Cloud & AI Storage Pricing Comparison 2026: AWS, Azure, GCP, OCI

Cloud storage is one of the most deceptively complex line items in any cloud bill. The headline price per gigabyte is only a fraction of what you actually pay. Retrieval fees, egress charges, API operation costs, and replication overhead routinely push real-world spend two to five times higher than raw storage rates. This guide breaks down what AWS S3, Azure Blob Storage, and Google Cloud Storage actually cost in 2026 — and what you can do to reduce that number.

Why Cloud Storage Costs Are Hard to Predict

Most engineers and finance teams underestimate cloud storage costs because they anchor on a single number — the per-GB storage rate. In practice, cloud providers charge for at least five distinct cost dimensions:

• Storage capacity — the raw per-GB-per-month charge, usually tiered by total volume
• API operations — every PUT, COPY, POST, LIST, GET, HEAD, and DELETE request carries a per-1,000-requests fee
• Data retrieval — relevant for infrequent-access and archive tiers; providers charge per-GB to retrieve data from cold storage
• Data transfer out (egress) — sending data to the internet or across regions incurs charges; often the most surprising line item
• Replication and cross-region transfer — replicating objects across availability zones or regions adds both transfer fees and request fees

Understanding all five dimensions is the only way to model what a workload will actually cost. The sections below cover each major provider in that framework.

This is part of a series of articles about Cloud Cost Management. 

Cloud Storage Pricing at a Glance (May 2026)

Pricing is for US regions, single-region redundancy, excluding requests, retrieval, and egress. All figures are monthly per-GB.

Sources: AWS S3, Azure Blob, Google Cloud Storage, Oracle OCI. Prices subject to change; always verify against official pricing pages for your region and redundancy tier.

AWS S3 Pricing (2026)

Amazon S3 is the reference standard for object storage, and its pricing is more layered than most competitors. All prices below are for the US East (N. Virginia) region, which typically has the lowest S3 rates globally. Other regions run 5–20% higher.

S3 Storage Tiers

What The Per-GB Rate Doesn't Tell You

• Request charges. PUT/COPY/POST/LIST is $0.005 per 1,000 requests on Standard. GET is $0.0004 per 1,000. High-frequency applications (log ingest, ML training loops) can spend more on requests than on storage.
• Egress. Data transfer out to the internet is $0.09/GB for the first 10 TB/month. Cross-region replication adds per-GB transfer cost too.
• Retrieval fees. Glacier Flexible expedited retrieval is $0.03/GB. Deep Archive standard retrieval is $0.02/GB plus $0.10 per 1,000 requests.
• Minimum storage duration. Standard-IA is 30 days. Glacier Flexible is 90. Deep Archive is 180. Delete early and you pay as if you hadn't.

A 100 TB S3 Standard workload looks like $2,304/month on paper. Add 500M GET requests ($200), 50M PUT requests ($250), and 5 TB egress ($450), and you're at $3,200. That's the real number FinOps has to budget against.

Azure Blob Storage Pricing in 2026

Azure's four-tier model (Hot, Cool, Cold, Archive) gives you more granularity than AWS on the warm end — Cold at $0.0045/GB splits the difference between IA and Glacier Instant — but redundancy drives the bill as much as the tier does.

Redundancy multiplies fast. LRS (Locally Redundant Storage) is the baseline. ZRS is roughly 1.25× LRS. GRS (Geo-Redundant) is roughly 2× LRS. RA-GRS (Read-Access Geo-Redundant) is about 2.5× LRS. If your compliance team defaults every bucket to GRS, your Hot tier is effectively $0.036/GB — double the rate card.

Azure also applies a 128 KiB minimum billable object size to Cool, Cold, and Archive tiers. Store 1 million 4 KB log files in Cool and you're billed for 128 GB instead of 4 GB.

Google Cloud Storage Pricing in 2026

GCS is the most structurally simple of the big four. Four classes, priced by location type (regional, dual-region, multi-region). Google made two notable 2026 changes: Nearline multi-region pricing went up (from $0.010 to $0.015/GB), and Archive multi-region pricing went down (from $0.004 to $0.0024/GB in US/EU). Run the new numbers before you assume your old lifecycle policies still make sense.

GCP's differentiator is that inter-region traffic inside the same multi-region is free for reads — a real savings for globally distributed apps. The catch: multi-region storage itself is priced higher than regional, and operation charges (Class A and Class B ops) can spike on list-heavy workloads.

Oracle OCI Object Storage Pricing in 2026

Oracle's pitch is consistent global pricing — the same rate in Frankfurt, Ashburn, Tokyo, and Sao Paulo — and an aggressive egress allowance (10 TB/month free across all OCI services, then $0.0085/GB). That's roughly 10× cheaper than AWS egress at scale, which makes OCI genuinely attractive for egress-heavy workloads.

The tradeoff: OCI Standard is the most expensive hot tier of the big four. OCI wins on egress-heavy patterns (video delivery, data distribution, multi-cloud architectures) and loses on storage-heavy, low-egress patterns (backups, cold archives). Don't pick OCI on the storage rate. Pick it on the total cost including transfer.

AI Storage Pricing: Vectors, Tables, and the New Line Items

Until 2025, AI storage was just "more S3 or GCS, billed the same way." That's no longer true. AWS added two AI-native storage classes that sit outside the object-storage pricing model, and they're reshaping how teams budget for RAG pipelines and data lakes.

AWS S3 Vectors

Amazon S3 Vectors went GA in December 2025 and is now live in 17+ regions. It's purpose-built for vector embeddings used in retrieval-augmented generation (RAG) and semantic search — the workloads that have been eating Pinecone, Weaviate, and managed vector DB bills alive.

