Data Storage Devices: All You Need to Know in 2024

Every few years we write a version of this post because the storage landscape shifts enough that the old one is no longer correct. The fundamentals don't change much, but the specific products, the price per gigabyte, and the pragmatic choices move around. Here is the state of data storage devices as we see it in 2024, written for the IT decision-maker who has to choose where data actually lives.

The four layers of the storage hierarchy

The way we actually think about storage today has four layers. Every workload lives at one of them, and the cost and performance math changes by roughly an order of magnitude at each step.

Layer 1 — Hot, latency-sensitive: NVMe flash. Sub-millisecond read latency, hundreds of thousands of IOPS per device, and a cost per GB that is no longer absurd. This is where databases live, where the hot working set of a VDI environment lives, where any application that users are actively waiting on should live. Enterprise NVMe drives in 2024 run roughly $0.10 to $0.25 per usable GB depending on endurance class and form factor.

Layer 2 — Warm, mixed workload: SATA/SAS SSD and high-density NVMe. Slightly higher latency than Layer 1, much higher capacity, meaningfully cheaper. QLC NAND has made this layer very interesting — you can buy all-flash capacity storage at roughly $0.05 to $0.10 per usable GB, fast enough for nearly any file serving or virtualization workload that isn't transaction-database-heavy. This is the layer that killed the 10K and 15K SAS hard drive.

Layer 3 — Cold, capacity-first: nearline HDD. Spinning disk is not dead. It has been pushed out of the performance tier, but for bulk storage — backup staging, media archives, object storage tiers, video surveillance, scientific data — 20 TB and 24 TB helium-filled nearline drives at roughly $0.012 to $0.018 per GB are unmatched. For workloads measured in petabytes, HDD remains the cost-efficient choice and will be for the foreseeable future.

Layer 4 — Archive: tape and deep-archive object storage. LTO-9 tape at $0.004 per GB for the media and a few dollars per tape slot in a robotic library is still the cheapest way to store data you rarely read. Cloud deep-archive tiers (AWS Glacier Deep Archive, Azure Archive Blob) are in a similar cost range with different trade-offs — no capital expense, but retrieval times measured in hours and retrieval fees that can hurt if you actually need the data back.

Every workload belongs on one of these layers. The common mistake is putting data at the wrong layer — most often putting cold data on Layer 1 flash because that is where it was first created and nobody moved it, or putting hot data on Layer 3 HDD because that is what the storage array had capacity for.

What changed recently

Three things are worth knowing if you haven't updated your storage priors in the last couple of years.

QLC NAND made all-flash capacity affordable. The arrival of quad-level cell NAND from Solidigm, Samsung, Micron, and Kioxia pushed the cost per GB of enterprise SSD down to the point where the argument for performance HDDs in a production array is mostly gone. You can buy an all-flash array with usable capacity in the hundreds of terabytes for prices that would have required a hybrid design a few years ago. If you are planning a storage refresh in 2024 and the vendor is quoting you a hybrid (flash cache plus HDD capacity) design for general VM or file workloads, get a second quote on all-flash.

NVMe-oF matured and is shipping. NVMe over Fabrics — NVMe/TCP specifically, since it runs on existing Ethernet — lets an array share NVMe-class performance to compute hosts without the CPU overhead of iSCSI. Every major storage vendor now supports this in production. For high-performance database workloads that need sub-millisecond latency from an external array, NVMe/TCP has become the default choice and iSCSI is legacy.

Object storage is the new file share. For any workload that doesn't need POSIX semantics — backups, logs, media, data lakes, analytics — S3-compatible object storage has become the default destination. On-prem options (MinIO, Ceph, Pure FlashBlade, Dell ECS, NetApp StorageGrid) and cloud options (S3, Azure Blob, Wasabi, Backblaze B2) all speak the same protocol. This is a significant change from five years ago, when file and block storage were still the default answers for most use cases.

The questions to ask before buying storage

Storage decisions get made badly more often than any other infrastructure decision we see, because the comparison between products is genuinely hard and the vendors all sound similar. Here are the questions we actually use.

What is the true cost per usable GB, over three years, including all the software licenses? Raw drive cost is a small part of the story. Data reduction ratios (dedupe, compression) can make or break the economics. Software features — replication, snapshots, encryption, tiering — are often licensed separately. Support renewals in years two and three usually jump. Get the three-year TCO including everything, not just the year-one list price.

What does the performance look like under your actual workload? Synthetic benchmarks from the vendor are marketing artifacts. Ask for a proof of concept with your real data and your real workload pattern. Most serious storage vendors will do this for free on a loaner system. The ones that refuse are telling you something.

What happens when a component fails? Rebuild time, performance impact during rebuild, and failure modes beyond a single component all matter. An array that takes 24 hours to rebuild a drive while delivering 40 percent of its usual performance is not the same product as an array that rebuilds in an hour with no noticeable impact.

What is the upgrade and migration story? Storage gets bought for three to five years. What happens at year four when you need more capacity or need to move to the next generation? The best vendors now offer in-place upgrades or non-disruptive data-in-place migrations between generations. The worst still require forklift upgrades that take a weekend and a migration project.

Where does the data actually encrypt and who holds the keys? Self-encrypting drives are table stakes. Key management is where the difference shows up. If you are in a regulated industry, the answer to "who holds the encryption key" needs to be "we do, in our own KMS," not "the array vendor, we think."

What we recommend for common scenarios

A mid-market organization with a few hundred VMs and a few hundred terabytes of data: an all-flash storage array (Pure, Dell PowerStore, NetApp AFF, Nutanix) for the production VM estate, plus an object storage tier for backups and cold data. Skip the hybrid designs unless you have a specific reason.

An organization with petabyte-scale media or analytics data: object storage on nearline HDDs, either from a traditional vendor or built on MinIO or Ceph. Keep a small flash tier for working datasets and analytics acceleration. Do not store petabytes on a transactional all-flash array — you will pay three to five times as much as you need to.

A backup target: object storage with immutability support, either on-prem or in the cloud. Match it with a backup software product that understands object lock and can write immutable backup sets. This is the single most important architectural decision for ransomware resilience.

A deep archive of regulated or historical data: tape or cloud archive tier. The decision between them usually comes down to capital preference, retrieval patterns, and the comfort level with a cloud provider holding the data.

Three takeaways

1. All-flash is the default now for primary storage. QLC NAND closed the cost gap enough that hybrid designs rarely make sense for mid-market VM and file workloads. Nearline HDD still dominates bulk capacity.
2. Object storage is where most new data belongs. Backups, logs, media, analytics, and anything else that doesn't need POSIX semantics should go to S3-compatible storage by default.
3. Three-year TCO, not sticker price. Storage vendors live in year two and year three. Make sure your comparison does too.