AI

CoreWeave pricing and the Nvidia supply squeeze

CoreWeave's 8x H100 node costs $49.24 per hour on demand, though B200 nodes offer better value per training result. Nvidia provides a $6.3 billion capacity backstop to help manage CoreWeave's $51.61 billion in debt and lease obligations.

CoreWeave pricing and the Nvidia supply squeeze

CoreWeave’s 8x H100 node is $49.24 per hour on demand. This price covers the full instance rather than a single chip. If you provision an 8x H100 node but only use three GPUs, your effective rate climbs because the bill remains the same. A half-utilized H100 node is more expensive per useful hour than a fully utilized B200. You should plan your budget using node costs instead of per-GPU quotes to avoid surprises. Egress is free at CoreWeave, which removes the lock-in tax found at AWS or Azure. A 100 TB hot dataset is $6,144 monthly under binary GB accounting. CPU support nodes for training pipelines are $3.36 to $9.31 per hour. Spot rates are up to 60 percent below on-demand, while reserved contracts are up to 60 percent cheaper. In North America, an H100 node is $19.71 for spot, while the European spot rate is $19.51. The 8x L40 node is $10.00, and the 8x B200 node is $68.80. Because the B200 is 2.5x the H100 in training performance, it is often cheaper per result despite the higher sticker price.

The Nvidia relationship is a safety net. Nvidia agreed to buy $6.3 billion of unused capacity from CoreWeave through April 2032, per an SEC 8-K filing. This backstop is a way to manage the $51.61 billion debt and lease obligations CoreWeave reported as of June 30, 2026. While H100 rental prices fell in 2025, rental rates for H100s spiked 22% in the month leading to September 7, 2026. While the market saw price drops earlier, the current supply freeze for older silicon drives these costs back up. This volatility is present alongside a $104.2 billion revenue backlog. Large clients like Meta committed $21 billion for AI cloud capacity through 2032. OpenAI also reached a cumulative $22.4 billion in contracts as of September 2025. Investors weigh these contracts against a $4.2 billion principal repayment due later this year and a $55.6 billion revenue backlog. In January 2026, Nvidia injected an additional $2 billion into CoreWeave at a price of $87.20 per share. The company is a user of Delayed Draw Term Loans to finance GPU purchases, such as a $2.0 billion draw at S+4.00%.

I recommend CoreWeave for large-scale training because its Kubernetes-native orchestration handles massive distributed workloads better than simpler platforms. Lambda is a better choice for researchers that want to avoid Kubernetes complexity or need lower H100 prices. Lambda’s H100 instances are $2.99 to $3.29 per hour, whereas CoreWeave’s H100 rates are $4.25 to $6.16, a difference that compounds at scale. The H200 is 2.4 percent more expensive than the H100. CoreWeave also provides Flex Reservations, which are a capacity ceiling with a lower 24/7 holding fee. The Zero Trust AI Framework is a set of six layers of customer-controlled security, including Zero Key Access and Zero Node Access. Lambda’s 1-Click Clusters are 16 GPUs or more, while CoreWeave is a provider of whole 8-GPU servers.

GPU Type On-Demand Node Price Single GPU Inference
8x H100 $49.24 $6.16
8x B200 $68.80 $8.60
8x L40 $10.00 N/A
GB200 NVL72 $42.00 N/A

CoreWeave targets $30 to $35 billion in capital expenditures for 2026. You should notice that CoreWeave’s revenue is from long-term, take-or-pay contracts where clients pay for capacity whether they use it or not. This helps manage the $51.61 billion debt load. Will the massive debt load eventually force a consolidation of these specialized clouds?