How much do data egress fees add to the cost of moving AI training data between cloud providers in 2027?

The outcome you should expect

When teams plan a 2027 migration of AI training datasets between cloud providers, the egress line item is rarely the single largest cost — but it is the most underestimated. For a mid-sized training corpus of 500 TB, list-price egress alone lands between $25,000 and $60,000 depending on the source provider's tier and the destination's region. Add request charges, cross-AZ traffic, and re-downloads after failed transfers, and the real number often lands 20–30% higher than the initial quote.

The outcome most organizations should plan for is not a single egress bill but a compounding one. Training data is not moved once. It is staged, re-staged, versioned, and re-pulled every time a preprocessing pipeline changes. A dataset that costs $40,000 to move once can cost $150,000–$250,000 over a full training cycle if pipelines re-read from the origin cloud instead of a local cache. The providers that win migrations in 2027 are the ones whose customers treat egress as a recurring operational cost, not a one-time exit fee.

Expect three distinct outcomes depending on your posture. First, unplanned migrations — where teams move data reactively after a contract dispute or a capacity crunch — routinely see egress consume 35–50% of the total migration budget because there is no time to negotiate or compress. Second, planned migrations with 6–12 months of lead time typically land egress at 15–25% of total migration cost, because teams can use committed-use discounts, compression, and incremental transfer windows. Third, hybrid architectures that never fully leave the origin cloud can keep egress under 10% by moving only the deltas — but they trade that savings for architectural complexity and dual-vendor management overhead.

The single most important expectation to set with finance is that egress is not a fixed fee. It scales linearly with data volume, which means it scales with model size, dataset size, and retraining frequency. A team that doubles its training corpus in 2027 doubles its egress exposure unless it renegotiates. Budget owners who treat egress as a one-time line item will be surprised every quarter.

What drives that outcome

The egress bill in 2027 is driven by five compounding factors, and understanding how they interact is the difference between a $30,000 migration and a $300,000 one.

Volume and velocity. Egress is priced per gigabyte, so the first driver is simply how many bytes leave the origin cloud. AI training data is unusually heavy: multimodal datasets with video, high-resolution imagery, and long-context text can run 10–100x larger than traditional enterprise datasets. A single training run on a frontier-scale multimodal model can involve petabytes of data movement. Even at the low end of 2027 pricing — roughly $0.05 per GB for high-volume committed transfers — a petabyte costs about $50,000 to move once.

Distance and direction. Traffic between regions within the same provider is cheaper than traffic between providers. Cross-region egress within a single cloud might run $0.01–$0.02 per GB, while inter-provider egress — the exact scenario this page addresses — runs $0.05–$0.12 per GB at list. The premium exists because inter-provider traffic leaves the provider's network entirely, and providers have historically priced that as a competitive disincentive to leave.

Request and operation charges. Egress is not just bytes. Every GET request, every list operation, and every API call that retrieves data carries its own micro-charge. For training data stored as millions of small objects — common with image or audio datasets — request charges can add 10–20% on top of the per-GB fee. A dataset with 50 million 100 KB images incurs 50 million GET requests, and at $0.0004 per 1,000 requests, that is $20,000 in requests alone before a single byte of egress is billed.

Compression and deduplication. The single largest lever a practitioner controls is how much data actually crosses the boundary. Training datasets are often 60–80% redundant — duplicate images, near-identical text passages, repeated frames. Deduplicating before transfer can cut egress volume by half or more. Compression adds another 2–5x reduction for text and tabular data, less for already-compressed media. A team that deduplicates and compresses aggressively can move the same effective dataset for 20–30% of the naive egress cost.

Negotiated discounts and committed-use agreements. List price is rarely what enterprises pay. Providers routinely offer 40–70% discounts on egress in exchange for committed spend, multi-year agreements, or reciprocal traffic commitments. The catch is that these discounts are negotiated, not published, and they often come with minimum volume thresholds that smaller teams cannot meet. A startup moving 50 TB pays list price; an enterprise moving 5 PB negotiates.

The diagram above shows why egress is not a single number. Each input either increases the bytes moved or changes the rate applied to those bytes. Practitioners who model all five inputs before committing to a migration typically find that the naive estimate — volume times list rate — overstates or understates the real bill by 30–50% depending on which levers they pull.

