Compute & Cost
Storage Costs Nobody Includes in the Quote
· 9 minute read
The model is the headline. The corpus, the embeddings, the traces and the 180-day logs are the invoice that grows. Put storage in the quote before it becomes a SAN surprise.
AI quotes love compute. They mention storage the way a wedding card mentions parking. Then the corpus is digitised, the embeddings are three copies, the traces are kept for quality, CERT-In's 180-day floor applies to specified ICT logs, the departmental retention schedule applies to the file, and someone asks the SDC for another tray of disks in month five.
This guide puts storage back in the BoQ. It is not a price list for rupees-per-GB — those belong on GeM and on your SDC rate card the week you buy. It is a list of objects that occupy bytes, a retention map that decides whether they occupy them for 180 days or for seven years, and a method to stop a vendor from treating your corpus as a free side effect of their licence.
Personal data in logs and embeddings is still personal data under the Digital Personal Data Protection Act, 2023. Cheap unlimited retention is not a privacy design. Expensive unlimited retention is not an audit design. Write the schedule.
The objects that occupy bytes
Source corpus: PDFs, scans, MIS extracts, audio if you are unwise. This is often the largest durable pile. Deduplicate before you embed. A second copy in the agent because the model needs it local is a choice, not a law. Embeddings and indexes can be a fraction of the corpus or, with naive overlap and many models, larger than the corpus. Each blessed model version can demand its own index. Pin versions or you will store history as RAM.
Traces and packets: prompts, retrieval sets, proposals, officer actions. This is the reconstructable file. It is not optional on rights-affecting workflows. It is also not a reason to keep raw audio of every citizen forever. ICT logs: CERT-In's 28 April 2022 directions set a 180-day retention floor for specified logs in Indian jurisdiction. That is a floor. Departmental record rules and court holds can be longer. Conversation content may be personal data. Split the stores.
Model artefacts: weights, adapters, tokenizers, eval caches. Air-gap bags multiply copies. Count the registry, not only the live node. Backups and snapshots. A 30-day snapshot policy on a growing trace store is how a 4 TB corpus becomes 40 TB of SAN without anyone training anything.
| Object | How to estimate | Retention driver |
|---|---|---|
| Corpus (authoritative) | MIS + scan inventory, after dedupe | The record schedule of the underlying file |
| Working copy for retrieval | Corpus × formats you actually parse | Purpose; delete when the file leaves purpose |
| Embeddings / index | Measure on 1,000 docs, then scale; count versions | Live model versions only, plus one rollback |
| Decision packets | Peak events × packet size from a sample | Same as the administrative file |
| ICT / security logs | Nodes × events × 180 days, then holds | CERT-In floor + your SOC + legal hold |
| Snapshots | Change rate × policy | Write the policy; default is how SAN dies |
Who pays and who hosts
If the vendor hosts, you will meet egress, snapshot and hot versus cold SKUs. Explode them. A cheap GPU hour next to a dear object store is a classic cloud shape. If you host, you will meet the SDC's rate for capacity and the wait for trays. Put a capacity reservation in the AI file so storage is not a surprise ticket in the hall. If you embed in a SaaS vector host, you have a third landlord. That landlord is a processor. Price plus DPDP plus exit. Three reasons to be deliberate.
How to stop the growth without deleting the file
Purpose tags on packets. No tag, no retain beyond the ICT floor. One live index per blessed model. Rollback is a deliberate copy, not an accident. Do not embed annexures you can retrieve by file number from the MIS. Compress scans once, well. Do not keep the raw and the AI-optimised and the thumbnail pyramid without a reason.
Prcept will export packets to a store you own and will not treat unlimited vendor-side trace retention as a feature. Storage is your record. We should not be its only landlord.
Exit is a storage problem
The last week of a contract is when you discover the embeddings are in a proprietary index and the packets are only visible in a vendor console. Write the export format in the BoQ: corpus hashes, embedding files or a documented dump, packets as JSON the department can reopen, logs in a format the SOC already stores. Time the export once in staging. If it needs a vendor engineer, it is not your store.
Price egress now. A cheap GPU hour next to a surprise egress bill is a classic cloud shape. If the landlord will not quote egress, assume it is hostile and prefer your SDC.
