NVIDIA Vera Rubin NVL72 on CoreWeave Cloud: Built for Agentic AI
NVIDIA Vera Rubin NVL72 packs 72 Rubin GPUs and 36 Vera CPUs into a single rack, built for pretraining, post-training, test-time scaling, and agentic scaling. Versus Blackwell: four times faster training, one-fourth the GPUs, 10 times higher inference throughput per watt, one-tenth the cost per million tokens. 100 radiology scans an hour on Blackwell becomes 1,000 on Vera Rubin, same power draw.
CoreWeave optimized every layer of the stack to unlock it, from CoreWeave Mission Control's liquid cooling and rack management to SUNK's topology-aware scheduling and CoreWeave ARENA's cost transparency and expert access.
1
00:00:02,500 --> 00:00:03,133
Meet the most
2
00:00:03,133 --> 00:00:06,133
advanced rack scale AI system ever built.
3
00:00:06,633 --> 00:00:09,633
Nvidia Vera Rubin NVL72.
4
00:00:09,900 --> 00:00:14,266
With 72 Rubin GPUs and 36 Vera CPUs per rack.
5
00:00:14,633 --> 00:00:18,000
Vera Rubin is designed for pretraining, post-training,
6
00:00:18,000 --> 00:00:20,400
test-time scaling and agentic scaling.
7
00:00:20,933 --> 00:00:22,733
Compared to Nvidia Blackwell,
8
00:00:22,733 --> 00:00:26,666
Vera Rubin provides AI training with one-fourth the GPUs,
9
00:00:27,300 --> 00:00:30,000
ten times higher inference throughput per megawatt
10
00:00:30,000 --> 00:00:35,633
and one-tenth the cost per million tokens for highly interactive deep reasoning agentic AI.
11
00:00:35,633 --> 00:00:38,600
To unlock the full potential of the Vera Rubin platform,
12
00:00:38,933 --> 00:00:42,266
CoreWeave optimized every layer of our stack.
13
00:00:42,866 --> 00:00:46,433
With our software defined liquid cooling system, Valvey
14
00:00:46,466 --> 00:00:49,100
and unified rack manager, Racky
15
00:00:49,100 --> 00:00:53,733
CoreWeave sustains peak system performance at lower TCO.
16
00:00:53,733 --> 00:00:57,200
CoreWeave mission control automates the entire deployment pipeline
17
00:00:57,266 --> 00:00:59,700
from install to discovery.
18
00:00:59,700 --> 00:01:02,566
SUNK plugins provide a modular framework for
19
00:01:02,566 --> 00:01:06,900
topology aware scheduling, so workloads run at optimal performance.
20
00:01:06,900 --> 00:01:11,833
And CoreWeave Arena gives the signals, cost transparency and expert access
21
00:01:11,833 --> 00:01:13,066
you need.
22
00:01:13,066 --> 00:01:15,633
Together, these innovations help you deliver
23
00:01:15,633 --> 00:01:20,366
and accelerate everything from research to business outcomes.
24
00:01:20,900 --> 00:01:23,866
CoreWeave was one of the first clouds to bring rack scale systems
25
00:01:23,866 --> 00:01:29,500
like the Nvidia GB200 and GB300 NVL72 to production at scale,
26
00:01:29,633 --> 00:01:32,733
and we’re one of the first clouds to bring up and validate Vera Rubin,
27
00:01:33,000 --> 00:01:36,400
empowering AI pioneers with a full stack.
28
00:01:36,566 --> 00:01:41,066
A system this powerful deserves the world's number one AI cloud.
29
00:01:41,066 --> 00:01:42,566
We’re built for this.
30
00:01:42,566 --> 00:01:43,533
CoreWeave
31
00:01:43,533 --> 00:01:45,966
The essential cloud for AI.