Benchmarks and realistic ranges

Concrete numbers matter more than principles here, so the following ranges reflect what teams actually report paying in 2026–2027 for inter-provider AI training data movement. These are list-price and lightly-negotiated figures; heavily negotiated enterprise rates can be lower.

Per-GB egress rates. Inter-provider egress at list price in 2027 runs $0.05–$0.12 per GB. The low end applies to high-volume committed transfers (multi-petabyte agreements), the high end to on-demand, low-volume transfers. Cross-region within the same provider runs $0.01–$0.02 per GB. Internet egress to a non-cloud destination runs $0.08–$0.12 per GB. These rates have been declining slowly — roughly 5–10% per year — but not fast enough to offset growing data volumes.

Total cost for representative dataset sizes. A 100 TB training dataset moved between providers at list price costs $5,000–$12,000 in egress alone. A 500 TB dataset costs $25,000–$60,000. A 1 PB dataset costs $50,000–$120,000. A 5 PB dataset costs $250,000–$600,000. These figures exclude request charges, which add 10–20% for small-object datasets, and exclude re-downloads, which can double the figure if pipelines are not cached.

Egress as a share of total migration cost. For a planned migration with 6–12 months of lead time, egress typically represents 15–25% of total migration cost. Total migration cost includes engineering labor, dual-running infrastructure during cutover, data validation, and re-training of pipelines. For reactive migrations with no lead time, egress can represent 35–50% of total cost because there is no time to negotiate, compress, or stage.

The compression and dedup multiplier. Teams that deduplicate before transfer report moving 40–60% fewer bytes. Teams that compress text and tabular data report 2–4x reduction on top of dedup. Combined, aggressive preprocessing can reduce effective egress cost to 20–35% of the naive estimate. This is the single highest-leverage action available to a practitioner, and it costs engineering time rather than vendor negotiation.

Negotiated discount ranges. Enterprise agreements in 2027 typically secure 40–70% off list egress rates. The discount depends on total committed spend, contract length, and whether the customer commits to reciprocal traffic. A customer committing $10M+ annually across compute and storage can often negotiate egress to $0.02–$0.04 per GB. A customer committing $500K annually typically gets 20–30% off list, not more.

The re-download penalty. The most commonly underestimated cost is re-download. If a training pipeline reads from the origin cloud on every run instead of caching locally, a dataset moved once can be effectively moved 5–20 times over a training cycle. At 10 re-downloads, a $40,000 egress bill becomes $400,000. Caching the dataset in the destination cloud after the first transfer eliminates this entirely, at the cost of destination storage — typically $0.02–$0.03 per GB per month, which for 500 TB is $10,000–$15,000 per month.

Sanity-check heuristics. A quick way to estimate 2027 egress cost: take dataset size in TB, multiply by $50–$120, and that is the list-price egress range. Multiply by 0.3–0.6 if you have negotiated discounts, and by 0.2–0.35 if you deduplicate and compress aggressively. If your estimate is below $20 per TB, you are probably missing request charges or re-downloads. If it is above $150 per TB, you are probably quoting on-demand rates for a workload that should be committed.

Risks, edge cases, and failure modes

Egress cost overruns rarely come from the per-GB rate. They come from the edges — the cases teams do not model until the bill arrives.

The small-object tax. Datasets stored as millions of small files — images, audio clips, individual documents — incur request charges that can rival the per-GB fee. A 200 TB image dataset with 100 KB average object size contains 2 billion objects. At $0.0004 per 1,000 GET requests, that is $800,000 in request charges alone. The fix is to repack small objects into larger archives before transfer, which can reduce request counts by 1,000x. Teams that skip this step routinely see egress bills 3–5x their per-GB estimate.

The re-download loop. If the destination pipeline is not configured to read from local storage, every training run re-pulls from the origin. This is the single most expensive failure mode because it is invisible until the bill arrives. A team that moves 500 TB once but re-downloads it 12 times pays 12x the egress. The fix is architectural: cache the dataset in destination object storage, and make the training pipeline read from that cache exclusively. The storage cost is trivial compared to the egress savings.

The partial-migration trap. Teams that move only part of their data — say, the training set but not the validation set, or the current version but not the historical versions — often find themselves re-pulling the missing pieces repeatedly. Partial migrations also complicate deduplication, because the dedup logic cannot see the full corpus. The fix is to move the entire dataset in one planned window, even if it costs more upfront, because incremental moves fragment the cache and multiply egress.

