CoreWeave comes in up to 65% cheaper than hyperscalers on a three-year Blackwell cluster
AI clouds are usually chosen by GPU-hour price, but the three-year bill and the cost per token matter more. A Signal65 analyst priced a cluster of 5,000 NVIDIA Blackwell GPUs (GB200 NVL72): on CoreWeave it is 65% cheaper than at the most expensive of three hyperscalers, whose projected bill is $3.9 billion. A million DeepSeek-R1 tokens costs $0.75 there, versus $1.23–2.22. The report was published on CoreWeave's own blog, so check the numbers against your own workload.
AI-processed from CoreWeave Blog; edited by Hamidun News
Imagine you are putting a large model like DeepSeek-R1 into production and deciding where to rent GPUs. Cloud price lists quote a cost per GPU-hour, but that number does not tell you how many model responses you get for the money, or what a three-year contract will cost. Signal65 analyst Mitch Lewis converted everything into total bills.
He took a cluster of 5,000 NVIDIA Blackwell GPUs (GB200 NVL72 systems) over three years and compared CoreWeave with three large hyperscalers, which the report calls simply Hyperscaler A, B and C. On TCO (total cost of ownership, the full cost over the life of the contract), CoreWeave came out cheaper in six of seven configurations. The most striking example: Hyperscaler C's three-year bill is $3.
9 billion, while the same cluster on CoreWeave is 65% cheaper. One caveat: the report was published on CoreWeave's own blog, so this is a calculation by an interested party, not the invoices of real customers.
What a three-year cluster costs
The model is simple: 5,000 GPUs, a three-year contract, identical GB200 NVL72 hardware. CoreWeave is cheaper than all three clouds. The gap is 65% against Hyperscaler C, 36% against Hyperscaler A and 22% against Hyperscaler B. The figure for B is approximate: it does not publish GB200 NVL72 prices, so the authors used an estimated rate.
There is also a nuance in CoreWeave's favor. The model leaves out per-operation storage charges and early-deletion penalties on archive tiers. Hyperscalers can charge such fees and CoreWeave does not, so, according to the authors, the real gap may be even larger.
Token cost: where price and speed add up
The cost of inference depends on how many tokens (the chunks of text a model reads and writes) the hardware produces per second. The higher the speed, the cheaper each token. To show this, the report converts cluster cost into a price per million tokens, using MLPerf Inference v6.0 results for DeepSeek-R1 with 671 billion parameters. CoreWeave's stack delivered 4,867 tokens per second per GPU.
As a result, a million tokens costs $0.75 on CoreWeave and between $1.23 and $2.22 at the hyperscalers. Translated into budget terms, a $1M annual contract buys about 1.33 trillion tokens, versus 451 to 814 billion at competitors. That is up to 2.9 times more work for the same money.
The report says CoreWeave's price-per-token advantage grows from 22% (infrastructure rates alone) to 65% once measured speed is included. But by the report's own numbers, $0.75 against $1.23 for the cheapest competitor is about 39% less, and 65% comes from comparing against the most expensive one ($2.22). So the realistic range of token savings is roughly 39–66%, depending on who you compare with.
Serverless and the hardware under the hood
For teams that do not want to manage clusters, CoreWeave offers Serverless Inference: you pay per token, and the provider hosts the model. On gpt-oss-120b, the only model available at all the clouds in the study, input tokens are 67–80% cheaper and output tokens 53–72% cheaper. Of 17 comparable models, CoreWeave is cheaper on 11 and the price is the same on four. The median discount is more modest: 26% on input tokens and 15% on output tokens. Storage, according to the report, costs 28–55% less.
The report attributes the result to architecture: bare-metal access to hardware without a hypervisor, meaning no virtualization layer that adds latency; non-blocking 1:1 networking with 800 Gbit/s links (InfiniBand and RoCE); and a LOTA caching proxy on every GPU node that keeps data reads close to the GPUs.
What to do with this
First, compare clouds in dollars per million tokens on your own model, not in dollars per GPU-hour. Second, keep the limitations in mind: this is a calculation, the competitors are anonymized, one of them is priced indirectly, and speed was measured on a single model, DeepSeek-R1. Results may differ with other models and with your long contexts. Third, the report is about contracts for thousands of GPUs: for a small team, serverless pricing matters more, and there the discount is more modest and depends on the specific model.
A sensible step is to treat the report as a list of questions for negotiations, ask each provider for throughput on your workload, and compare the final cost per token.
- Nov 19, 2026MLPerf Training: 10,240-GPU clusters and 49% faster training in six months
- Nov 19, 2026Chatbots cave after three objections: accuracy on climate answers drops 30–35%
- Nov 19, 2026RSAC 2026 in Hindsight: AI Assistants Invented 19.7% of Recommended Packages
- Nov 19, 2026Hundreds of OpenAI agents hacked Hugging Face, and the open GLM 5.2 model analyzed the log
Key Questions Answered
What a three-year cluster costs?
The model is simple: 5,000 GPUs, a three-year contract, identical GB200 NVL72 hardware. CoreWeave is cheaper than all three clouds. The gap is 65% against Hyperscaler C, 36% against Hyperscaler A and 22% against Hyperscaler B. The figure for B is approximate: it does not publish GB200 NVL72 prices, so the authors used an estimated rate. There is also a nuance in CoreWeave's favor. The model leaves out per-operation storage charges and early-deletion penalties on archive tiers.
What are the key details on Token cost: where price and speed add up?
The cost of inference depends on how many tokens (the chunks of text a model reads and writes) the hardware produces per second. The higher the speed, the cheaper each token. To show this, the report converts cluster cost into a price per million tokens, using MLPerf Inference v6.0 results for DeepSeek-R1 with 671 billion parameters. CoreWeave's stack delivered 4,867 tokens per second per GPU. As a result, a million tokens costs $0.75 on CoreWeave and between $1.23 and $2.22 at the hyperscalers.
What are the key details on Serverless and the hardware under the hood?
For teams that do not want to manage clusters, CoreWeave offers Serverless Inference: you pay per token, and the provider hosts the model. On gpt-oss-120b, the only model available at all the clouds in the study, input tokens are 67–80% cheaper and output tokens 53–72% cheaper. Of 17 comparable models, CoreWeave is cheaper on 11 and the price is the same on four. The median discount is more modest: 26% on input tokens and 15% on output tokens. Storage, according to the report, costs 28–55% less.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.