pricingcloudAI

How to Estimate Egress Costs When Hosting Large Datasets for AI Developers

UUnknown

2026-03-09

5 min read

Hook: Why egress will eat your dataset budget (and what to do about it)

If you host large datasets for AI developers, your storage bill is only the start. The real budget shock usually arrives when developers download, stage, or stream data — that's egress. With AI workloads moving petabytes of data between storage, training clusters, and edge cache, companies that ignore egress end up paying 3–10x more than expected. This guide — current for 2026 — gives storage providers and SMBs a practical, numbers-first playbook: how to estimate egress, compare storage vs egress vs compute fees, build simple calculators, and negotiate better contracts.

Top-line trends (2025–2026) that change egress math

Before we dive into formulas, understand the marketplace drivers that shape price and negotiating leverage in 2026:

Data marketplaces and monetization: Cloudflare’s January 2026 acquisition of Human Native signaled major platforms are enabling creators and dataset owners to package, license, and monetize training data — which raises expectations for transparent per-download pricing and payment flows.
Neoclouds and AI stacks: Full‑stack AI vendors (e.g., new entrants and neoclouds) are bundling storage, networking, and GPU instances. That trend makes collocated compute cheaper and reduces egress when customers train in the same network.
Better peering and regional pricing: Providers now offer targeted peering, reserved interconnects, and regional egress credits for AI workloads — these can dramatically lower long‑tail egress cost if you structure traffic accordingly.
CDN + cold storage combos: More hosting platforms couple object stores with CDN caches purpose-built for dataset shards. Caching repeated reads reduces egress by orders of magnitude for active subsets of data.

Core concepts you must separate

Storage cost (per GB-month): what you pay to hold data. Typical ranges in 2026: $0.004–$0.04/GB-month depending on hot vs cold tier and provider.
Egress (per GB transferred out): charged when data leaves the storage region or provider’s free egress scope. Typical ranges in 2026: $0.02–$0.12/GB on major clouds, lower with reserved contracts or peering.
Compute fees: GPU/CPU time to process data. For dataset hosting, compute matters because colocated compute avoids egress; remote compute increases it.
Bandwidth profile: number of downloads/streams, concurrency, and access patterns (full dataset reads vs sampled reads).

Step-by-step: How to estimate egress costs (practical)

1) Inventory your dataset

Start with precise counts:

Dataset size (GB or TB)
Number of file objects and average file size (affects TLS and request overhead)
Compression ratio if you serve compressed archives

2) Model your access patterns

Estimate these dimensions for a typical billing period (usually 30 days):

Downloads per period: how many times will the full dataset be pulled?
Partial vs full reads: percent of sessions that read only sample subsets
Concurrent sessions: peak simultaneous readers (affects CDN load)
Training reuse: does the same data get read for many epochs inside cloud compute (no extra egress if compute is collocated)?

3) Choose an egress price baseline

Gather published egress rates for the providers you might use. Use ranges if you don’t have contracted rates. Example (2026):

Major cloud public egress: $0.04–$0.12/GB
CDN or aggregated peered egress: $0.01–$0.04/GB
Reserved/committed egress discounts: can reduce prices by 30–70%

4) Calculate expected monthly egress

Formula (simple):

Monthly egress (GB) = dataset_size_GB × full_downloads_per_month + partial_read_GB

Partial read calculation example:

If dataset is 10,000 GB (10 TB) and 5 users download the entire dataset monthly, plus 50 users read 10% each (1 TB total), monthly egress = 10,000 × 5 + 1,000 = 51,000 GB.

5) Convert to cost

Multiply monthly egress (GB) by chosen egress price. Then add storage cost and any compute-related transfer fees.

Total monthly cost = storage_cost + (monthly_egress_GB × egress_price_per_GB) + compute_transfer_fees

Quick worked example

Assume:

Dataset: 20 TB = 20,480 GB
Storage price: $0.01/GB-month
Egress price: $0.06/GB
Downloads per month: 3 full downloads + 20 partial users reading 5% each (0.05 × 20 TB × 20 = 2 TB)

Monthly egress = 20,480 × 3 + 2,048 = 63,488 GB

Costs:

Storage = 20,480 × $0.01 = $204.80
Egress = 63,488 × $0.06 = $3,809.28
Total = $4,014.08 / month

Notice: egress >> storage. That’s typical for active AI datasets.

Three practical cost calculators (unit, monthly, fulfillment)

Unit calculator (per-GB basis)

Formula:

Unit cost per GB served = storage_price_per_GB_month × (1 / expected_retention_months) + egress_price_per_GB + per-GB_compute_fee

Use when you want to set a marketplace price for a dataset download or estimate per-download margins.

Monthly estimate (for budgeting)

Inputs required:

dataset_size_GB
storage_price_per_GB_month
egress_price_per_GB
expected_full_downloads_per_month
expected_partial_read_percentage and count
compute_transfer_fees (if applicable)

Compute exactly as earlier formula and present a 3-scenario table: minimal, expected, worst-case. That helps procurement negotiate caps.

Fulfillment (per-order pricing for marketplaces)

When you sell datasets (marketplace model) you need a fulfillment price per order that includes packaging, egress, and payment/contract costs.

Fulfillment fee per order = (avg_data_served_per_order_GB × egress_price_per_GB) + per_order_storage_amortization + payment_fee + overhead

Example: average order serves 250 GB, egress $0.04/GB, storage amortization $0.50/order, payment fees $0.30, overhead $0.20. Fulfillment fee = 250×0.04 + 0.50 + 0.30 + 0.20 = $11.80.

How compute placement changes egress cost (the 3 real scenarios)

Compute colocated in same provider/region: minimal egress, often only intra-region network cost (sometimes $0). Big opportunity: offer hosted training or VM bundles so your customers pay compute but you keep egress low.
Compute in same provider but different region: inter-region egress usually cheaper than public internet but still billed; model per-GB inter-region rates.
Compute outside provider: full public egress applies. This is the worst-case for dataset owners who sell downloads.

Recommendation: advertise a co-located training plan or offer staging in a compute-friendly zone to reduce billed egress.

Advanced strategies to reduce billable egress (and negotiation levers)

The following tactics turn a variable liability into predictable, or lower, costs. Use these when negotiating with cloud providers or setting prices for customers.

Bundle compute and storage: Offer
Related Reading

Unknown

Contributor

Senior editor and content strategist. Writing about technology, design, and the future of digital media. Follow along for deep dives into the industry's moving parts.