Storage on KnightLi Blog

Why AI Data Centers Are Driving HDD Demand Again

Sat, 16 May 2026 21:02:33 +0800

Over the past two years, most AI infrastructure discussions have focused on GPUs, HBM, advanced packaging, and power supply. Behind training and inference systems, however, there is another bottleneck that is easier to overlook: storage.

A large model does not finish its work with a single computation inside a GPU. During training, it continuously produces checkpoints, optimizer states, training logs, dataset versions, and intermediate results. During inference, it also generates user interaction records, compliance archives, audit data, and system logs. These datasets do not always need to sit on the fastest media, but they often cannot be deleted immediately.

That is why hard drives are becoming important again.

AI Training Creates Massive Cold Data

Large model training needs to save checkpoints regularly. A checkpoint is essentially a saved state of the training process: if a training run crashes halfway through, the system can resume from a checkpoint instead of starting over.

For a large model, a single checkpoint can be several terabytes. A full training run may last weeks or even months, producing many checkpoints along the way. Even if some are later cleaned up, experiment replay, reproducibility, rollback, and model audits still require large amounts of data to be retained.

Training data itself is also expanding. High-quality text, images, videos, and code need to be cleaned, deduplicated, split, and versioned. As synthetic data, reinforcement learning data, and multimodal data become part of training pipelines, storage pressure will keep increasing.

This kind of data has several traits:

It is enormous in volume;
It is not always accessed frequently;
It needs long-term retention;
It is highly sensitive to cost per unit of capacity.

This data does not make sense to store entirely on expensive high-speed storage.

Why Not Use Only SSDs

SSDs are obviously faster, but data centers cannot optimize only for speed. For petabyte-scale cold data and anything beyond that, cost per unit of capacity directly determines whether the system is sustainable.

Storage in an AI cluster can be divided into several tiers:

HBM and GPU memory handle the hottest and most urgent data;
DRAM handles temporary movement and staging;
SSDs handle frequently accessed data with stronger low-latency requirements;
HDDs handle massive cold data, backups, logs, checkpoint archives, and long-term retention.

In other words, SSDs are important, but they cannot replace every tier. Truly large-scale systems usually need tiered storage: hot data prioritizes speed, while cold data prioritizes capacity, cost, and reliability.

As AI companies start retaining training residue, model versions, synthetic data, inference logs, and audit records for longer periods, the value of HDDs becomes more visible again.

Why HDD Supply Is Getting Tight

The hard drive market has not looked especially exciting for years, and consumer PCs have increasingly shifted to SSDs. Data centers follow a different demand logic.

Cloud providers and AI companies need high-capacity nearline drives with predictable delivery and low cost per terabyte. For hard drive vendors, these customers usually sign long-term supply agreements and receive higher priority than fragmented consumer channels.

That leads to several effects:

Capacity for high-capacity enterprise drives is locked in early by large customers.
Consumer hard drives and ordinary retail channels receive less supply.
New capacity takes time to come online, so short-term shortages are hard to fix quickly.
Hard drives move from low-attention hardware into part of AI infrastructure.

More importantly, the hard drive industry itself is already highly concentrated. There are only a few mainstream suppliers, and ramping production of advanced high-capacity drives is not as simple as building more factories. Technologies such as HAMR can increase capacity per drive, but moving from technical mass production to stable large-scale delivery still takes time.

Storage Price Increases Can Reach Consumers

AI data centers are not only absorbing GPUs and power. They can also affect the storage supply chain.

When more enterprise SSD, memory, and HDD capacity flows toward cloud providers and AI infrastructure, the consumer market may begin to feel price pressure. Higher retail prices for SSDs, memory, or hard drives are not always just retail volatility. They may come from upstream capacity being reallocated.

This effect is usually not linear. Large customers sign long-term agreements with more stable pricing, delivery, and capacity planning. Consumers are more exposed to spot-market fluctuations. The result is a familiar pattern: rising AI data center demand eventually makes storage devices more expensive for ordinary buyers too.

The Investment View Requires More Caution

AI-driven storage demand is real, but that does not mean every storage-related company will benefit over the long term.

Hard drives and flash memory still have cyclical characteristics. Rising prices, tight capacity, and long-term customer contracts can improve short-term performance. But once new capacity comes online or demand growth slows, the industry may return to supply-demand rebalancing. For hardware companies, the most important questions are not about one price increase, but whether demand can persist, margins can improve, capacity expansion becomes excessive, and the customer mix remains healthy.

A steadier interpretation is that AI is changing the demand structure of the storage industry. In the past, outsiders paid more attention to compute. Now more costs are shifting toward data retention, data governance, and model lifecycle management.

Conclusion

AI does not only consume compute. It also keeps producing data.

