Key Statistics
Key Takeaways
- Market trajectory: the near-memory compute in-SSD processor market reaches USD 211 million in 2026 and USD 1,280 million by 2036 at 19.8% growth, supported by selective adoption of data-local processing rather than conventional flash-capacity expansion alone.
- Computational storage devices: the largest type, with an estimated ~44% of 2026 revenue, bring processing into a drive-shaped product. Controller-embedded solutions account for ~31%, while near-storage processors and NAND-integrated engines remain smaller development paths.
- AI and machine-learning inference: the largest application, at an estimated ~36% of 2026 revenue, is linked to vector search, feature access and context data. Database filtering and transparent compression remain nearer-term cases with easier measurement of host-CPU savings.
- North America: the leading regional market with an estimated ~43% of 2026 revenue, benefits from hyperscale qualification programs, storage-software ecosystems and domestic SSD controller design. Asia-Pacific, at ~32%, is critical for NAND and drive manufacturing.
- Interoperability and return on investment: SNIA and NVMe standardization are reducing software integration risk, but deployment decisions still depend on demonstrated workload benefits, security isolation, endurance and total cost of ownership across production storage fleets.
Near Memory Compute (NMC) In-SSD Processor Market Overview
The near-memory compute (NMC) in-SSD processor market stood at USD 211 million in 2026 and is projected to reach USD 1,280 million by 2036, expanding at a CAGR of 19.8% during 2027–2036. North America accounted for an estimated ~43% of base-year revenue, while Asia-Pacific supplied much of the underlying NAND and SSD manufacturing capacity.
Near-memory compute in SSD processors places execution resources within the drive controller, an attached storage processor, or a programmable computational-storage device rather than transferring every data block to a general-purpose host. The economic boundary covers incremental silicon, firmware and specialized drive hardware sold for data-local computation, not the total enterprise SSD market. That distinction matters because ordinary NVMe drives already perform error correction and housekeeping, but those functions do not by themselves constitute customer-programmable computational storage.
The Semiconductor Industry Association reported USD 791.7 billion in worldwide semiconductor sales for 2025, including USD 223.1 billion in memory products; those industry totals provide scale context and must not be confused with the narrower NMC opportunity. The market has two commercialization tracks. Drive-resident transparent functions such as compression, data placement and encryption can fit existing operating-system and database interfaces, making qualification relatively straightforward. Programmable analytics, predicate pushdown and AI feature processing can offer larger host-resource savings, but they require an application execution model, isolation controls and reproducible benchmark evidence. ScaleFlux positions its CSD3000 around transparent compression, while Samsung demonstrated programmable SmartSSD computation with Arm cores and FPGA acceleration.
Hardware and software standards now offer a stronger basis for multi-vendor deployment. SNIA published Computational Storage API version 1.1 on 31 May 2025 and its Architecture and Programming Model version 1.2 on 21 April 2026. NVM Express released NVMe 2.4 specifications in August 2026, including the Computational Programs Command Set. These documents separate the device discovery, execution and transport layers, reducing the risk that storage applications must be rewritten for every controller family.
Capacity density and computational offload are related but not interchangeable. Solidigm introduced a 122 TB PCIe SSD in November 2024, and ScaleFlux describes up to 256 TB effective capacity on selected CSD5000 configurations through write-reduction technology. Neither capacity figure is equivalent to NMC revenue. The addressable opportunity instead emerges where a drive processor measurably reduces data transfers, extends endurance, saves host CPU cycles or lowers energy use without unacceptable latency or software migration cost.
AI inference is expanding the set of candidate workloads because vector indexes, retrieval-augmented generation and context stores repeatedly access large data structures that may not fit economically in accelerator memory. NVIDIA announced a BlueField-4-based context memory storage platform in January 2026, illustrating the wider move toward specialized storage data paths; it is not itself proof of an in-SSD processor sale. Buyers will distinguish between functions executed inside SSD silicon and acceleration provided by separate DPUs or storage servers.
