Skip to content
AI.info

The Pulse

CoreWeave cites study showing up to 65% lower GB200 NVL72 cloud costs

A Signal65 analysis modeled a three-year, 5,000-GPU deployment against three unnamed hyperscaler clouds. CoreWeave says the study also found a lower cost per token after accounting for inference throughput.

CoreWeave cites study showing up to 65% lower GB200 NVL72 cloud costs

CoreWeave says a modeled 5,000-GPU deployment of Nvidia GB200 NVL72 systems cost up to 65% less over three years on its cloud than on three unnamed hyperscaler clouds. The company highlighted the findings from an analysis by Signal65 performance analyst Mitch Lewis in a Nov. 16 post.

Signal65's analysis used public on-demand prices, MLPerf Inference v6.0 throughput results and published serverless token rates. Across seven configurations, it found CoreWeave had the lowest total cost of ownership in six.

What the cloud comparison measured

For Nvidia GB200 NVL72 systems, the highest-cost comparison was an unnamed provider called Hyperscaler C, whose projected three-year bill was $3.9 billion, according to CoreWeave. An identical deployment on CoreWeave was modeled at 65% less.

CoreWeave's modeled cost was also 36% lower than Hyperscaler A's and 22% lower than Hyperscaler B's. The comparison with Hyperscaler B used an estimated rate because that provider does not publish GB200 NVL72 pricing.

The model excludes per-operation storage charges and early-deletion penalties on archive tiers, CoreWeave says. The company says it charges neither fee and that the cost gap in practice is likely wider than the modeled figures.

The study also found cost differences for other Blackwell systems. Signal65 calculated up to 41% lower cost on Nvidia HGX B200 and up to 52% lower cost on Nvidia HGX B300 over three years.

Throughput and token costs

To calculate inference costs, the analysis combined cluster costs with MLPerf Inference v6.0 results for DeepSeek-R1, a 671-billion-parameter reasoning model. CoreWeave says its system delivered 4,867 tokens per second per GPU running the model on GB200 NVL72.

The resulting modeled cost was $0.75 per million tokens on CoreWeave, compared with $1.23 to $2.22 across the hyperscalers, according to the company's post. Against the most competitively priced hyperscaler, Signal65 calculated that the cost advantage grew from 22% based on infrastructure pricing alone to 65% after performance was included.

The analysis translates those costs into an estimate of the tokens a fixed annual commitment could buy. CoreWeave says $1 million would purchase about 1.33 trillion tokens on its cloud, compared with 451 billion to 814 billion on the other providers.

Hosted model pricing

The study separately compared serverless inference, in which customers access hosted models through an API rather than manage infrastructure. For gpt-oss-120b, the one model available on every cloud in the study, CoreWeave says its input tokens cost 67% to 80% less and its output tokens 53% to 72% less.

Across 17 models with a direct hyperscaler comparison, CoreWeave's serverless service was lower-priced on 11 and identically priced on four, Signal65 found. Median input token costs were 26% lower and median output token costs were 15% lower across those comparisons.

CoreWeave also points to its bare-metal servers, networking and storage as parts of the infrastructure behind the results. Its Local Object Transport Accelerator runs as a caching proxy on GPU nodes in a CoreWeave Kubernetes Service cluster, and the company says the study found its storage costs were 28% to 55% lower than those of the hyperscalers.

Sources

  1. CoreWeave
  2. Signal65

Companies in this story

Related stories

Explore

More articles