GPUs handle computation, HBM feeds data at high speed, SSDs support hot data access, and hard drives carry the enormous cold data base. As long as large model training, synthetic data, inference logs, and compliance retention continue to grow, data centers will need large amounts of low-cost, high-capacity storage media.

Hard drives may not look like the star hardware of the AI era, but they are becoming an indispensable layer of AI infrastructure. The more advanced the model, the more it depends on massive storage systems. The more expensive the compute, the more it needs reliable checkpoints and archives to protect the cost already invested.

Seagate Exos 2X14 Dual-Arm HDD: Cheap, Fast, Large, and Not for Everyone

Fri, 01 May 2026 11:05:45 +0800

The Seagate Exos 2X14 / Mach.2 is a rather unusual enterprise mechanical hard drive. Its key selling point is not just capacity, but its dual-arm design: a single 14TB drive appears to the system as two 7TB logical disks, with each side using its own head-arm assembly for parallel access.

That makes it look very tempting on the used market. When ordinary 14TB enterprise drives are being pushed up in price, these retired dual-arm drives can sometimes offer 14TB of capacity for less money while delivering sequential speeds close to SATA SSD territory. But this is not a normal NAS drive that everyone should buy as a drop-in replacement. Its strengths and risks are both obvious, and they need to be understood before buying.

Why It Can Be So Fast

A normal mechanical hard drive has one head-arm assembly, so reads and writes are mainly handled by one group of heads at a time. The Exos 2X14 / Mach.2 splits one physical drive into two 7TB logical units, allowing two head-arm groups to work at the same time.

Each 7TB logical disk can reach roughly 250MB/s in sequential reads and writes, similar to many large enterprise HDDs. If the two 7TB logical disks are combined as RAID 0, large sequential transfers can approach 500MB/s when there is no network bottleneck. That is close to mainstream SATA SSD performance and enough to saturate a 2.5G network.

However, this boost mainly applies to large sequential transfers. 4K small files, fragmented files, and random access do not suddenly become fast just because of the dual-arm design and RAID 0. It is still a mechanical hard drive at heart, so small-file workloads remain slow.

The Biggest Barrier Is Hardware Compatibility

Most low-priced units available on the used market are retired server drives with SAS interfaces, not SATA models. That means they are not suitable for most off-the-shelf NAS systems.

Consumer and prosumer NAS devices from brands such as UGREEN, Synology, and QNAP are usually designed around SATA backplanes, and SAS drives often cannot be used directly. DIY NAS users also need extra hardware:

An available PCIe slot on the motherboard
A SAS HBA or RAID card, such as an LSI 2008 flashed to IT mode
A server SAS backplane, or cables such as SFF-8087 to SFF-8482 adapters
An operating system that can correctly identify the two logical disks from the same physical drive

The controller card itself may not be expensive. The real hassle is the full chain of cabling, compatibility, cooling, power, and available PCIe space. If the machine is already tight on space, airflow, cables, or expansion slots, this drive adds real friction.

Do Not Treat the Two 7TB Disks as Backups of Each Other

The easiest mistake is treating the two 7TB logical disks shown by the system as two independent drives. They are not two physical disks; they are two logical units inside the same physical hard drive.

If these two 7TB logical disks are configured as RAID 1, it may look like a mirror, but the practical value is very limited. They share the same enclosure, controller board, power path, and part of the same mechanical environment. If the physical drive, board, or power path fails, the original data and the supposed backup can disappear together.

RAID 5 and RAID 6 also require caution. Traditional RAID 5 usually tolerates the failure of one physical disk, but when a dual-arm drive fails, the array may effectively lose two 7TB logical disks at the same time. If the array design does not account for this failure model, redundancy can be defeated very easily.

So this type of drive is better understood as one fast, large, risk-concentrated disk, not as two independent disks that can safely protect each other.

How to Use It in NAS Systems

If hardware passthrough and driver recognition work properly, systems such as FNOS, TrueNAS, Unraid, and Windows can usually see the two 7TB logical disks. The key question is not whether the system can see them, but how they should be used.

In FNOS, the two 7TB logical disks can be combined into a RAID 0 storage space to gain sequential throughput.

In TrueNAS, creating a storage pool may trigger a prompt about using disks with the same serial number. If you are sure you want to use this drive, you can allow it, place the two logical disks into the same vdev, and choose Stripe instead of Mirror. Mirror creates an unreliable illusion of redundancy here.

In Unraid, it is not recommended to put the two logical disks into the main array with a parity disk. A better approach is to create a separate cache pool or dedicated high-speed pool, use Btrfs or ZFS for an independent RAID 0, and assign it to temporary high-speed data, downloads, transcoding cache, or data that can be rebuilt.

Windows can also identify and use the disks, with similar logic: striping can improve sequential performance, but it should not be treated as a real two-drive backup plan.

