The Future of HDDs: Balancing Skyrocketing Capacity with Limited IOPS

Future of HDDs: IOPS Limitations, Sequential Access, and the Decline of Tape

As data growth accelerates, hard disk drives (HDDs) are reaching unprecedented capacities – but their input/output operations per second (IOPS) have remained nearly flat. This growing imbalance poses a critical question for storage planners: Can HDDs continue to scale effectively, or will new access patterns and technologies be required to bridge the performance gap? In this article, we explore how sequential access methods (like those used in Host-Managed Shingled Magnetic Recording, HM-SMR) might become mainstream to cope with the IOPS-to-capacity challenge. We also examine whether massive HDDs could encroach on tape’s archival role in an era of more active data retrieval, and how HDDs stack up against flash storage in performance and cost for various tiers.

The IOPS-to-Capacity Dilemma

Modern 3.5-inch nearline HDDs have exploded in capacity – from hundreds of gigabytes in the mid-2000s to 20 TB, 30 TB, and beyond by the mid-2020s. However, the mechanical nature of HDDs means their random IOPS have improved only marginally. A 7,200 RPM disk still delivers on the order of 100–200 random IOPS, similar to its predecessors a decade ago. As capacities push past 10 TB and 20 TB, the IOPS-per-terabyte “density” plummets to single digits. In fact, once drives exceed 20 TB, they provide fewer than 10 IOPS per TB, suitable only for the coldest data.

This widening gulf between capacity and performance creates a bottleneck for applications. Even workloads traditionally thought of as sequential – like video streaming – can overwhelm large drives when many users access different files concurrently. For example, a video-on-demand server might serve thousands of streams; a single HDD must constantly seek between Frozen, The Mandalorian, or whatever content is being watched. In effect, what should be sequential turns into a random workload at scale, stressing the drive’s limited IOPS. The result is that conventional random-access use of ultra-large HDDs can lead to unacceptable latency or require enormous DRAM/SSD caches to buffer hot data, which is costly and complex (The Diminishing Performance of Disk-Based Storage | VAST Data).

Sequential Access Patterns as a Solution (SMR and Beyond)

One way to mitigate the IOPS shortfall is to change how data is written and read on these high-capacity disks. Enter sequential access patterns – writing data in long, continuous chunks – which maximize the HDD’s strength (high sequential throughput) and minimize expensive random head movements. Technologies like Shingled Magnetic Recording (SMR) embody this approach. In SMR, tracks on the platter are written to overlap like shingles, squeezing ~15–20% more data per platter by using the narrow read head width as the limiting factor rather than the wider write head (30TB HAMR disk drives get less shingled capacity increase than you’d think – Blocks and Files). The trade-off is that new data must be written sequentially on top of overlapping tracks, since random writes would overwrite adjacent data.

 Host-Managed SMR (HM-SMR) drives take this concept further by exposing zones that require sequential writes. The host system is responsible for organizing data into sequential streams and managing when to rewrite “shingled” zones, rather than the drive doing it internally. This avoids the unpredictable performance of device-managed SMR, at the cost of additional complexity in the software stack. Once the host is SMR-aware and writes only sequentially, the performance can be made consistent, often comparable to conventional drives for reads, and sequential write throughput close to normal HDD speeds.

Major cloud companies have already embraced this model. For instance, Dropbox engineered its storage system to use host-managed SMR drives at scale. They write data in large sequential chunks aligned to SMR zones, and after the initial write, most accesses are random reads – which SMR handles well . Deletes and rewrites are batched and managed by the host in the background to reclaim space methodically. This allowed Dropbox to deploy 14 TB SMR drives (Western Digital Ultrastar Hs14) across their infrastructure, gaining ~40% more capacity per drive for lower cost/TB. As the WD team noted, SMR capacity isn’t “free” – it demands investment in software – but at cloud scale this pays off in significant total cost of ownership (TCO) savings (Making Host Managed SMR Work for You – Dropbox’s Successful Journey – Western Digital Corporate Blog).