The cross-AZ surprise. Traffic between availability zones within the same region is often billed as egress, even though it never leaves the provider. Teams that assume "same region equals free" are surprised by cross-AZ charges of $0.01–$0.02 per GB. For a distributed training job that shuffles data across zones, this can add 5–15% to the bill. The fix is to pin training data and compute to the same AZ where possible, or to accept the cross-AZ cost as a known line item.

The contract-exit timing risk. Egress discounts are often tied to committed-use agreements. If a team exits a contract early, the discount may be clawed back, and the remaining egress may be billed at list price. Teams that plan migrations around contract renewal dates can avoid this; teams that migrate reactively often pay a premium. The fix is to align migration timelines with contract cycles and to negotiate exit terms explicitly.

The vendor-lock-in asymmetry. Some providers charge more for egress to specific competitors than to others. This is not always published, and it can make one destination significantly more expensive than another for the same data volume. The fix is to request egress pricing for each specific destination during negotiation, not just a generic per-GB rate.

The compliance and residency edge case. Training data that contains regulated information — health records, financial data, personal data — may be subject to residency requirements that prevent it from leaving certain jurisdictions. In those cases, egress is not just expensive; it may be prohibited. Teams must confirm residency rules before planning any inter-provider move, because a dataset that cannot legally leave the origin cloud cannot be migrated at any price.

The validation cost. After transfer, teams must validate that the data arrived intact. For large datasets, validation itself requires reading the data, which can incur additional egress if validation runs from the origin. The fix is to validate in the destination cloud, using checksums computed before transfer. This eliminates validation egress entirely.

A practical rollout plan

The following plan reflects what teams that successfully control egress costs in 2027 actually do. It assumes a planned migration with at least six months of lead time.

Phase 1: Inventory and model (weeks 1–4). Before moving anything, inventory the dataset. Count objects, measure average object size, identify redundancy, and estimate compressed size. Model the egress cost using the five drivers above. Identify which datasets are read-mostly (candidates for one-time move) versus read-write (candidates for hybrid). Output: a per-dataset egress estimate and a prioritized migration order.

Phase 2: Preprocess and compress (weeks 4–12). Deduplicate the dataset. Compress text and tabular data. Repack small objects into archives. This phase typically reduces effective transfer volume by 50–70%. It costs engineering time but no vendor fees, and it is the highest-leverage action in the entire plan. Output: a preprocessed dataset ready for transfer, with checksums.

Phase 3: Negotiate and stage (weeks 8–16). Negotiate egress discounts with the origin provider, using the modeled volume as leverage. Request destination-specific rates. Stage the transfer in windows that align with contract terms. Set up destination storage and caching before the first byte moves. Output: a signed egress agreement and a destination environment ready to receive.

Phase 4: Transfer and validate (weeks 16–20). Execute the transfer in planned windows. Validate checksums in the destination. Configure the training pipeline to read from destination storage exclusively. Decommission origin-side read paths to prevent accidental re-downloads. Output: a fully migrated dataset with no origin dependency.

Phase 5: Monitor and optimize (ongoing). Track egress spend monthly. Watch for re-downloads, cross-AZ traffic, and request-charge spikes. Re-negotiate discounts annually as volume grows. Output: a recurring egress cost that stays within 15–25% of total migration cost.

The plan above assumes the migration is planned. For reactive migrations, compress phases 1–3 into days rather than weeks, accept higher egress cost, and prioritize eliminating re-downloads above all else, because that is the failure mode that compounds fastest.

What to do differently at scale. Teams moving more than 1 PB should consider a physical transfer appliance rather than network egress. Providers offer physical data transfer services that bypass per-GB egress fees entirely, at the cost of shipping time and handling. For datasets above 5 PB, physical transfer is often 5–10x cheaper than network egress, even accounting for appliance rental and shipping. The trade-off is latency: physical transfer takes days to weeks, while network transfer can complete in hours.

What to do differently for small teams. Teams moving less than 50 TB rarely have leverage to negotiate egress discounts. For them, the levers are compression, deduplication, and avoiding re-downloads. A small team that deduplicates aggressively and caches in the destination can keep egress under 20% of migration cost even at list prices. The mistake small teams make is assuming they cannot control egress cost because they cannot negotiate; in practice, preprocessing discipline matters more than discount rate for volumes under 50 TB.