Lifespan, Temperature, and Power Still Matter

The dual-arm design improves performance, but it also makes the mechanical structure more complex. These drives contain two head-arm assemblies and more head components, which naturally means more mechanical risk points. Since most units on the used market are retired enterprise drives, real power-on hours, previous workload, handling history, and transport condition are often unclear. Specification-sheet endurance numbers should not be trusted blindly.

Power and cooling also matter. Under full load, power consumption may exceed 11W, and shell temperatures above 40 degrees Celsius are not unusual. If the drive is installed in a small case, dense drive cage, or low-airflow NAS, sustained heat can further affect stability and lifespan.

Who Should Buy It

This drive is suitable for people who are comfortable building a DIY NAS, already have the right SAS hardware, understand the risks of RAID 0, and mainly handle large sequential files. Media libraries, download caches, temporary project storage, and rebuildable data pools can all benefit from its capacity and speed.

It is not suitable for:

Users of off-the-shelf branded NAS systems
Users who do not want to buy a SAS card and adapter cables
Primary storage with high data-safety requirements
Users who want to use the two 7TB logical disks as a RAID 1 backup
Workloads dominated by small files, virtual-machine random I/O, or databases

If the goal is low-cost capacity and fast sequential reads and writes, the Exos 2X14 / Mach.2 is genuinely interesting. But its proper role is a cheap, fast, tinkerer-friendly special-purpose drive, not a universal NAS disk that everyone can buy without thinking.

Hard drives have prices; data has value. This kind of dual-arm drive can be worth buying, but it is best used only for recoverable, rebuildable, or separately backed-up data. Truly important files should still live inside a clear and reliable backup strategy.

A Practical Guide to Common U.2 Enterprise SSD Series

Wed, 15 Apr 2026 22:19:10 +0800

If you are shopping for used servers, enterprise SSDs, or planning to add U.2 NVMe drives to a workstation, NAS, or storage node, you quickly run into model names like P5510, P5620, PM9A3, SN640, 7450 PRO, and CD8.

Those names are not very intuitive on their own. Just from the number, it is often hard to tell whether a drive is aimed at high performance, high endurance, large capacity, read-heavy workloads, or mixed workloads. This article organizes several common U.2 series by vendor and positioning, so it is easier to build a practical mental map before comparing exact capacity, endurance, and pricing.

One thing to note up front: each series usually includes multiple capacity points, firmware variants, interface generations, and endurance tiers. The goal here is to explain the general role of each series, not to serve as a full SKU-by-SKU spec sheet.

01 A Quick Way to Think About U.2 Drives

Enterprise U.2 SSDs can be roughly grouped into a few categories:

General-purpose models: suitable for most servers and virtualization workloads, with balanced read/write behavior.
Read-optimized models: better for read-heavy databases, object storage, content delivery, and cache layers.
Mixed-workload models: better for databases, logs, and virtualized environments with meaningful write pressure.
High-endurance models: better when write volume is high and low latency matters.
High-capacity QLC models: better when cost per TB matters more than sustained heavy writes.

For individuals and small teams, the most common mistake is not buying something “too slow,” but buying the wrong class of drive. A large-capacity QLC drive is not ideal for heavy write workloads, and a high-endurance Optane drive is usually overkill for pure archival storage.

02 Solidigm / Intel

Solidigm inherited Intel’s NAND SSD business, so these lines are often discussed together.

D7-P5510 / P5620

These are both very typical PCIe 4.0 datacenter NVMe series and are common in general servers, virtualization platforms, and enterprise storage nodes.

D7-P5510 is generally seen as a more general-purpose or read-oriented option.
P5620 is usually treated as a more write-capable mixed-workload tier with higher endurance.

If you want a drive family that works well across a wide range of enterprise use cases, these are usually safe options. Their appeal is not that one specific metric is extreme, but that they are balanced, stable, and widely available.

D5-P5316

D5-P5316 stands out because it follows a high-capacity QLC approach.

Its main attraction is not extreme write performance, but density and cost per TB. For object storage, colder data, large read-heavy datasets, or environments where rack-level capacity matters a lot, this type of drive is very attractive.

Its limits are also clear: it is not a great choice for sustained heavy writes, demanding random-write workloads, or frequent rewriting. It is better understood as a high-density capacity drive than a high-endurance performance drive.

Optane DC P4800X

P4800X belongs in a completely different category. It is not a normal NAND product, but part of Intel’s Optane / 3D XPoint line.

This type of drive is usually known for:

Very low latency
Excellent small-block random performance
Extremely high write endurance

If your workload is heavy logging, metadata-intensive work, low-latency databases, cache layers, or very high write pressure, Optane behaves very differently from ordinary NAND SSDs. The downside is just as obvious: capacity is usually limited and pricing is not friendly. Today, many people think of it as a special-purpose “legendary” drive rather than a normal high-capacity enterprise SSD.

