CoreWeave Bills Inactive AI Data at Lower Rates, Claims Up to 75% Savings

Three billing levels based on access

According to CoreWeave, its AI Object Storage uses real-time access tracking to adjust billing across three pricing levels: Hot at $0.06/GB/mo, Warm at $0.03/GB/mo and Cold at $0.015/GB/mo.

Data moves from Hot to Warm after seven days of inactivity, and from Warm to Cold after more than 30 days. The company says this happens without manual management or rehydration delay.

CoreWeave stresses that the system does not archive or move data. In its description, the service works out which data is inactive and simply does not charge a premium for it. Every object remains accessible at full line-rate performance, with no retrieval, egress or request fees, the company states.

The problem: AI artifacts pile up in hot storage

CoreWeave argues that each fine-tune produces gigabytes of "just-in-case" artifacts such as checkpoints, model weights and datasets. Teams tend to keep them in hot storage to avoid rehydration delays, which drives costs up, clogs containers and slows directory listings.

The company cites an industry trend suggesting around 80% of data is inactive at any given time, while noting that no public benchmarks quantify this at scale. It also shares a table of three anonymized customer scenarios:

• Customer A: 65 PB total, 5 PB active (7%), 60 PB inactive (93%)
• Customer B: 8 PB total, 2 PB active (25%), 6 PB inactive (75%)
• Customer C: 17.5 PB total, 3.5 PB active (20%), 14 PB inactive (80%)

CoreWeave says deleting old checkpoints means losing data for future training and creates problems such as lost audit trails, inconsistent retention and extra on-call load from maintaining deletion scripts.

How CoreWeave compares itself with hyperscalers

The company says traditional cloud tiering was designed for general-purpose object storage rather than iterative machine learning. It lists three shortcomings of intelligent tiering: it is not fully automatic and needs opt-in tagging and bucket policies, it adds management, request, retrieval and egress fees, and it can make read times inconsistent when checkpoints are loaded from an inactive tier.

To illustrate the cost gap, CoreWeave provides a comparison based on its own assumptions: 20 PB of data, a 1 MB average file size, 20% of data uploaded monthly, 30% read monthly and 30% egressed out of the cloud monthly. The totals it reports are:

• CoreWeave: $544,526, or $0.026 per GB per month
• AWS S3 Standard: $779,027, or $0.037
• AWS S3 Express: $2,643,683, or $0.126
• GCP GCS Standard: $1,009,713, or $0.048
• Azure Blob Hot LRS: $786,113, or $0.037

In this scenario, the competitor totals include egress charges of $314,573 for AWS and Azure and $503,316 for GCP, while CoreWeave lists no egress, request, retrieval or tiering fees. These figures come from the company's own analysis.

What it means for AI teams

For teams running large training and fine-tuning pipelines, the pitch is that they can keep full checkpoint histories for audit and rollback without paying hot rates for data that is rarely read. CoreWeave also lists consistent performance across active and inactive data, no manual cleanup scripts and simpler governance as benefits.

The savings claims rest on CoreWeave's own assumptions and customer examples, so actual results will depend on how much of a team's data is inactive and how often it is accessed. The company has also published an ebook, Beyond the Hot Tier: Cut AI Storage Costs, Accelerate Breakthroughs, for a deeper look.