iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
The cheapest AI storage architecture is not simply the one with the lowest price per terabyte. Reduce retained bytes first, then place each dataset, checkpoint, cache and model in the least expensive tier that still meets its latency, throughput, retrieval and durability requirements. Keep frequently read shards and active checkpoints close to accelerators; deduplicate shared data; convert suitable tabular data to compressed columnar files; and automate transitions and deletion for everything else.
Storage bills reflect capacity, redundancy, access tier, retrieval, network transfer and operations. A design that saves on capacity but leaves GPUs waiting for data, or incurs frequent archive retrieval and cross-region egress, can cost more overall. The workflow below gives you a repeatable way to lower the bill without deleting unique, regulated or expensive-to-recreate data.
What actually drives AI storage cost
For a useful comparison, calculate total cost of ownership rather than multiplying terabytes by a headline storage rate. At minimum, include:
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →- Retained bytes: raw inputs, cleaned data, feature sets, training and validation shards, checkpoints, model artifacts, logs, caches and intermediate ETL output.
- Redundancy: copies within a service, backups, snapshots, replicas and copies kept in other regions.
- Access tier: hot, standard, infrequent-access, nearline, coldline, archive or an equivalent provider tier.
- Retrieval and requests: charges that apply when cold objects are read, restored or listed.
- Network movement: cross-region transfer, internet egress and repeated copying between training jobs.
- Performance and operations: file-system capacity, provisioned throughput, cache infrastructure, administration, monitoring and the cost of idle accelerators.
There is no single current dollar-per-terabyte figure that can be applied across providers. Region, class, retention period, request volume, replication and performance requirements change the result. Build a workload-specific estimate with the provider calculator for the region you will actually use, and include retrieval and egress rather than comparing capacity rates alone.
#1 Best Overall
- MEET THE NEXT GEN: Consider this a cheat code; Our Samsung 990 PRO Gen4 SSD helps you reach near max performance with lightning-fast speeds; Whether you’re a hardcore gamer or a tech guru, you’ll get power efficiency built for the final boss
- REACH THE NEXT LEVEL: Gen4 steps up with faster transfer speeds and high-performance bandwidth; With a more than 55% improvement in random performance compared to 980 PRO, it’s here for heavy computing and faster loading
- THE FASTEST SSD FROM THE WORLD'S FLASH MEMORY BRAND: The speed you need for any occasion; With read and write speeds up to 7450/6900 MB/s you’ll reach near max performance of PCIe 4.0 powering through for any use
- PLAY WITHOUT LIMITS: Give yourself some space with storage capacities from 1TB to 4TB; Sync all your saves and reign supreme in gaming, video editing, data analysis and more
- IT’S A POWER MOVE: Save the power for your performance; Get power efficiency all while experiencing up to 50% improved performance per watt over the 980 PRO; It makes every move more effective with less consumption
Step 1: inventory and classify every object
Begin with an inventory, not a deletion campaign. Give each object a type, owner and lifecycle state so that a cheap copy is not accidentally treated as the system of record.
| Data class | Questions to record | Typical decision |
|---|---|---|
| Raw source data | Is it legally required, unique, reproducible and still needed for lineage? | Retain the authoritative copy; archive or delete only governed duplicates. |
| Cleaned and feature data | Can it be regenerated from retained raw data? How often is it reused? | Keep frequently reused versions hot; compress and tier older versions. |
| Training, validation and test sets | Which exact version produced a published result? Is there a retention obligation? | Pin reproducible releases; expire abandoned experiments. |
| Checkpoints and model artifacts | Is this the latest recoverable checkpoint, a promoted model or an obsolete experiment? | Keep active recovery points fast; move historical milestones to cold storage. |
| Logs, caches and ETL intermediates | Can the item be recreated safely, and what is the maximum useful age? | Use short expiration rules; avoid backups unless they have business or compliance value. |
For each class, record size, lineage, sensitivity, retention requirement, regeneration cost, access frequency, region and owner. Google Cloud sustainability guidance recommends automatic validation and cleaning at ingestion and periodic review of “dark and shadow data”—objects that exist but have no active owner or documented purpose. Make those fields tags or metadata so lifecycle rules can act on them.
Separate the source of truth from working copies
Keep one governed source of truth in shared object storage whenever possible. A notebook, worker or temporary cluster should read that source or a deliberately scoped cache rather than creating a full private copy. Name releases and manifests so a training run can be reproduced without retaining every transient export.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsStep 2: remove bytes that add no value
Deduplicate before changing storage classes
Use content hashes, dataset manifests or provider inventory reports to find identical files and snapshots. Deduplicate raw and processed datasets separately: two files may have different names but identical bytes, while two versions with the same name may not. Preserve the canonical copy, its lineage and the permissions required by downstream jobs before deleting duplicates.
