Artificial intelligence news.
Straight from the original announcements.

CoreWeave Claims Lower Blackwell AI Costs Than Hyperscalers

November 18, 2026, 14:00By the AI News Bites Editorial TeamSource: CoreWeave
CoreWeave Claims Lower Blackwell AI Costs Than Hyperscalers
Image: CoreWeave

CoreWeave says a Signal65 study found its cloud delivered the highest per-GPU inference throughput and the lowest total cost of ownership in six of seven NVIDIA Blackwell configurations. The analysis modeled a 5,000-GPU deployment over three years and compared CoreWeave with three traditional hyperscaler clouds.

Key points
  • Signal65 modeled a 5,000-GPU NVIDIA Blackwell deployment over three years.
  • CoreWeave had the lowest TCO in six of seven configurations.
  • Its NVIDIA GB200 NVL72 deployment was modeled as 65% cheaper than Hyperscaler C.
  • CoreWeave reported 4,867 DeepSeek-R1 tokens per second per GPU.
  • Serverless Inference was cheaper on 11 of 17 directly comparable models.

CoreWeave reports a major three-year cost gap

The study, titled CoreWeave Delivers Higher Throughput and Lower Cost per Token on NVIDIA Blackwell, examines cloud economics for enterprise AI infrastructure. It was written by Mitch Lewis, a Performance Analyst at Signal65, and models the cost of operating a 5,000-GPU NVIDIA Blackwell deployment over three years.

According to CoreWeave, the provider had the lowest total cost of ownership in six of seven configurations. In the NVIDIA GB200 NVL72 comparison, CoreWeave says its projected cost was lower than all three evaluated hyperscalers. Hyperscaler C's projected three-year cost reached $3.9 billion, while an identical deployment on CoreWeave was 65% less expensive.

The analysis also found CoreWeave was 36% less expensive than Hyperscaler A and 22% less expensive than Hyperscaler B. CoreWeave says the Hyperscaler B comparison uses an estimated rate because that provider does not publish GB200 NVL72 pricing.

Higher throughput lowered the modeled cost per token

The report links infrastructure economics to inference throughput, using MLPerf® Inference v6.0 results for the 671-billion-parameter DeepSeek-R1 reasoning model. CoreWeave says its stack delivered 4,867 tokens per second per GPU on NVIDIA GB200 NVL72.

When throughput is included, CoreWeave's cost per million tokens was $0.75, compared with $1.23 to $2.22 on the hyperscalers. The company says its advantage over the most competitively priced hyperscaler increased from 22% based on infrastructure pricing alone to 65% after throughput was factored in.

CoreWeave also says a $1 million annual commitment would buy about 1.33 trillion tokens on its platform, compared with 451 billion to 814 billion elsewhere. That represents up to 2.9x the tokens for the same spend, according to the company.

Signal65's model excludes some fees that CoreWeave says are common on hyperscaler clouds, including per-operation storage charges and early-deletion penalties on archive tiers. CoreWeave says it charges neither, meaning the real TCO difference could be wider than the modeled results.

Serverless pricing and infrastructure design are part of the comparison

The study also evaluates CoreWeave Serverless Inference for teams that want hosted open-source models without managing clusters. For gpt-oss-120b, the only model available on every cloud in the study, CoreWeave says its input tokens cost 67% to 80% less and output tokens cost 53% to 72% less. Across 17 directly comparable models, it was lower-priced on 11 and identically priced on four, with median input token costs 26% lower and median output token costs 15% lower.

CoreWeave attributes the results to a platform built around bare metal servers, high-bandwidth networking and local storage. Its network uses NVIDIA Quantum InfiniBand and RDMA over Converged Ethernet fabrics with 800 Gbps of full-duplex bandwidth. The company says its Local Object Transport Accelerator keeps storage reads local to GPU nodes, while storage costs in the study were 28% to 55% below those of the hyperscalers.

For AI teams, the findings support CoreWeave's argument that hourly infrastructure rates alone do not capture the cost of running inference. The study instead combines deployment costs, throughput, storage and token pricing, giving enterprises a basis for comparing both managed infrastructure and pay-per-token access.

Related news

NVIDIA and Microsoft Extend AI Infrastructure From Azure to Windows PCsTop story
Chips and infrastructureNovember 19, 2026, 21:00

NVIDIA and Microsoft Extend AI Infrastructure From Azure to Windows PCs

NVIDIA and Microsoft are expanding their accelerated computing partnership across Azure, industrial systems and Windows PCs, adding access to Vera Rubin infrastructure…

CoreWeave reports first multi-rack NVIDIA Vera Rubin MLPerf results
Chips and infrastructureNovember 16, 2026, 17:15

CoreWeave reports first multi-rack NVIDIA Vera Rubin MLPerf results

CoreWeave says it is the first cloud provider to deliver multi-rack NVIDIA Vera Rubin results in MLPerf Training v6.1, including a 17.65-minute run for DeepSeek-V3 671B…

CoreWeave posts 441,740 tokens per second on DeepSeek-R1 in MLPerf Endpoints
Chips and infrastructureOctober 22, 2026, 18:19

CoreWeave posts 441,740 tokens per second on DeepSeek-R1 in MLPerf Endpoints

CoreWeave says it delivered leading results in the inaugural MLPerf 0.7 Endpoints benchmark, sustaining 441,740 output tokens per second with the DeepSeek-R1 reasoning…