• Upload (PUT): $0.20/GB of vectors uploaded
• Storage: $0.06/GB/month
• Query: Per-API-call pricing plus a $/TB charge based on index size

Storage is roughly 3× the rate of S3 Standard. That looks expensive — until you compare it to a managed vector database running on dedicated compute. AWS claims up to 90% cost reduction versus traditional vector DBs for upload, storage, and query combined. For RAG pipelines with billions of embeddings and moderate QPS, the math usually checks out.

AWS S3 Tables (Apache Iceberg)

S3 Tables is the first object store with native Apache Iceberg support — a storage class purpose-built for analytical queries against tabular data at object-store scale.

• Storage: $0.0265/GB/mo for the first 50 TB (15% above S3 Standard)
• Requests: $0.005 per 1,000 operations
• Maintenance/processing: $0.05/GB processed for compaction and optimization

The 15% storage premium is fine. The surprise is the $0.05/GB processed charge for maintenance operations — compactions and snapshot expiration can run continuously on active tables. Teams have reported 20× bill increases when they forgot that managed Iceberg is still billed operationally. Model the maintenance cost before you migrate.

Training and Inference Data Storage

Model artifacts, training datasets, and inference logs share a common pattern: write-heavy, read-heavy in bursts (training runs), then cold for long stretches. The default of leaving everything in Standard is the single biggest AI storage mistake we see.

A better pattern: Standard during active training, Intelligent-Tiering or Nearline for the 90 days after, Glacier/Archive after that. Apply lifecycle policies at the prefix level so training checkpoints, raw datasets, and inference logs each follow their own decay curve. A 500 TB training bucket stored entirely in Standard is $11,500/month. The same bucket tiered correctly is closer to $2,800/month — and the accessed subset is still hot.

The Hidden Costs That Blow Up Storage Budgets

If the rate card were the whole bill, FinOps wouldn't exist. Here's what actually shows up on the invoice:

1. Egress. Cross-AZ, cross-region, and outbound-to-internet transfer fees frequently exceed storage charges for analytics, CDN, and AI workloads. Model egress before you pick a provider.
2. Request charges. High-frequency reads, list operations, and log writes can multiply costs silently. S3 log-heavy buckets sometimes spend 2× on requests versus storage.
3. Retrieval fees. Glacier, Coldline, and Archive tiers look cheap until you actually need the data. Retrieval costs $0.02–$0.05/GB plus per-request fees.
4. Minimum storage durations. Deleting Standard-IA data in 20 days still bills you for 30. Deep Archive deletions before 180 days bill for the full period.
5. Minimum object size charges. Azure's 128 KiB minimum in cool tiers and GCP's 128 KB minimum in Nearline/Coldline/Archive punish small-file workloads.
6. Redundancy. GRS/RA-GRS in Azure and multi-region in GCS roughly double the storage rate. Default-on redundancy is a common FinOps gotcha.
7. Cross-service integration. S3 Analytics, S3 Storage Lens, Intelligent-Tiering monitoring — the observability tools that help you save money also cost money. Budget them.
8.  

How Mature FinOps Teams Control Storage Cost

The best FinOps practices for storage don't start with "pick the cheapest tier." They start with allocation — knowing which team, product, and workload owns each byte.

1. Allocate before you optimize. If you can't attribute a 200 TB bucket to a team, you can't ask that team to clean it up. This is where Virtual Tags matter — you can reallocate cost ownership based on prefix, account, or workload pattern without waiting on engineering to re-tag resources. Allocation that updates in hours, not quarters.

2. Automate lifecycle policies. Every tier transition should be automatic. Manual "let's archive Q3 logs" projects never happen. Lifecycle rules at the bucket or prefix level move data through tiers as it ages.

3. Monitor access patterns, not just storage size. The 500 TB bucket growing at 20 TB/month isn't the problem. The 50 TB bucket where 80% of data hasn't been read in 6 months is. Access pattern visibility is how you decide what to tier.

4. Model egress and requests alongside storage. Total cost of ownership for storage is storage + requests + egress + retrieval. Pick providers and architectures based on all four.

5. Apply unit economics. What does it cost to store one customer's data? One year of logs? One trained model? FinOps maturity is measured in whether you can answer those questions on demand — and whether engineering, finance, and product all see the same number.

Related content:

• Read our guide on What Is Cloud Computing

• Read our guide on Cloud Tags

Why Finout for Cloud Storage Cost Management

Storage is one of the easiest FinOps wins — if your cost data is clean, allocated, and visible. It's one of the hardest if it isn't.

Finout is built for FinOps teams managing multi-cloud storage at scale:

• MegaBill unifies AWS, Azure, GCP, OCI, and every AI and SaaS service into one cost view. No more reconciling four invoices to answer one question.
• Virtual Tags let you allocate storage cost by team, product, environment, or business unit — without waiting on engineering to re-tag a single bucket. Ownership updates in hours, not quarters.
• CostGuard surfaces storage waste (unused volumes, unattached disks, mis-tiered objects) from day one, with remediation recommendations tied to real dollar impact.
• Anomaly Detection catches storage cost spikes the day they happen, not at month-end close.
• AI Cost Management brings the same discipline to vector storage, training data, and inference workloads — the fastest-growing line items on the 2026 cloud bill.

If your FinOps team is still stitching together dashboards from AWS Cost Explorer, Azure Cost Management, and a Google Sheet, you don't have a FinOps practice — you have a spreadsheet habit. Finout is the system of record for allocation, ownership, and unit economics in the agentic era. It's what the best FinOps teams run on.