Do not deduplicate away information that appears identical but has different legal, temporal or scientific meaning. A test set, an audit snapshot or a signed release may need to remain independent even when its bytes match another object.
Delete safely regenerable data
Ephemeral caches, failed-run output, temporary shuffles and intermediate ETL files are usually the first candidates for expiration. Set a short retention period based on the time needed to debug or rerun a pipeline. Before deletion, verify that the pipeline can recreate the item from retained inputs and that no active job or legal hold references it.
Rank #2
- Ideal for high speed, low power storage
- Gen 4x4 NVMe PCle performance
- Up to 6,000MB/s read, 4,000MB/s write
- Includes Acronis cloning software
- 5-year limited warranty
AWS Well-Architected guidance recommends shared file systems or object storage for common data, right-sized volumes, deduplicated backups and backups only for data with business or compliance value. Apply the same principle to snapshots: keep recovery points that protect an actual recovery objective, not every automatically generated snapshot forever.
Recommended Free Tools
Step 3: use efficient formats and compression
For tabular data, benchmark a columnar format such as Parquet against the current representation using representative columns, data types and query patterns. AWS Machine Learning Lens guidance reports that, compared with CSV, Parquet can reduce storage by up to 87%. That is a published AWS figure, not a guarantee: the result depends on the corpus, schema and compression settings.
Choose partitioning that matches how training and analytics jobs read data. Excessively tiny files increase metadata and request overhead; very large files can reduce parallelism and make retries expensive. Test a realistic sample for read throughput, split size, decompression CPU and end-to-end accelerator utilization before converting the entire corpus.
Choose a codec by workload, not by ratio alone
AWS also states that Zstandard produces 10–15% smaller files than Gzip at the same compression speed. Treat that as guidance for the conditions described by AWS, not a universal benchmark. Image, audio, video and already-compressed formats may gain little from another compression pass and can incur extra CPU cost.
- Measure compressed size, read bandwidth, CPU utilization and wall-clock training time.
- Keep the codec available in every environment that must read the data.
- Record format and codec in the dataset manifest so future jobs do not guess.
- Recompress cold historical data only when the storage saving exceeds the rewrite, validation and temporary-space cost.
Step 4: match storage tier to access pattern
Use the least expensive class that satisfies the workload’s latency, throughput, retrieval and minimum-retention requirements. The right answer can differ within one dataset: active shards may be hot while old releases are archival.
Free tools Windows power users keep installed
One-click scans. No signup required.
| Option | Best fit | Strengths | Costs and risks to check |
|---|---|---|---|
| Object storage | System-of-record datasets, shared artifacts, backups and large-scale capacity | Scales broadly, supports lifecycle policies and separates storage from compute | Request, retrieval and egress charges; object-listing and small-file overhead; performance varies by class |
| Managed parallel file system | High-concurrency training phases requiring file semantics and very high throughput | Shared namespace, low latency and parallel reads/writes; Google recommends Managed Lustre for sub-millisecond, high-concurrency phases | Provisioned capacity and throughput, idle cost, administration and data-ingest or export charges |
| Local NVMe SSD | Hot cache, shuffle space or checkpoint workspace beside an accelerator | Very low latency and high local bandwidth; can prevent expensive accelerator idle time | Usually ephemeral, limited capacity and endurance; you must price replication, rebuilds and data loss risk |
| Nearline, cold or archive object tier | Older datasets, completed experiments and backups that are rarely read | Lower storage rate and automated transitions | Retrieval latency and fees, minimum-retention rules, restore workflow and egress |
Keep hot data close to compute
Storage capacity is not the only performance variable. A cheap remote tier that starves GPUs or TPUs can cost more than a faster cache because the accelerator sits idle. Google Cloud TPU guidance gives starting estimates of 2 TB of dataset storage and 200 GB of checkpoint storage per TPU for LLM pre-training, 12 TB and 1 TB for multimodal training, and 1 TB and 1 GB for inference. These are starting estimates, not quotas or universal sizing rules; adjust them for replica count, retained checkpoints, sharding and concurrency.
Rank #3
- SPEED UP PROJECTS. Launch creator applications fast with uncompromising PCIe 4.0 read speeds up to 7,100MB/s,[2] (1TB and 2TB[1] models) and write speeds up to 6,700MB/s[2] (1TB[1]-4TB[1] models).
