The hidden costs of GPU cloud are posted by the same providers that post the GPU-hour rate: data transfer, storage for datasets and checkpoints, public IP addresses, support plans and tax. On 8 November 2026, AWS posted its 8-GPU H100 instance in US East (N. Virginia) at $55.04 an hour [3, 6], or $40,179.20 for a 730-hour month (computed for this report). Its other price pages listed $0.09 per GB for the first paid tier of internet egress, $0.01 per GB in each direction between Availability Zones and $0.005 an hour for each public IPv4 address [1, 2, 4].
This explainer reads those line items from price pages as posted on 8 November 2026, at hyperscalers, the large general-purpose clouds, and at neoclouds, the clouds built around GPU capacity. It prices each with examples computed for this report and sets out what exit waivers and the EU Data Act change. From 12 January 2027 the Data Act bars charges for switching providers, egress for the switch included. Egress to another provider used at the same time may still be charged at cost [25, 27].
What the GPU-hour price includes
An 8-GPU node's price bundles CPU cores, host memory and local disks, and the bundle differs by provider.
| Provider and node | Posted price, 8 November 2026 | vCPUs | Host memory | Local storage |
|---|---|---|---|---|
| AWS p5.48xlarge, US East | $55.04 an instance-hour, $6.88 per GPU-hour computed [3, 6] | 192 | 2 TiB | 8 x 3.84 TB NVMe SSD [6] |
| CoreWeave HGX H100, North America | $49.24 a node-hour, $6.155 per GPU-hour computed [19] | 128 | 2,048 GB | 61.44 TB "Local Storage" [19] |
| Lambda 8x HGX H100 | $4.59 per GPU-hour from 16 November 2026 [20] | 208 | 1,800 GiB | 22 TiB SSD [20] |
| Nebius HGX H100 | $4.50 per GPU-hour [22] | 16, listed with the per-GPU price | 200 GB, listed with the per-GPU price | No local disk listed [22] |
AWS's local disks keep nothing once an instance stops: every block of instance store is erased when an instance stops, hibernates or terminates, and AWS's guide says not to rely on it for valuable, long-term data [7]. Durable copies of datasets and checkpoints go to object, file or block storage, billed apart.
Billing increments differ too. AWS bills Linux instances per second after a 60-second minimum [1]. Lambda bills on-demand instances in one-minute increments for as long as they run, whether or not they are in use [21]. AWS prices exclude VAT and sales tax, and Lambda's table adds "applicable sales tax/VAT/GST" [1, 20].
Hidden costs of GPU cloud, line by line
| Line item | Posted rate, 8 November 2026 | Source |
|---|---|---|
| Internet egress, AWS US East | $0.09 per GB for the first paid tier, after 100 GB a month free | [1, 2] |
| Internet egress, Google Premium Tier, to North America | $0.12 per GiB to 1,024 GiB, then $0.11 | [12] |
| Internet egress, Azure, from North America or Europe | $0.087 per GB for the first paid tier, after 100 GB a month free | [16] |
| Between zones in a region | $0.01 per GB each way at AWS, $0.01 per GiB at Google | [1, 12] |
| Between regions | From AWS US East, $0.01 per GB to Ohio and $0.02 to Oregon. At Google, $0.02 per GiB within North America to $0.14 to South America | [2, 12] |
| Object storage | $0.023 per GB-month, S3 Standard | [8] |
| Block storage | $0.08 per GB-month, EBS gp3 | [9] |
| Public IPv4 address | $0.005 an hour at AWS | [4] |
| NAT gateway | $0.045 an hour plus $0.045 per GB, AWS's US East (Ohio) example | [4] |
| Support | AWS Business Support+, 9% of the first $10,000 a month, then 7% to $80,000 | [10] |
Google's Premium Tier row is the default table on its page, and its rates depend on the source region [12].
Egress fees and free egress claims
A 2022 study of bulk transfers found egress priced by the volume moved, with no matching charge for data coming in [34]. AWS's 100 GB a month is one global allowance across its services and regions [1]. Google's Standard Tier gives 200 GiB a month per account, then charges $0.085 per GiB [12]. Oracle's first 10 TB a month of outbound transfer is free [24].
Four neoclouds post no network egress charge. CoreWeave lists transfer to the internet and within CoreWeave as free [19]. Lambda advertises "no egress fees" [20]. Crusoe's FAQ opens with "At this time" and says it does not charge for ingress or egress [23]. Nebius lists egress and ingress traffic as free, and its Standard object storage class still charges $0.015 per GiB of egress traffic [22].