Objections you will hear — and what to do with them
These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.
Storage is cheap
Capacity is cheap until the hall is full or the cloud invoice discovers snapshots. Process is never cheap. Write the schedule.
We must keep everything for CAG
CAG wants a reconstructable case, not a swamp. Packets plus the administrative file beat a year of raw prompts with no index.
The vendor's unlimited storage is included
Included until the SKU changes, or until exit. Unlimited on someone else's disk is a dependency, not a gift.
CERT-In 180 days means we delete on day 181
Only for the logs that floor covers, and only if no other schedule or hold applies. Decision packets usually live with the file.
A two-week storage BoQ
- Week 1: inventory objects, measure a sample, draft a retention row for each against CERT-In, DPDP purpose and the departmental schedule.
- Week 2: add capacity and snapshot rows to the quote. Name the landlord. Price egress and exit. Delete one unnecessary copy before you buy trays.
How this shows up in the file
Subject: Storage and retention for [agent].
Objects and landlords: [table]. Retention: ICT logs at least 180 days per CERT-In 28.04.2022; packets per the administrative file; embeddings limited to live plus one rollback. Snapshots: [policy]. Personal data in traces: [purpose, access, erasure]. Capacity request to SDC / cloud: [TB, date].
This guide is not legal advice. Unlimited vendor-side retention is not a feature and not a privacy design.
This article is informational field guidance for Indian public institutions, not legal, procurement, tax, accounting, tariff or engineering advice. Confirm against the current Gazette, GFR, GeM term, SERC tariff order, IndiaAI portal rule, CAG mandate, DPDP text, departmental finance manual and your counsel before you file it. Figures are methods and order-of-magnitude illustrations, not a dataset of real deployments and not a substitute for a live quote.
How to put this in the finance note
A P2 Procurement searching “storage cost AI deployment” needs a number a CFO can defend, not a GPU brand. “Storage Costs Nobody Includes in the Quote” belongs in a cost model with people, power, idle time, AMC and the cost of a failed pilot.
The model is the headline. The corpus, the embeddings, the traces and the 180-day logs are the invoice that grows. Put storage in the quote before it becomes a SAN surprise. IndiaAI subsidy, if you use it, is a live notice — not a permanent discount. On-prem TCO includes ops headcount. Do not invent Rs/hour. Cite the source of every rupee.
- Separate capex, opex, and one-time cleanup.
- Show utilisation, not just peak GPUs.
- Price the human fallback, not only inference.
- Date every tariff and subsidy assumption.
Close this loop before the next CAB
Put “Storage Costs Nobody Includes in the Quote” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P2 Procurement, not “the vendor.”
Revisit the item when the model, the GeM term, the region, or the SI changes. “storage cost AI deployment” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.
What must be true before you file this
If “Storage Costs Nobody Includes in the Quote” is only a heading, it will not survive a file inspection. A P2 Procurement should be able to attach one artefact that proves “storage cost AI deployment”: a log export, a clause, a scored row, a dated notice, or a refusal rule.
Write three dated sentences: what was decided, who owns it, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.
- Name the owner of “storage cost AI deployment” inside the institution.
- Attach one artefact a stranger can open next year.
- Revisit when the model, the notice, or the SI changes.
- Do not treat a vendor slide as evidence.
Questions this usually raises
- What storage should an AI quote include?
- Corpus working copy, embeddings, decision packets, ICT logs, model registry and snapshots — each with a landlord and a retention row.
- Are embeddings personal data?
- They can be, when they are about an identifiable person. Treat them as in-scope until you have a written reason not to.
- How much storage will we need?
- Measure a sample of your own files. We will not print a fake GB-per-officer national average.
- Can we keep traces forever for improvement?
- Not as a default. Improvement is a purpose. It needs a basis, a minimisation, and usually a separate store from the administrative packet.
- Will Prcept host our corpus?
- We prefer your store. If we host a working copy, it is purpose-limited and exportable. Unlimited vendor-side retention is not a feature we sell.
- Does CERT-In 180 days mean delete on day 181?
- Only for the logs that floor covers, and only if no other schedule or hold applies. Decision packets usually live with the file.