- CREATE AND STORE MORE. Make more room for your 4K videos and high-resolution images with capacities from 500GB[1] up to 4TB[1] on M.2 2280 built with our trusted 8th generation SANDISK BiCS QLC 3D CBA NAND.
- IT GOES WHERE YOU GO. With an all-new power efficient design, your drive delivers high performance with low power, giving you more time to be productive while on the go.
- UNCOMPROMISED RELIABILITY. With up to 1,200 TBW[3] (4TB[1] model) endurance rating, your drive is designed for creators.
- KEEP YOUR DRIVE UPDATED. Monitor your SSD’s performance and check for updates with the downloadable SANDISK Dashboard application.[5]
NVIDIA describes a hierarchy from storage devices through a data fabric to GPU memory and discusses NVMe, NVMe-over-Fabrics and GPUDirect Storage for high-performance paths. A local NVMe device is sensible when it serves a bounded hot cache or checkpoint workspace and measurably reduces accelerator idle time. It is not a replacement for durable, governed storage unless you add a tested replication and recovery path.
Step 5: automate transitions and expiration
Manual cleanup fails as experiment volume grows. Create rules using object age, tags, dataset state and legal-hold status. Separate two actions: transition retained data to a cheaper class, and expire data that has no remaining purpose.
Google Cloud Storage
Object Lifecycle Management can move older or unused objects to cheaper storage classes or delete them according to rules you set. Google Cloud Autoclass can move objects based on observed access patterns. Use a hot class for active shards and checkpoints, add a transition for completed or aging objects, and define an expiration for temporary artifacts. Simulate or dry-run the policy, inspect which objects match, then monitor retrievals and deletions after activation.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Amazon S3
S3 Lifecycle rules can transition objects through classes such as Standard-IA and S3 Glacier tiers, while S3 Intelligent-Tiering automatically moves objects among access tiers according to access. Apply tags for environment, owner, retention class and data state so a development cache cannot inherit a production retention policy. Check retrieval charges and minimum-retention conditions before selecting an archival class.
Azure Blob Storage
Azure Blob lifecycle management can move blobs to Cool and Archive tiers or delete them when a rule matches. Microsoft advises validating lifecycle automation before committing live assignments. Test with a noncritical prefix, verify that legal holds and immutability controls are respected, and document how an archived object is restored before relying on it for recovery.
A safe policy rollout
- Export an inventory and identify owners, legal holds, published releases and active jobs.
- Write transition and expiration rules in a simulation or dry-run mode where the service supports it.
- Review the matched-object report for false positives, small files and objects with unusual access patterns.
- Enable the policy on a test prefix or account, then observe retrieval latency, request costs and deletion results.
- Promote the rule gradually and keep an exception mechanism for irreplaceable, regulated or actively cited data.
When should old checkpoints move to cold storage?
Move a checkpoint after it leaves the active recovery window, not merely because it is old. Keep the latest checkpoints in a fast tier while a run is in progress and until you have validated a newer checkpoint. After a model is promoted, retain milestone checkpoints needed for reproducibility, rollback or audit; transition the rest when the probability of retrieval is low and the restore time fits the recovery objective.
Rank #4
- HUGE SPEED BOOST: Get random read/write speeds that are 40%/55% faster than 980 PRO; Experience up to 1400K/1550K IOPS, while sequential read/write speeds up to 7,450/6,900 MB/s reach near the max performance of PCIe 4.0*
- BREAKTHROUGH POWER EFFICIENCY: Use less power and get more performance; Enjoy up to 50% improved performance per watt over 980 PRO, plus optimal power efficiency with max PCIe 4.0 performance**
- SMART THERMAL CONTROL: Samsung's own nickel-coated controller delivers effective thermal control; With its slim size, 990 PRO is a perfect fit for desktops and laptops that meet the PCI-SIG D8 standard***
- THE CHAMPION MAKER: Up to 65% improvement in random performance enables faster loads for an ultimate gaming experience on PS5 and DirectStorage PC games****
- SAMSUNG MAGICIAN SOFTWARE: Get the most out of your SSD with Samsung Magician's advanced yet intuitive optimization tools; Monitor drive health, protect valuable data, and receive important updates for your 990 PRO
Before moving a checkpoint, store its model version, code revision, data manifest, hyperparameters, tokenizer, hardware assumptions and integrity checksum alongside it. Test at least one restore from the target tier. If restoring a checkpoint takes longer than the incident or training schedule can tolerate, it is not an adequate hot recovery copy, regardless of its low storage rate.
