AI Term of the Moment

Leonardo AI


Look Up Another Term


Definition: Vera Rubin


NVIDIA's flagship AI hardware as of 2026. With 10x greater performance, Vera Rubin supersedes Blackwell. However, Blackwell products are expected to be available through 2027. The name comes from American astronomer Vera Florence Cooper Rubin, who pioneered work on the rotation rates of galaxies. Vera is the CPU and Rubin is the GPU. See Blackwell.

Often sold through system integrators, Vera CPUs can be purchased separately; however, Rubin GPUs are sold in various configurations. For example, the Vera Rubin Superchip includes one Vera CPU and two Rubin GPUs. HGX modules have eight GPUs and no CPUs, while the NVL72 rack (see below) includes 36 Vera CPUs, 72 Rubin GPUs and 20.7 terabytes of HBM4 memory and liquid cooling. The NVL144 doubles the number of chips. NVL is NVIDIA's high-speed interface between chips (see NVLink).




Vera Rubin NVL72 Rack
Comprising more than a million different components and selling for roughly $8 million, a small data center may have 10 to 100 of these racks while a major AI factory may use thousands. At $8 million per rack, it is no wonder NVIDIA became the most valuable company in the world in late 2026. Because they are self-contained, Vera Rubin racks can be installed in a data center in less than an hour. (Image courtesy of NVIDIA Corporation.)






Two GPUs for Disaggregated Inference
When a user asks a question, the inference software processes the entire prompt first, known as the "prefill stage," and CPX GPUs are optimized for very large inputs. The "decode" stage generates the answer tokens, and fast HBM memory connected to regular Rubin GPUs speeds up the process. See inference. (Image courtesy of NVIDIA.)





A
B
C
D
E
F
G
H
I
J
K
L
M
N
O
P
Q
R
S
T
U
V
W
X
Y
Z
misc