The short answer: a GPU cloud provider's headline ₹/hour rate covers the GPU itself, not what it charges for moving data into and out of its network — and in India's GPU cloud market, those two are usually billed on different terms. Ingress (data moving into the provider's environment) is close to universally free. Egress (data moving out, typically to the public internet) is frequently metered separately once usage crosses a threshold that isn't standardised across providers. AI and ML workloads generate more of exactly the traffic this billing structure targets than most general-purpose cloud workloads do, which makes the mechanics worth checking before treating a GPU-hour rate as the full cost.
Why transfer is billed separately from compute in the first place
A provider generally pays its own upstream transit cost for traffic that leaves its network for the public internet, which is the practical reason egress tends to carry a charge while traffic staying inside the provider's own infrastructure often doesn't. This isn't universal — some providers fold a data allowance into the base rate — but where it applies, it means two listings with an identical ₹/GPU-hour figure can carry very different total costs depending on how much data a workload actually moves out.
Why AI workloads are more exposed to this than most
Several patterns specific to GPU rental generate transfer volume that a typical web-hosting workload doesn't. Model checkpoints and full weight files run from single-digit gigabytes to hundreds of gigabytes, and downloading a trained model out of the provider's environment is an egress event each time it happens. Multi-node distributed training synchronises gradients between GPUs continuously; if all nodes sit inside the same provider's network this traffic is usually free, but splitting a job across regions or providers turns that same synchronisation into billable transfer. Moving a finished model to a different environment for serving, or replicating datasets across regions for redundancy, are further one-off but often large transfers. A live inference endpoint adds a different pattern again: individual responses are small, but at production scale the aggregate looks like ordinary metered API egress, billed continuously rather than as a one-time event.
Bundled, itemised, and the effect on comparisons
Because there's no standard schema for this, providers structure it in genuinely different ways. Some quote a single all-inclusive rate with a generous data allowance folded in. Others quote a leaner GPU-hour figure and meter transfer per GB beyond a free tier, which can be generous or tight depending on the provider. A third pattern makes transfer within a region or private network free while charging only for cross-region or public-internet traffic. None of these are wrong, but they mean the same headline ₹/GPU-hour number can represent a materially different total cost for a transfer-heavy workload — the same caution that applies to comparing billing granularity and hardware specs applies here too; see our guide to what a ₹/GPU-hour figure hides for the related pitfalls.
GST treatment doesn't change based on how the bill is itemised
Whether data transfer appears as a separate line item or is folded into the base charge doesn't change its tax treatment. For an India-incorporated provider, transfer charges are part of the same composite supply as the compute rental and are invoiced under the same SAC code and 18% rate — see our GST invoicing guide for the mechanics. For a foreign marketplace billing in USD, a separately itemised transfer charge is still subject to the same Reverse Charge Mechanism self-assessment as the rest of the invoice for a GST-registered Indian buyer. Itemisation is a billing-presentation choice, not a tax category.
Where the traffic actually goes matters more than the bill
Data-transfer billing tells you what you'll pay; it doesn't tell you where the data physically goes, and for personal data under the DPDP Act, that distinction matters more than the invoice line. Egress from an India-based provider's own regional infrastructure is a different residency picture from egress routed through a global CDN or a backup location outside India, even if both show up as an identical-looking "data transfer" charge. If checkpoints, training data, or logs contain personal data in scope of the Act, it's worth confirming the actual destination of any egress traffic, not just its billing category — our DPDP compliance checklist covers what to verify with a provider directly.
What to check before signing on
- Free allowance and marginal rate. How much egress is included before metering starts, and the per-GB rate beyond it.
- Intra-account and cross-region transfer. Whether moving data between your own resources on the same provider is free, and whether that changes across regions.
- Checkpoint and storage overlap. Whether checkpoint storage is billed as storage, as transfer, or both, since providers don't draw this line consistently.
- Serving-endpoint egress. Whether a live inference endpoint's outbound traffic is billed at the same rate as batch or one-off transfers, or differently.
- Where the traffic is routed. For personal data specifically, the physical destination of egress traffic, not just its price.
Frequently asked questions
Does a lower ₹/GPU-hour rate always mean a lower total cost?
Not for a transfer-heavy workload. A provider with a slightly higher GPU-hour rate but a larger free egress allowance, or free intra-region transfer, can come out cheaper overall for a workload that regularly moves large checkpoints or serves high-volume inference traffic.
Is egress pricing standardised across GST-registered providers in India?
No. There's no common formula or free-tier size across providers, which is exactly why it needs checking per provider rather than assumed from one listing to the next.
Does billing data transfer separately change its GST or DPDP treatment?
No. GST treatment follows the provider's incorporation and registration status regardless of how a charge is itemised, and DPDP obligations follow where personal data actually goes, not how the invoice line is labelled.
This is a general framework for evaluating data-transfer cost structures, not a claim about any specific provider's current egress pricing. Check the compare tool for current listings and the cost calculator to model total cost for your expected transfer pattern before making a purchasing decision.