TAIPEI — You’ve gotta love Jensen Huang. The man walks on stage in his trademark leather jacket, hands us the keys to Blackwell—a tech marvel we’ve barely even figured out how to cool yet—and then casually drops, “Oh by the way, here’s what’s replacing it.”
Enter Rubin.
Before the ink is even dry on the Blackwell purchase orders, NVIDIA is already ushering its current flagship toward the exit. It’s a level of tech-industry flex that borders on disrespectful, but looking at the specs for the next-generation Rubin architecture, it’s hard to blame them for bragging. If Blackwell was a massive leap forward, Rubin is looking like a straight-up teleportation device.
How Much Better Are We Talking?
To understand how absurdly powerful Rubin is, you have to look at what’s under the hood. While Blackwell is just starting to stretch its legs in data centers globally, Rubin is already flexing numbers that make the current generation look like a collection of pocket calculators.
Here is how the next-gen titan stacks up against the brand-new Blackwell:
| Feature | Blackwell Architecture | Next-Gen Rubin Architecture |
| Memory Tech | 8-Hi HBM3E | Next-Gen HBM4 (8-Hi & 12-Hi) |
| Process Node | TSMC 4NP (Custom 4nm) | TSMC 3nm (Enhanced) |
| Form Factor | Ultra-dense Dual-Die | Superchip (Rubin Ultra + Vera CPU) |
| Bandwidth | Mind-boggling | Literally breaking the physics bottleneck |
The real kicker here is the jump to HBM4 memory. Blackwell uses HBM3E, which is blisteringly fast, but Rubin is moving to a native 3nm process and packing true next-generation high-bandwidth memory.
We are talking about a generational leap that doesn’t just tweak performance metrics by 10% or 20%—it completely obliterates the data bottlenecks that currently keep AI engineers up at night.
The Rubin Ultra and the “Vera” Tag Team
NVIDIA isn’t just dropping a new GPU; they are rewriting the entire system architecture. Rubin will eventually evolve into the Rubin Ultra, paired with a brand-new centralized processor dubbed the Vera CPU.
The Rumor Mill: Word on the street is that the Rubin platform will leverage advanced packaging that makes current chip-on-wafer tech look downright vintage. It’s designed to handle advanced AI training and real-time reasoning at a fraction of the power-per-token cost of Blackwell.
So, if you are a tech executive who just spent hundreds of millions of dollars securing your allocation of Blackwell chips, you might want to take a deep breath. Yes, your new hardware is still incredible. But Jensen just put an expiration date on it before you even got it out of the box.
Welcome to the relentless pace of the AI arms race. Grab some popcorn.
This article is from a submission, please keep the link for forwarding: https://www.36ti.com/nvidia/nvidia-just-teased-the-rubin-architecture-and-honestly-blackwell-isnt-even-warm-yet-21084710