Going forward, such zoned sequential-write approaches may become the norm for the largest drives. Industry experts predict that as data “deluge” continues and businesses make more of their archival data active, SMR (and similar techniques) will be pervasive in order to maximize capacity economically. In practice this means future high-capacity HDD deployments could resemble something between today’s disks and tape – written in append-only logs or large sequential objects, with random writes buffered or rerouted to faster tiers. Filesystems and object stores are evolving (with technologies like Zoned Storage APIs) to facilitate this by design. The upside is that sequential access patterns let us fully utilize multi-terabyte drives’ space without being bottlenecked by ~150 IOPS limits, as long as workloads can be buffered and made sequential. The downside is inflexibility – not every application can easily adopt a write-once, mostly-read-many pattern.

Advances in HDD Technology: Vendor Roadmaps and Multi-Actuators

Aside from changing access patterns, HDD vendors are attacking the problem from technological angles. Higher areal density remains a core driver of capacity: both Seagate Technology  and WD (WD) have clear roadmaps to 30, 50, even 100+ TB drives by the end of the decade. Seagate’s recently published strategy outlines ~50 TB HDDs by 2026 and 120 TB or more by 2030, using technologies like Heat-Assisted Magnetic Recording (HAMR) and even Bit-Patterned Media in the long term. HAMR drives use a tiny laser to momentarily heat the disk surface while writing, enabling much tighter bit packing (Seagate’s current 20 TB HAMR drives are a first step, with ~90 TB per drive on the horizon) (Seagate’s Roadmap: The Path to 120 TB Hard Drives). Western Digital, after initially pursuing microwave-assisted recording, is also planning to introduce HAMR-based drives by 2026 (starting around 36 TB for conventional recording, 44 TB with SMR) and foresees 80 TB CMR and 100 TB SMR drives around 2030 (Western Digital to unveil 44TB HAMR HDDs in 2026, 100TB in 2030 | Tom’s Hardware). Toshiba is similarly leveraging energy-assisted recording (Flux-Control MAMR) in its 20 TB drives and will likely adopt next-gen techniques to keep pace.

While these technologies boost capacity, they don’t inherently improve random I/O performance – an area where vendors are experimenting with other ideas. One such approach is multi-actuator drives. By adding a second independent actuator arm inside a drive (as in Seagate’s “MACH.2” technology), an HDD can perform two separate reads/writes in parallel . A dual-actuator 20 TB disk appears to the host as two logical 10 TB drives, roughly doubling IOPS and throughput. This is a welcome improvement, but it’s only a one-time gain – e.g. from ~150 IOPS to ~300 IOPS – which still leaves large drives orders of magnitude behind solid-state drives. Seagate noted that even a 20 TB dual-actuator HDD delivers <400 IOPS, less than 1/1000th the IOPS of an equivalently sized QLC flash SSD. Multi-actuator mechanisms also add cost and complexity, so we may see them only in niche high-performance capacity drives if at all widely.

Another innovation is marrying flash with disk internally. Western Digital’s OptiNAND is one example – it incorporates a small NAND flash buffer in the HDD to store metadata and absorb bursts of writes, smoothing out performance especially on SMR drives. This doesn’t increase sustained random IOPS much, but it can reduce worst-case latencies and improve reliability (caching write data in case of power loss, for instance). Techniques like track-at-once caching, write coalescing, and better firmware algorithms are being used to make huge drives more performant without host changes. Still, none of these fully escape the physics of spinning media – which is why the sequential, smarter usage of HDDs is becoming so important.

Importantly, the big three HDD manufacturers are confident that these innovations will keep HDDs relevant. Seagate’s CTO emphasized that HDDs will continue to deliver the lowest TCO for mass storage in the next decade, even as capacities push to 50 TB and beyond (Seagate’s Roadmap: The Path to 120 TB Hard Drives). Western Digital likewise touts its 32 TB Ultrastar DC HC690 (an 11-platter SMR drive) and 26 TB CMR drive, introduced in 2024, as offering superior TCO for hyperscale and enterprise customers who need petabytes at low cost (Western Digital Delivers World’s Highest Capacity ePMR HDDs to Meet Growing Nearline Demand | Western Digital). These latest drives combine multiple advances – energy-assisted recording (ePMR), more platters in the same 3.5″ enclosure, OptiNAND, and improved servo controls – to keep increasing capacity without increasing the drive’s physical size or power draw (Innovating to 11: Western Digital Increases HDD Capacity, Not Size  – Western Digital Corporate Blog). In short, manufacturers are squeezing every drop out of HDD technology: more bits, clever caching, maybe two heads working at once, and expecting the users of these drives to adapt by handling data in large sequential chunks when possible.

