissioned a monster named "Colossus" in Memphis in late 2025. This supercomputer runs on 100,000 Nvidia H100 chips. The market impact was immediate : each of these chips consumes an astronomical amount of memory that manufacturers no longer produce for us. Worse, Musk is already planning to double this capacity to reach 200,000 chips during 2026. That's a lot of silicon that will never end up in your future DDR5 stick.
Meta (Mark Zuckerberg) : The Extreme Stockpiler
Mark Zuckerberg was one of the first to set the tone : Meta wants to own the largest computing power in the world for its Llama models. Today, Meta already owns over 600,000 enterprise-class GPUs. With every server update, they siphon off Samsung and SK Hynix stocks before distributors like Amazon even have time to place an order.
Samsung, SK Hynix & Micron : The Curtain is Drawn
The verdict from manufacturers is in, and it is chilling : the two Korean giants have officially declared that their HBM memory stock for the entire year of 2026 is already SOLD OUT. Everything has been pre-sold to the "AI ogres" (Meta, Google, Amazon). For its part, Micron has shifted its production lines toward these lucrative contracts, progressively abandoning the general public. The industrial calculation is cold but effective : why bother selling a 16 GB stick to a gamer for $100 when you can sell the same component ten times more expensive to OpenAI or Mistral AI ?
Campus Note : Our gamer hearts ache seeing our trusted brands turn their backs on us for the billion-dollar arms of Silicon Valley giants. The rift is complete, and the PC player is no longer the industry's priority.
The Investigation Continues
In the next article : But concretely, why are these machines so hungry ? What is it in the code of these AIs that requires so many gigabytes that we can't even find them in stores anymore ? We dive into the engine of the ogre to understand this technological "Black Hole."
Read next : Why AI devours so much RAM
And if you have any thoughts on this RAM war, feel free to talk about it in the comments right below !