Control replication, region and network charges
Replicate across regions only for a defined availability, disaster-recovery or data-residency requirement. Every additional copy increases retained bytes and may create ongoing replication or transfer charges. Place compute in the same region as the authoritative data when possible. If teams repeatedly copy a common corpus across regions or accounts, evaluate shared regional storage and managed data-sharing or transfer services instead of unmanaged full copies.
Include ingress, egress, inter-region transfer and retrieval in the architecture estimate. A cold tier can be cheaper for storage but more expensive when a training team repeatedly restores large shards. Conversely, paying hot-tier rates for a dataset read once per quarter is wasteful. Measure actual read frequency and size by job, not by intuition.
Compare architectures with a total-cost scorecard
Score each candidate—object storage, managed parallel file system, local NVMe cache and archival tiers—against the same workload:
- Monthly capacity and provisioned-throughput cost.
- Read and write throughput, latency and metadata concurrency.
- Retrieval, request and egress fees.
- Durability, replication and recovery behavior.
- Minimum-retention and early-deletion rules.
- Lifecycle automation, encryption, access control and governance.
- Portability between regions or providers.
- Engineering and operational effort.
- Effect on training time, inference latency and accelerator utilization.
Use measured job traces where possible: bytes read per epoch, concurrent readers, checkpoint frequency, cache hit rate, restore time and failed-retry volume. A design that costs slightly more per month but shortens training enough to release expensive accelerators earlier may have the lower total cost.
Common failure modes and fixes
“We moved everything to archive and training became unusable”
Archive is appropriate for rarely accessed data, not active shards or checkpoints. Keep a hot working set and a tested prefetch or restore process. Transition only after measuring access frequency and acceptable restore time.
Best Value
- This product has been replaced by our latest generation. Please search for the SANDISK Optimus GX 7100 NVMe SSD
- HIGH-OCTANE GAMING. Experience speeds up to 7,250MB/s read and 6,900MB/s write (1-2TB models), with up to 35% faster performance than previous generation.
- PURPOSE-BUILT. Designed for serious on-the-go gamers, with a PCIe Gen4 interface and SANDISK’s next generation TLC 3D NAND.
- MORE TIME TO CLEAR THAT CHECKPOINT. Built with laptops and handheld gaming devices in mind, with up to 100% more power efficiency over the previous generation.
- DO MORE WITH DASHBOARD. Ensure your drive is optimized for prime performance with the downloadable WD_BLACK Dashboard (Windows only).
“Our storage rate fell but the bill did not”
Inspect retrieval, request, replication and egress line items. Large numbers of small objects can create request overhead, and repeated cross-region copies can dominate capacity savings. Consolidate files where safe, colocate compute and data, and revise transitions using actual access logs.
“Lifecycle deleted a dataset we still needed”
Require owner, retention class and legal-hold metadata before an object can match expiration. Start with a report-only policy, use an exception prefix or tag, and retain manifests and checksums so an accidental deletion is detectable.
“Compression saved space but slowed the pipeline”
Benchmark decompression CPU, file sizes, partitioning and end-to-end throughput on representative samples. A smaller file is not a saving if it extends accelerator time or forces serial reads. Keep an uncompressed or faster intermediate only when its performance value is demonstrated.
“Local NVMe lost the only copy of a checkpoint”
Treat local NVMe as a cache or workspace unless its durability is explicitly engineered. Upload checkpoints and manifests to durable object storage, verify checksums, and test recovery after instance or node replacement.
A practical operating cadence
At ingestion
- Validate schema, deduplicate where appropriate and write an owner, lineage and retention tag.
- Convert suitable tabular data to the chosen columnar format and record codec and partitioning.
- Place the authoritative copy in the designated region and prevent uncontrolled notebook copies.
During training and serving
- Keep active shards, current checkpoints and serving models in the hot path required by throughput targets.
- Measure cache hit rate, accelerator idle time, bytes read and checkpoint restore time.
- Expire failed-run intermediates and caches on a schedule short enough to prevent accumulation.
After a run or release
- Mark the final model, reproducibility manifest and required milestone checkpoints.
- Transition inactive but valuable artifacts to the appropriate infrequent-access or archive class.
- Delete abandoned experiments, duplicate exports and snapshots without business or compliance value.
At least quarterly
- Review dark and shadow data, owners, access logs, lifecycle matches and retrieval costs.
- Recalculate total cost using current regions, tiers, replication and workload volume.
- Restore a representative archived dataset or checkpoint to verify that documentation and automation still work.
The decision in one sentence
Retain unique and required data, eliminate redundant and safely regenerable bytes, compress active and retained datasets appropriately, automate tiering and expiration, and pay for high-performance storage only where it protects throughput or recovery time.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