HDDs vs. Tape: Will Disks Become the New Archive Medium?

In parallel to HDD improvements, a debate rages over the role of magnetic tape in modern data architectures. Tape storage has long been the go-to medium for cold archives and backups due to its extremely low cost per TB and power-off storage capabilities. Today’s LTO tape cartridges cost only a few dollars per terabyte and can sit on a shelf using no energy, which is far cheaper and “greener” than keeping hard drives spinning in an archive. Tape also provides an inherent air-gap for security – data offline on a cartridge is safe from online ransomware attacks. These advantages ensure tape is not dead: the LTO roadmap now stretches to LTO-14 (planned ~ in the 2030s) with up to 1.4 PB of compressed data per cartridge. In 2023, LTO-9 (18 TB native) and upcoming LTO-10 (36 TB native, ~90 TB compressed) keep doubling capacity every few years. Tape technology has also improved retrieval times via the Linear Tape File System (LTFS), which indexes tape content for faster seek and file access, and new drive techniques that boost streaming speeds (What the new LTO roadmap means for tape storage | TechTarget). In other words, tape still offers unrivaled density and security for cold data.

However, the cons of tape are equally important in a world of more active data use. Tape is a sequential medium – accessing arbitrary data can be orders of magnitude slower than disk. A cartridge may need to be loaded and wound to the correct position, introducing minutes of latency before data begins streaming. Even with LTFS, random small reads from tape are impractical; tape truly excels at large, linear reads or write-once, read-rarely scenarios. This is where modern “active archives” are challenging tape. Many organizations find that archived data is no longer write-and-forget: it might be mined for analytics, used to train AI models, or re-used in new services. For example, big science projects or media companies might dip into years-old data repeatedly (multiple reads) to feed new analyses. Similarly, cloud providers offering “Cold” storage tiers are noticing customers still expect reasonably timely access for restores or data lake queries, not multi-hour delays.

Because of these trends, HDDs have started to encroach on tape for large archives that require more frequent reads. Hard drives, especially if used with SMR (meaning they are written sequentially like tape but can be read randomly), present an attractive middle ground: they are slower and cheaper than primary storage, yet provide online access (milliseconds latency, not minutes) and easier concurrent access than tape. With the cost per TB of huge HDDs dropping and approaches like powering down idle drives (MAID arrays) to save energy, some storage architectures use disks to implement “active archive” tiers. In such a design, tape might still be used as a true deep backup (for disaster recovery or regulatory retention), but the “everyday archive” lives on HDD-based object storage for faster retrieval.

For archival storage, magnetic tape has long been a formidable competitor to HDDs. Tape cartridges (such as the LTO series) offer extremely low cost per terabyte and have the advantage of consuming no energy when shelved. Historically, HDD-based archives struggled to match tape’s idle power efficiency, because spinning disks would still draw power and generate heat even when not actively in use. Around the mid-2000s, the industry introduced the concept of Massive Array of Idle Drives (MAID) to narrow this gap. MAID systems would power down most drives and spin them up only when needed, reducing power and cooling demands. However, early MAID implementations faced challenges: without an intelligent distributed file system, keeping data available with drives mostly off proved difficult. Redundancy requirements often meant drives had to spin up frequently to serve reads or rebuild data, undercutting the intended energy savings.

In recent years, we’ve seen a modern revival of the MAID idea in the form of Infinite  Cold Engine (ICE) technology from Leil Storage. Introduced in 2024, ICE is effectively “MAID 2.0” — it uses smarter software and drive management to achieve far greater energy efficiency without sacrificing performance. ICE builds on a distributed file system called SaunaFS , which allows the storage system to concentrate read/write activity on a small subset of drives at any given time. By doing so, the majority of drives can remain completely powered down for extended periods. This approach greatly cuts energy usage while still ensuring data is quickly accessible when needed. In essence, ICE treats a large HDD array a bit like a tape library: most media stay inactive until called upon. Yet unlike tape, when data is requested, the spin-up latency of a hard drive is much lower than the mount/load time of a tape cartridge, yielding faster access.

