Breaking

CoreWeave Reports First Live Benchmarks of NVIDIA Vera Rubin NVL72, Citing 10x Gain in Tokens per Megawatt

July 21, 2026 at 12:16 EDT

Infra & Chips
Foundation Models

CoreWeave has published what it describes as the first measured performance figures from NVIDIA's next-generation Vera Rubin NVL72 platform, reporting a tenfold improvement in tokens per second per megawatt over the current Blackwell generation when running the DeepSeek-R1 model. The company stressed that the numbers come from live silicon rather than projections, and NVIDIA amplified the claim through its own channels.

Vera Rubin NVL72 · First Measured Results

10× More AI Tokens Per Megawatt on First Live Silicon

CoreWeave reports the first measured figures from NVIDIA's next-generation platform: a tenfold gain in tokens per second per megawatt over Blackwell when running DeepSeek-R1 — a metric that matters now that power, not chip count, caps AI capacity.

1×

Blackwell
GB200 NVL72

10×

Vera Rubin
NVL72

Tokens per second, per megawatt

~1/10

cost per million tokens for highly interactive agentic AI

1/4

the GPU count needed to train Mixture-of-Experts models

up to 35×

throughput per MW for trillion-parameter models with dedicated inference accelerators

Inside the NVL72 rack

Rubin GPUs

Vera CPUs

6th-gen

NVLink switching

liquid

cooled, cable-free trays

Why it lands

Figures come from live silicon, not projections. With energy availability — not chip count — now governing how much AI capacity can come online, tokens-per-megawatt is the metric that matters.

The caveats

Independent at-scale evaluations remain limited. The vendor selected the model and configuration; pricing and deployment scale are undisclosed, with the platform still ramping into production through 2026.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.