Sending the FineWeb web corpus, 50,446.9 GB on disk, out of that Nebius storage class would cost about $705. Sending it once to the internet from AWS US East would cost about $4,324 (computed for this report from [1, 2, 22, 33]). For node prices side by side, see Neocloud and hyperscaler GPU prices, compared like for like.
Data transfer between zones and regions
Inside one provider, data crossing zones and regions is billed at the rates in the table above [1, 2, 12]. Azure lists transfer within one Availability Zone as free, $0.02 per GB between regions within North America or within Europe, and $0.05 from either to other continents [16].
The MLPerf Storage rules put a 70-billion-parameter model's checkpoint at 912 GiB [32]. Copying one from US East to US West (Oregon) at AWS costs about $19.59. At Google it costs $9.12 to another zone, $18.24 within North America and $45.60 to Europe (computed for this report from [2, 12, 32]). A January 2026 preprint on fine-tuning jobs run across regions on spot instances found egress for moving checkpoints at up to 16% of total cost [35], a result covered in What research shows about training on spot GPUs.
Checkpoint and dataset storage
A checkpoint is many times the size of the model's weights. Mixed-precision training with Adam needs 16 bytes of memory per parameter [31], and the MLPerf Storage 70B checkpoint works out to about 14 bytes per parameter (computed for this report from [32]).
| Storage class | Rate | Ten 70B checkpoints, one month |
|---|---|---|
| Nebius Standard object storage | $0.0147 per GiB-month [22] | $134.06 |
| Amazon S3 Standard, first 50 TB | $0.023 per GB-month [8] | $225.23 |
| CoreWeave object storage, hot | $0.06 per GB-month [19] | $547.20 |
| Amazon EBS gp3 | $0.08 per GB-month [9] | $783.40 |
Amounts are computed for this report. CoreWeave counts storage in binary gigabytes [19].
Storage bills run while data sits idle. Lambda bills a filesystem for as long as it exists, mounted or not [21]. AWS bills EBS by the GB provisioned until the storage is released [9].
IP addresses and support
At AWS, a public IPv4 address costs $0.005 an hour, in use or idle [4], or $3.65 over 730 hours (computed for this report). CoreWeave charges $4.00 per address a month, and Nebius lists public IP addresses as free [19, 22].
Support is a percentage of the bill. AWS Business Support+ on one p5.48xlarge month comes to $3,012.54, about 7.5% added to the instance bill (computed for this report from [3, 10]).
Exit waivers and Data Act egress fees
AWS, Google Cloud and Azure credit egress for customers who move their data off. AWS credits the data migrated once AWS Support approves a request, and its 8 October 2026 update gives eligible customers outside the EU 180 days to complete the move [5]. Google requires an Exit Notice and credits the transfer cost on the final invoice [14]. Azure credits up to 60 calendar days of egress charges from the start date the customer names [18].
The EU Data Act sets the rule for switching. Article 29 allows reduced switching charges, capped at the costs directly linked to the switch, until 12 January 2027, and from that date bars any switching charges [25]. Article 2 counts data egress charges for the switch as switching charges and leaves out standard service fees and early termination penalties [25]. Article 34 lets a provider charge egress for in-parallel use of several providers, "without exceeding such costs" [25]. The Commission's FAQ says this in-parallel egress may still be billed at cost after 12 January 2027 [27].
What the UK CMA found about egress fees
The UK Competition and Markets Authority (CMA) found in July 2025 that egress fees are a key commercial barrier to switching and multi-cloud [29]. In its estimates, published as ranges, 10-20% to 20-30% of customers would have had to pay egress fees above 5-10% of a year's cloud spend to move all their data within one year [30]. The review covered UK cloud customers across all workloads.
Putting the line items on one bill
For one p5.48xlarge in US East for a month, ten 70B checkpoints in S3 Standard add 0.6%, one cross-zone read of FineWeb served from instances adds 2.5%, Business Support+ adds 7.5% and sending FineWeb once to the internet adds 10.8% (computed for this report from [1, 2, 3, 8, 10, 32, 33]). What posted GPU cloud prices tell you, and what they do not lists billing rules among the seven adjustments a fair comparison needs. Reading GPU prices from cloud pricing APIs covers reading provider prices from pricing APIs and price files.
Two prices compare fairly only on one basis. Rillor's GPU and neocloud pricing method converts each price to U.S. dollars per GPU-hour and tags each offer with GPU variant, interconnect, node size, region and term.
Read the full report
The full report adds rate tables by provider, every worked example with its arithmetic, the exit terms and Data Act articles, its limits and all 35 sources.