ICE also incorporates intelligent performance rebalancing.

Drives are organized into write groups, and the system dynamically rotates which group is active (“hot”) so that no single set of disks bears all the workload permanently. This rotation extends drive life and spreads out wear, all while keeping response times within a predictable range. Thanks to such smart orchestration, Leilreported that their ICE-enabled system can save up to 43% in energy consumption compared to traditional always-on HDD arrays. The concept has garnered attention in the storage industry—Leil’s ICE was unveiled at the SNIA Storage Developer Conference in 2024, indicating that experts are actively exploring energy-efficient HDD solutions in enterprise settings. The positive reception of ICE underscores a key trend: reducing operational costs and environmental impact has become just as important as raw capacity in future storage solutions.

The upshot is that HDDs are closing the gap with tape in cold storage scenarios. With technologies like ICE, massive HDD archives can be nearly as energy-frugal as tape libraries, while offering much faster random access to data. That said, tape is not standing still—new tape generations (LTO-9 and beyond) keep increasing cartridge capacity (on the order of dozens of terabytes) and remain unmatched for absolute storage cost in deep archive use cases. It’s likely that both HDDs and tape will coexist in large-scale archival systems, each playing to its strengths: tape providing economical deep storage for infrequently accessed data, and HDDs (especially with intelligent power management) providing a more readily accessible archive layer.

HDDs vs. Flash: Performance and Cost at Scale

With flash memory costs falling over time, another common question is whether flash-based solid state drives (SSDs) will eventually supersede HDDs even for bulk storage. In high-performance tiers, this has largely happened – faster 10k and 15k RPM enterprise disk drives gave way to SSDs years ago for transactional databases, virtual machines, and other IOPS-intensive workloads. But for capacity-oriented storage, flash has not yet dethroned spinning disks. The reasons come down to cost and endurance.

On pure performance, SSDs win by orders of magnitude. A single SATA or NVMe SSD can deliver tens of thousands to millions of IOPS, compared to a few hundred from an HDD. Even the slowest, high-density QLC NAND drives can outperform multi-actuator HDDs by a factor of 1000× in random I/O. Flash also has near-zero seek latency, which simplifies data management – no need for sequential hacks like SMR to get throughput. In environments that demand high IOPS or low latency at scale (for example, web-scale databases or real-time analytics), architects increasingly opt for all-flash arrays or flash-based object stores despite the cost, because the performance benefits and simpler scaling justify it.

However, flash memory is still significantly more expensive per TB than disk (and tape). As of 2025, high-capacity enterprise HDDs might be roughly $15–$20 per TB for large buyers, whereas high-capacity SSDs (using QLC, the cheapest NAND) could be several times that – and even more once you factor in redundancy and wear provisioning. HDD vendors argue that this cost gap will persist. Seagate and WD have publicly stated that HDDs will maintain a TCO advantage for at least the rest of this decade for bulk storage (Seagate’s Roadmap: The Path to 120 TB Hard Drives). One reason is that while NAND flash manufacturing is improving (more layers, bits per cell, etc.), HDD capacity is also climbing, so both mediums are moving targets. A 20 TB HDD today might have an SSD equivalent at 15 TB, but by the time a 30 TB SSD is feasible, HDDs might be 60 TB. Simply put, flash isn’t catching up to HDD on capacity fast enough to eliminate the cost gap. Furthermore, extremely high-density flash (e.g. QLC and upcoming PLC with 5 bits per cell) suffers from lower endurance. That makes it less ideal for “cold” archival use where data might sit for years – ironically, HDDs and tapes have better archival retention characteristics without power. To mitigate this, vendors use techniques like error-correcting codes and overprovisioning, but that adds to complexity and cost.

