Enterprise storage pricing is once again under pressure as artificial intelligence and data centers drive up demand for high-capacity SSDs. Against that backdrop, buying ever-faster NVMe drives doesn’t guarantee you’ll actually get their full performance: the RAID controller or the CPU resources needed to manage storage can become the real bottleneck in the system.
NVMe performance in 20 seconds
- AI demand keeps pushing up prices and shipments of enterprise SSDs.
- Traditional RAID controllers can become a bottleneck against large pools of NVMe drives.
- Software RAID avoids that limit, but it uses server CPU.
- Graid proposes offloading part of that work to a dedicated GPU.
- The performance figures published by the company should be understood as results from its own testing.
The issue is especially relevant when sizing new servers. An enterprise PCIe 4.0 SSD can move several GB/s and hundreds of thousands, or even more than a million, input/output operations per second (IOPS). Once you install 16, 24, or 32 drives in a single system, the potential aggregate performance grows far faster than what SATA and SAS generations — which many conventional RAID architectures were designed around — ever offered.
On top of that, the cost of getting it wrong is higher. TrendForce notes that server storage demand remains supported by investment in AI infrastructure and data centers. For the third quarter of 2026, it expects NAND Flash contract prices to rise between 10% and 15% versus the previous quarter, though at a slower pace than in prior quarters. Enterprise SSDs remain among the segments with the strongest demand.
That’s why it’s worth examining what happens between the SSDs and the applications before simply adding more drives.
RAID Built for Hard Drives Meets Dozens of NVMe Drives
For years, a hardware RAID controller solved two problems fairly comfortably.
The server handed storage operations off to a dedicated card. Its processor calculated the parity needed for RAID 5 or RAID 6, managed the disks, and kept much of that work off the main CPU.
That architecture made sense when each drive delivered a few hundred MB/s.
NVMe has changed the proportions.
A single modern NVMe SSD can use several PCIe lanes and reach several GB/s. Group 24 drives together, and the theoretical aggregate throughput can comfortably exceed 100 GB/s, depending on the SSDs and the platform.
The controller sitting in front of them then has to be able to process that volume of traffic.
Graid Technology calls the performance or compute capacity lost between what SSDs can physically deliver and what applications actually receive a “storage tax.” The term comes from the company itself and should be understood as a marketing way of describing a real technical problem: any intermediate layer can turn into a bottleneck once storage performance climbs high enough.
According to figures published by Graid, certain NVMe configurations sitting behind conventional RAID controllers can end up delivering only 12% to 18% of the drives’ theoretical aggregate speed.
That’s not a figure that can be generalized to every hardware RAID setup. The result depends on the controller, the PCIe interface, the number and model of SSDs, the RAID level, block size, access pattern, queue depth, and server architecture.
But the underlying problem is real: if 24 SSDs can collectively deliver more traffic than the controller can handle, installing faster drives stops producing a proportional improvement.
Software RAID Moves the Bottleneck Elsewhere
One alternative is dropping the traditional RAID controller and managing storage through software instead.
Linux has supported this for years through technologies like mdadm, alongside distributed and software-defined storage systems that take different approaches.
The advantage is that there’s no longer a central controller that every operation necessarily has to pass through.
But the work doesn’t disappear.
It shifts to the server’s CPUs.
The cost isn’t just about calculating XOR operations to generate parity, either. In systems capable of processing millions of IOPS, queues, interrupts, completions, data movement, and read-modify-write operations all come into play.
A small RAID 5 write, for example, can require several physical operations to update the data and its corresponding parity.
Graid says it has measured, in real customer workloads, CPU usage equivalent to between 18% and 28% of AMD EPYC processor cores due to the I/O processing associated with software RAID.
Again, that’s a measurement from the company itself, not a rule that applies to every server. CPU consumption can vary considerably depending on configuration and workload.
The economic angle, though, is easy enough to grasp.
On a virtualization, database, or AI inference server, cores used to manage storage are cores no longer available for applications.
A GPU Dedicated to RAID as a Third Option
Graid proposes solving this dilemma with SupremeRAID, an architecture that uses a NVIDIA GPU installed via PCIe to offload part of the RAID processing.