Related questions

Do egress fees apply to data moved between regions of the same provider?

Yes, but at lower rates. Cross-region egress within a single provider typically runs $0.01–$0.02 per GB in 2027, versus $0.05–$0.12 per GB for inter-provider transfers. The premium for inter-provider traffic reflects the fact that it leaves the provider's network entirely.

Can egress fees be eliminated entirely?

Rarely. Some providers offer free egress for specific services or under specific commitments, but inter-provider AI training data movement almost always carries a per-GB fee. Physical transfer appliances can bypass network egress fees for very large datasets, but they introduce shipping time and handling costs.

How does compression affect egress cost?

Compression reduces the number of bytes that cross the boundary, which directly reduces the per-GB charge. Text and tabular data compress 2–4x; already-compressed media compresses little. Combined with deduplication, compression can cut effective egress cost to 20–35% of the naive estimate.

What is the biggest hidden egress cost?

Re-downloads. If the training pipeline reads from the origin cloud on every run instead of a local cache, a dataset moved once can be effectively moved 5–20 times over a training cycle. Caching in the destination cloud eliminates this entirely at a small storage cost.

Are egress rates expected to fall by 2027?

Rates have been declining roughly 5–10% per year, but data volumes are growing faster. The net effect is that total egress spend continues to rise even as per-GB rates fall. Teams should plan for rising total egress cost, not falling.

FAQ

How much do data egress fees add to the cost of moving AI training data between cloud providers in 2027? They typically add 15–40% to total migration cost, or roughly $50–$120 per TB at list prices. Negotiated discounts can cut that by 40–70%, and aggressive deduplication and compression can cut it further. The exact figure depends on volume, distance, object count, and contract terms.

What is the single biggest factor driving egress cost? Effective bytes moved. Deduplication and compression can reduce transfer volume by 50–70%, which is a larger lever than any negotiated discount for most teams. The second biggest factor is re-downloads, which can multiply the bill 5–20x if pipelines are not cached.

Do all cloud providers charge the same egress rates? No. List rates vary by provider and by destination. Inter-provider egress to a specific competitor may be priced higher than generic internet egress. Teams should request destination-specific rates during negotiation rather than assuming a single per-GB number.

Can physical transfer replace network egress for large datasets? Yes, for datasets above roughly 1 PB. Physical transfer appliances bypass per-GB network egress fees at the cost of shipping time and handling. For very large datasets, physical transfer is often 5–10x cheaper than network egress, though it takes days to weeks rather than hours.

How should teams budget for egress in 2027? Budget egress as a recurring operational cost, not a one-time migration fee. Model it using the five drivers — volume, distance, request count, compression, and discounts — and assume 15–25% of total migration cost for planned moves, 35–50% for reactive ones. Re-negotiate annually as volume grows.

What is the most common egress mistake teams make? Assuming egress is a one-time cost and failing to cache the dataset in the destination cloud. The second most common mistake is ignoring request charges for small-object datasets, which can rival or exceed the per-GB fee. Both are preventable with preprocessing and pipeline discipline.

Sources

• AWS Data Transfer Pricing: https://aws.amazon.com/ec2/pricing/on-demand/
• Google Cloud Network Pricing: https://cloud.google.com/vpc/network-pricing
• Microsoft Azure Bandwidth Pricing: https://azure.microsoft.com/en-us/pricing/details/bandwidth/
• Cloudflare Bandwidth Alliance: https://www.cloudflare.com/bandwidth-alliance/
• Gartner Cloud Infrastructure Cost Research: https://www.gartner.com/en/infrastructure-and-it-operations-leaders
• FinOps Foundation Cloud Cost Guidance: https://www.finops.org/
• Uptime Institute Data Center Cost Analysis: https://uptimeinstitute.com/
• IDC Cloud Infrastructure Spending Tracker: https://www.idc.com/

Related on PULSE

• How do cloud providers price inter-region data transfer in 2027?
• What is the real cost of re-downloading training data from object storage?
• How does dataset deduplication reduce AI training infrastructure cost?
• When does physical data transfer beat network egress for large datasets?
• How should RevOps model cloud egress in a multi-year AI budget?
• What contract terms reduce data egress fees in enterprise cloud agreements?