That’s not to say flash won’t continue eating into HDD territory. Hybrid tiered designs are common, where frequently accessed data resides on SSDs while HDDs store the bulk of colder data. As the price of NAND drops and technologies like computational storage and high-capacity NAND modules (e.g. 3D NAND drives in “ruler” form factors) mature, some secondary storage workloads may go all-flash. In fact, all-flash object stores and backup repositories are emerging, often using compression/deduplication to offset cost. Companies like VAST Data have even claimed that with modern data reduction and cheap QLC flash, one can achieve an all-flash system with cost comparable to HDD-based systems, but with far superior performance. These claims usually assume certain data compressibility and very large scale – so results vary. In practice, HDDs remain the workhorse for capacity at scale in most enterprises and clouds, especially where access patterns are mostly sequential or read-heavy (which alleviates the IOPS issue). Flash is the go-to for speed, but at petabyte scales, the budget usually dictates a tiering approach rather than pure flash.

Conclusion and Future Outlook

HDDs are at a crossroads. On one hand, they are becoming massive libraries capable of storing tens of terabytes on a single device, advancing toward 50–100 TB with evolving recording technologies. On the other hand, their access performance (especially for random I/O) can’t keep up with these capacity gains, forcing a paradigm shift in how we use them. The industry response is twofold: make the drives bigger and a bit smarter (through HAMR, SMR, multi-actuators, etc.), and concurrently adapt software and data management to work with the drives’ sequential nature rather than against it. This means that future HDD deployments will increasingly treat hard drives not as random-access blocks, but as serial data streams – much like tapes, but with faster electromechanical access. Approaches like host-managed SMR, log-structured filesystems, and large object storage workloads align well with this future. We may also see more integration of HDDs into tiered systems where their weaknesses are shielded by front-end flash caches or by only using them for data that doesn’t need high IOPS.

In considering the broader storage landscape, it’s clear each medium – flash, disk, tape – will continue to play a role aligned to its strengths. Tape will remain the king of deep archive and offline security. Flash will keep pushing into larger capacities for hot data and may one day challenge HDDs even in the capacity tier if economics allow. HDDs, meanwhile, will occupy the “middle”: the bulk active data that is too large for flash economically, but that needs more accessibility than tape can provide. In that niche, hard drives can thrive if used optimally. In fact, HDD manufacturers are positioning their drives as crucial enablers for trends like big data and AI, which generate and consume enormous datasets. HDDs handle the “feast of data” on the input side (collecting and storing it cheaply) and on the output side (preserving vast trained models and content) (Western Digital Delivers World’s Highest Capacity ePMR HDDs to Meet Growing Nearline Demand | Western Digital). This suggests a healthy demand for high-capacity drives well into the future, as long as the cost per TB stays compelling.

Looking ahead 5–10 years, expect sequential-friendly HDD usage to become commonplace, multi-actuator and other enhancements to modestly improve performance, and drive capacities doubling or tripling. We’ll likely see 30+ TB drives standard in data centers, with specialized 50+ TB SMR units for those who can use them. Software ecosystems (from databases to Hadoop-like analytics platforms) will increasingly optimize for large block sequential I/O when targeting spinning disks, essentially treating HDDs as “slow, big pipes” rather than random-access disks. And when data truly needs random fast access, it will reside on SSD. In essence, the future of HDDs is one of specialization: they will not be general-purpose storage for all workloads like in the past, but they will excel in the capacity-centric roles they’re called to fill.

In summary, hard drives aren’t going away – but they are evolving. Sequential access patterns, tiered architectures, and continued innovation from drive makers are ensuring that HDDs can scale in the age of zettabytes. They may gradually take on characteristics of the tapes they aspire to replace (sequential writes, very high density), even as flash bites at their heels for performance. For storage professionals and IT planners, the key is to leverage each medium for what it does best. HDDs remain a cost-effective foundation for bulk storage, so long as we design systems to mitigate their limitations. By acknowledging the capacity-IOPS imbalance and architecting around it, we can continue to benefit from the incredible capacity growth of hard disks – without letting performance fall off a cliff as the terabytes soar.

Author: David Gerstein
David Gerstein is the CTO of Leil, an Estonia-based storage infrastructure company building HDD-native software for the exabyte era. Leil is in active co-engineering partnership with Western Digital, validated across the WD, Seagate, and Toshiba drive ecosystems.