The idea sounds unconventional, since GPUs are normally associated with graphics, artificial intelligence, or scientific computing. But their ability to run many operations in parallel can also be applied to certain storage-related tasks.
NVMe SSDs stay connected directly to the server’s PCIe platform. The GPU doesn’t simply act as the traditional physical path all data must cross; it works as the processing element for RAID operations.
Graid currently offers several SupremeRAID variants. Its technical documentation shows models based on the NVIDIA A2000, T1000, and T400 GPUs, supporting PCIe generation 3 through 5 NVMe drives and configurations of up to 32 physical drives, depending on the license and platform.
The goal is to keep the advantages of software-defined RAID while cutting down the work the CPU has to take on.
According to measurements the vendor has published, this architecture can recover the percentage of CPU that certain software RAID workloads would otherwise consume and bring storage closer to the aggregate performance available from the NVMe drives.
It’s a proposal that should be evaluated with benchmarks of the actual workload before making an investment. OLTP databases, virtual machines, sequential storage, model training, or inference can all generate very different I/O patterns.
Adding a dedicated GPU isn’t free, either. You have to factor in the card and license costs, power consumption, PCIe slot usage, support, compatibility, and operational complexity.
SSD Pricing Changes the Math
This debate arrives at a particularly interesting moment for the market.
TrendForce notes that the expansion of AI infrastructure is sustaining strong demand for high-capacity enterprise SSDs. Manufacturers are shifting more capacity toward server and data center products, while the consumer market shows a much weaker trend.
The firm’s latest data also shows the enterprise SSD market hit record revenue during the second quarter of 2026, boosted by both higher shipments and higher prices.
That helps put one of the more eye-catching claims used to promote the “storage tax” concept into perspective.
There isn’t enough evidence to claim, as a general rule, that all enterprise SSDs have doubled in price over the past year. The pattern varies by capacity, NAND technology, contract, manufacturer, and segment.
What current data does show is clear pressure on NAND and enterprise storage prices, paired with strong demand from servers and AI infrastructure. TrendForce believes the NAND market has stayed tight through 2026 and expects those constraints to start easing gradually during the second half of 2027.
In that environment, a storage platform’s efficiency carries more economic weight.
If a company buys 24 high-performance SSDs but another part of the system stops it from using them fully, adding more drives can be an expensive way to try to fix the wrong problem.
The review should start earlier: available PCIe lanes, NUMA topology, bandwidth between CPU and devices, controller, RAID level, block size, CPU used by storage, and application behavior.
That analysis also avoids the opposite mistake. A company that only needs 10 GB/s doesn’t necessarily gain an economic benefit just because its storage can deliver 100 GB/s.
The goal shouldn’t be hitting the theoretical maximum of the SSDs, but rather avoiding paying for performance the application needs that some intermediate component can’t actually deliver.
That’s where the genuinely interesting part of the so-called “storage tax” lies. More than a specific technology, it raises a question worth asking before renewing an NVMe platform: if SSDs keep getting faster and pricier, what percentage of that investment can the server actually use?
Frequently Asked Questions
Can a RAID controller limit the performance of multiple NVMe SSDs?
Yes. If the SSDs’ aggregate performance exceeds the controller’s capacity or that of its PCIe connection, the controller can become the bottleneck. The magnitude depends on each platform and can’t be pinned to a universal percentage.
Does software RAID use a lot of CPU?
It can consume noticeable resources under intensive IOPS and parity RAID workloads, though it depends on the processor, RAID level, software, and access pattern. The 18% to 28% figures cited by Graid come from the company’s own measurements.
What is SupremeRAID?
SupremeRAID is Graid’s technology that uses an NVIDIA GPU to offload RAID-related processing and reduce the work performed by server CPUs.
Are enterprise SSD prices rising because of AI?
Demand from data centers and AI infrastructure is sustaining the enterprise SSD market and putting pressure on NAND prices. TrendForce expects further increases during 2026, though the size varies by product and segment.
Source: storage newsletter