The 3-2-1 rule says to keep three copies of your data, on two different types of storage, with one copy off site. It is still the right foundation, and most small AI setups do not yet meet it. But it was written for drive failures and fires, and it has two gaps that ransomware exposed: all three copies can be online and reachable by the same attacker, and nothing in the rule requires anyone to check that a backup can be restored. The updated form, 3-2-1-1-0, adds one copy that is offline or cannot be altered, and zero errors confirmed by testing recovery. For AI infrastructure, where a restore means multi-gigabyte models and a working software environment, that final zero is the part most often missing.
What does each number mean?
| 3-2-1 | 3-2-1-1-0 | |
|---|---|---|
| 3 | Three copies: the original and two backups | Same |
| 2 | Two different storage types or devices | Same |
| 1 | One copy off site | Same |
| 1 | - | One copy offline, air-gapped or immutable |
| 0 | - | Zero errors: backups verified and restores tested |
| Protects against | Drive failure, device loss, site disaster | Also ransomware, compromised administrators, silent backup failure |
Why is plain 3-2-1 no longer enough?
- Everything is connected. A NAS, a second NAS at another office and a cloud sync folder can all be reached from one compromised administrator account. Three copies, one point of failure.
- Sync is not backup. A synchronised copy faithfully receives the encrypted or deleted version.
- Backups fail quietly. A job that has been erroring for three months, a full volume, a changed password. The first restore attempt is a bad moment to find out.
- Restores are harder than they look. Getting files back is one thing; getting a working AI service back is another.
How does it map onto an AI setup?
| Rule | In a small AI deployment |
|---|---|
| Copy 1 | Working data on the AI system's internal NVMe |
| Copy 2, second device | Scheduled, versioned backup to a NAS with redundant drives |
| Copy 3, off site | Replication to a second NAS or backup appliance at another location, or a cloud backup target |
| The extra 1 | Immutable snapshots or a protected backup appliance; or media physically disconnected after each run, such as tape or a rotated drive |
| The 0 | Automated verification of backup jobs, plus a timed restore of a model, a dataset and the serving environment |
One device can satisfy more than one number. An off-site backup appliance with immutability is both the off-site copy and the unalterable one.
Does everything need all five?
No. Apply the full rule to what cannot be recreated, and less to what can.
| Data | Suggested level |
|---|---|
| Fine-tuned models, adapters, labelled datasets, prompts, configuration, evaluation sets | Full 3-2-1-1-0 |
| Source documents and the company file share | Full 3-2-1-1-0 |
| Raw data that could be re-collected with effort | 3-2-1 |
| Public base models | One extra copy on the NAS, with version and source recorded |
| Vector index | One backup; it can be rebuilt from the documents |
| Intermediate checkpoints, scratch files | Retention rule, not full backup |
CALC. A team holds 6TB in total: 1TB irreplaceable, 2TB re-collectable raw data, 3TB of public models and scratch. Applying the full rule to everything means two further copies of 6TB, or 12TB of backup capacity. Tiering needs 2 x 1TB + 2 x 2TB + 1 x 3TB = 9TB, and only 1TB has to sit on the more costly immutable and off-site tier. The figures are illustrative; the method is to sort before you size. See how to back up AI models, datasets and checkpoints.
3-2-1 asks how many copies you have. The two extra digits ask the harder questions: can anyone destroy them all, and have you ever actually got one back?
What does the zero involve?
- Job monitoring. Alerts on failed or missed backups, and on backups that suddenly shrink or grow.
- Integrity checks. Scheduled verification that backup data is readable.
- File-level restore test. Monthly: recover a named file from last week.
- Service-level restore test. Twice a year: rebuild the AI service on a spare or loan machine from backups and written steps, and time it.
- A record. Date, what was restored, how long it took, what went wrong.
The timed result is your real recovery time. Compare it with what the business said it could tolerate; see backup is not disaster recovery.
Is tape still relevant?
For the offline copy, yes. A tape on a shelf cannot be encrypted over the network, and cost per terabyte is low at scale. It is slower to restore from and needs a drive and a routine. We weigh it in is LTO tape backup still relevant in 2027. For most small teams, immutable snapshots plus an off-site appliance or cloud target is simpler to operate.
Buying in Singapore
Copy 2: the Synology DS925+ (S$1,260) or Synology DS1825+ (S$2,148), with Seagate IronWolf Pro drives from S$390. Off-site and protected copy: the Synology ActiveProtect DP320 (S$2,611.20) or DP340 (S$6,526.80). Offline media: the HPE LTO-8 Ultrium 20-cartridge library pack (S$3,205), for sites that already run a tape drive (listed 12 Nov 2026; prices change; ask for a current quotation). To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is the 3-2-1 backup rule? Keep three copies of your data, on two different types of storage or devices, with one copy stored off site.
What is the 3-2-1-1-0 rule? It extends 3-2-1 with one copy that is offline, air-gapped or immutable, and zero errors, meaning backups are verified and restores are tested.
Is the 3-2-1 rule still enough? It remains a sound base but does not address ransomware that reaches every online copy, or backups that fail silently. The extra immutable copy and restore testing close those gaps.
Does every AI file need full 3-2-1-1-0 protection? No. Apply it to what cannot be recreated, such as fine-tuned models, labelled datasets and configuration. Public base models and scratch files need less.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