03 Samsung

Samsung enterprise NVMe drives are also common in servers and OEM platforms, especially in branded systems and cloud deployments.

PM9A3

PM9A3 is a common PCIe 4.0 enterprise series with a mainstream datacenter positioning. It is often compared to drives such as the P5510.

It is a good fit for:

General servers
Virtualization hosts
Balanced enterprise workloads

If you want an enterprise U.2 drive that is modern enough, performs well, and is relatively easy to find, PM9A3 is usually a strong candidate.

983 DCT

983 DCT is older, but many people still remember it because it was widespread in earlier enterprise platforms.

Its appeal today is maturity, broad market availability, and often better pricing. It works well in budget-sensitive environments where you still want a known and well-supported enterprise model. It feels more like a reliable veteran than a first choice for the newest platform.

PM1733 / PM1735

PM1733 and PM1735 are also representative Samsung enterprise NVMe families on the higher-performance side.

They are often associated with:

Strong sequential performance
Higher-end datacenter positioning
Better fit for bandwidth-heavy and high-IOPS workloads

If your host platform is already PCIe 4.0 and you care about databases, virtualization, compute nodes, or high-throughput storage, PM1733/PM1735 are often more attractive than entry-level or older enterprise SSDs.

04 Western Digital / HGST

Enterprise SSDs from the Western Digital / HGST ecosystem are also common, especially under the Ultrastar name.

Ultrastar SN640

SN640 is usually regarded as a read-optimized NVMe SSD.

It fits well in scenarios such as:

Content delivery
Read-heavy cloud storage
Boot or image drives
Read-heavy database replicas

The appeal of this class is usually the balance of capacity, power, and read-focused value. If your workload is mainly read-heavy, it can be more economical than a higher-endurance mixed-workload drive.

Ultrastar SN840

SN840 is generally understood as a higher-performance, higher-tier datacenter NVMe line.

If SN640 leans more toward read optimization, SN840 feels more like a performance-oriented enterprise NVMe option for heavier applications, virtualization, and data-platform workloads. For people targeting stronger platform capability, it is often worth a look, though price and availability may be less friendly.

05 Micron

Micron enterprise SSDs have also become very visible in the server market, and their product positioning is often relatively easy to understand.

7450 PRO / MAX

The 7450 family is a good example because the naming is direct:

7450 PRO: more mainstream enterprise usage, often suitable for general-purpose and read-oriented environments.
7450 MAX: more suitable for high-endurance and heavier-write scenarios.

That split is easy to understand. If you are deploying general servers, virtualization, or application infrastructure, PRO is often enough. If you are targeting databases, logs, or sustained write-heavy use, MAX is the better fit.

9400 Series

The 9400 family usually sits in a newer, stronger enterprise NVMe tier, aimed at higher throughput, higher IOPS, and more demanding server workloads.

If your goal is a newer platform with stronger performance and heavier business workloads, the 9400 line is often more attractive than the 7450. If you are only building a normal storage node or a home lab, it may not be the most cost-effective choice.

06 Kioxia

Kioxia is also a very common enterprise SSD vendor, especially in OEM servers, branded systems, and enterprise procurement channels.

CD6

CD6 is a typical PCIe 4.0 datacenter NVMe line with a mainstream enterprise positioning.

It is well suited to:

General servers
Cloud nodes
Enterprise application deployment
Balanced mixed workloads

If you want a drive family that is not overly specialized and tends to behave predictably in enterprise environments, CD6 is a reasonable candidate.

CD8

CD8 is generally viewed as a newer and higher-tier line with stronger platform specs and performance expectations.

If your focus is newer infrastructure, stronger performance targets, and a more modern datacenter setup, CD8 is usually more interesting than CD6. The tradeoff is that pricing is often higher as well.

07 A Fast Way to Narrow Down the Choice

If you simply want a quick starting point, this is a useful way to think about them:

General-purpose and safe choices: P5510, PM9A3, CD6
Mixed workloads and higher endurance: P5620, 7450 MAX
High capacity and lower cost per TB: D5-P5316
Ultra-low latency and extremely high endurance: Optane P4800X
Higher-performance modern datacenter drives: PM1733/PM1735, SN840, 9400, CD8
Mature older model with value pricing: 983 DCT

This is not a strict spec-sheet summary. It is better used as a practical orientation map.

08 Short Advice

If you are buying U.2 enterprise SSDs for a NAS, lab platform, or virtualization host, the three things to confirm first are:

Whether your backplane, cable, HBA, or motherboard actually supports U.2 NVMe.
Whether your workload cares more about capacity, endurance, or low latency.
Whether the drive is an OEM firmware variant that may affect compatibility or upgrades later.

The model absolutely matters, but so do interface support, cooling, power, and platform compatibility. If you understand the role of the series first, choosing the right capacity and price point becomes much easier.