Cloud object storage pricing has a history. It was designed for a specific category of customer — one who needs to store large volumes of data, access it rarely, and pay a predictable monthly rate tied to how many gigabytes sit in the bucket. Compliance archives. Cold backup. Regulatory retention. Data you hope never to read again but have to keep for seven years.
That model made sense for the workloads it was built around. The problem is that AI training isn’t one of them.
Explore how Unity NV-Series™ eliminates per-access billing for AI workloads →
Why does cloud storage cost more for AI training workloads?**
Cloud storage bills in five dimensions: storage capacity, API operations, data retrieval, data transfer out (egress), and replication overhead. AI training activates all five simultaneously — each epoch reads the full dataset, generating API calls, retrieval events, and egress charges in parallel. Real-world spend typically runs 2–5× the raw storage rate.
Cloud storage billing charges in five dimensions, not one
Most cost estimates for cloud storage anchor on a single number: the per-GB storage rate. AWS S3 Standard is approximately $0.023/GB per month. Azure Blob Hot is roughly $0.018/GB. Those numbers are easy to find and easy to plan around — which is why they dominate most initial cost models and almost no AI infrastructure budgets.
But cloud providers don’t bill on a single dimension. In practice, there are five:
- Storage capacity — the GB-stored rate, the number everyone quotes
- API operations — every GET, PUT, LIST, and COPY request against your data generates a charge; at scale, these aren’t small
- Data retrieval — some tiers charge specifically for bytes read back from the archive layer, before any byte leaves the cloud
- Data transfer out (egress) — the charge for moving data to compute, to another region, or to the internet; AWS runs $0.09/GB, Azure $0.087/GB
- Replication overhead — when data is replicated across regions, both storage and transfer charges apply to every copy
For data that sits still — a compliance archive, a cold backup set, a surveillance recording library — most of these dimensions stay quiet. You pay the storage rate, almost nothing else, and the bill matches your estimate. That’s what the pricing model was designed to produce.
AI training data doesn’t sit still.
What AI training workloads actually do to cloud storage
A training run against a large dataset doesn’t read the data once. It reads it once per epoch — and most training runs require dozens or hundreds of epochs before the model converges. Each pass activates the storage layer: API calls to retrieve file chunks, data transfer from storage to GPU nodes, retrieval charges if any of the data lives in a tiered layer, and egress fees if storage and compute sit in different availability zones.
On a 10TB training dataset running 50 epochs, that’s 500TB of data movement against a dataset occupying 10TB of storage. The storage bill is small. Everything else isn’t.
The 2025 State of AI Cost Management report documented this directly: data platforms are the number-one source of unexpected AI infrastructure spend, cited by 56% of enterprises surveyed. The spend isn’t in the storage line. It’s in the four billing dimensions that activate once the data starts moving.
As explained in why AI training egress generates unexpected cloud bills at scale, the architecture creates a billing event every time training data crosses the boundary between cloud storage and GPU compute. Distributed training across multiple nodes multiplies the effect — each node pulls its own copy of the batch.
The 5% monthly retrieval assumption that appears in most cloud vs. on-premises cost comparisons — including the widely-cited Wasabi benchmark, which the updated 2026 numbers substantially revise — was derived from archive use cases. At that rate, egress is a rounding error. At 20 to 50% monthly retrieval, which is realistic for active training pipelines, egress alone runs $108,000 to $270,000 per year on a petabyte of training data. Before API calls. Before retrieval fees. Before replication.
Talk to a Nexsan® expert about on-premises storage for AI workloads →
What cloud repatriation for AI workloads actually looks like
Organizations moving AI training workloads off cloud storage aren’t typically doing it because cloud infrastructure is bad. They’re doing it because the billing model was built for a different access pattern, and the gap between estimated and actual cost became difficult to defend to a finance team.
MinIO’s 2025 survey found 48% of enterprises taking a hybrid approach to AI and machine learning workloads. IDC projects that by 2028, 75% of enterprise AI workloads will run on hybrid infrastructure. The primary driver, in most cases, is cost predictability rather than performance alone — though performance improves substantially when training pipelines move from object storage latency to local NVMe throughput.
On-premises storage collapses all five billing dimensions into a single fixed infrastructure cost, with no ongoing charges per access, per retrieval, or per byte transferred. Organizations that have worked through the cloud repatriation calculus for AI specifically report that the training pipeline economics change materially once data is local. The training run doesn’t get cheaper. The cost of running it repeatedly does.
On-premises storage for AI training pipelines: fixed cost vs. per-access billing
For teams running regular training cycles against the same dataset, the billing model isn’t a minor line item — it’s a compounding variable that grows with every epoch, every run, and every new model iteration. Removing it changes the cost structure of the entire pipeline.
Nexsan Unity NV-Series™ is a unified NVMe all-flash storage platform delivering up to 20GB/s throughput and 1 million IOPS, with capacity scaling to 19.45PB. It supports NFS, SMB, iSCSI, Fibre Channel, and S3 protocols, and works with standard ML training stacks without reconfiguring pipelines. There are no per-access charges, no egress fees, and no API billing at the storage layer. Training pipelines read data at line rate, as many times as the workload requires.
The storage cost is fixed at acquisition. The access cost is zero. That’s a different kind of storage budget than most AI teams have been working with.
See Unity NV-Series specs and configuration options →