Segment Analysis: By Type
By type, computational storage devices hold an estimated ~44% of 2026 revenue, followed by SSD controller-embedded processors at ~31%, near-storage processing units at ~16% and NAND flash-integrated compute engines at ~9%. The categories differ in how tightly processing is coupled to the flash data path and whether software must explicitly invoke it. Forecast growth ranges from ~18.7% to ~22.5% across the four groups, reflecting maturity and integration barriers.
| Type | Function | Market position |
|---|---|---|
| Computational Storage Devices (CSDs) | SSD-form-factor products combine NAND storage with fixed-function or programmable processing. Samsung SmartSSD demonstrates a device-level compute approach, while ScaleFlux CSD families emphasize transparent data reduction. Enterprise buyers can qualify a single replaceable drive and retain established NVMe monitoring and lifecycle practices. For Computational Storage Devices (CSDs), enterprise engineers compare execution isolation, service-level latency, application portability, endurance and data integrity against conventional NVMe alternatives. For Computational Storage Devices (CSDs), this technical assessment determines whether the compute location creates a durable advantage for production storage fleets. | Largest deployable product class because a familiar U.2 or EDSFF envelope simplifies procurement, qualification and spare-parts management. The model assigns this class an estimated ~44% of 2026 revenue and ~18.7% CAGR; a strong installed drive base does not imply that every SSD has computational capabilities. The Computational Storage Devices (CSDs) business case requires a documented application interface and a reproducible performance-per-watt gain under realistic data distributions. For Computational Storage Devices (CSDs), supplier qualification, firmware recovery and customer-specific operating costs remain principal barriers to volume deployment. |
| SSD Controller-Embedded Processors | Programmable cores and accelerators are integrated into the SSD controller data path to handle compression, encryption, placement or selected scan operations. The value proposition depends on preserving flash translation, wear management and deterministic latency while exposing useful computation to the host. For SSD Controller-Embedded Processors, enterprise engineers compare execution isolation, service-level latency, application portability, endurance and data integrity against conventional NVMe alternatives. For SSD Controller-Embedded Processors, this technical assessment determines whether the compute location creates a durable advantage for production storage fleets. | Controller integration accounts for an estimated ~31% share and ~19.5% CAGR. Broadcom and Solidigm extended their controller supply arrangement in January 2025, illustrating the scale of enterprise controller ecosystems; the agreement is broader than the narrower NMC processor revenue boundary. The SSD Controller-Embedded Processors business case requires a documented application interface and a reproducible performance-per-watt gain under realistic data distributions. For SSD Controller-Embedded Processors, supplier qualification, firmware recovery and customer-specific operating costs remain principal barriers to volume deployment. |
| Near-Storage Processing Units (NSPUs) | Dedicated processing units adjacent to the SSD or storage subsystem provide a larger software execution environment than a constrained flash controller. They can support analytics and data-path acceleration without placing all compute logic in the NAND management silicon. For Near-Storage Processing Units (NSPUs), enterprise engineers compare execution isolation, service-level latency, application portability, endurance and data integrity against conventional NVMe alternatives. For Near-Storage Processing Units (NSPUs), this technical assessment determines whether the compute location creates a durable advantage for production storage fleets. | The segment represents an estimated ~16% share with ~21.5% CAGR, but classification requires care: a standalone DPU serving many SSDs is adjacent infrastructure, not automatically an in-SSD processor. Qualification focuses on host integration, scheduling and measurable reduction of storage traffic. The Near-Storage Processing Units (NSPUs) business case requires a documented application interface and a reproducible performance-per-watt gain under realistic data distributions. For Near-Storage Processing Units (NSPUs), supplier qualification, firmware recovery and customer-specific operating costs remain principal barriers to volume deployment. |
| NAND Flash-Integrated Compute Engines | Research architectures put limited computation very close to NAND arrays or inside memory-oriented processing blocks. This can reduce movement for specialized scans, but process integration, firmware support and wear behavior complicate the move from laboratory prototypes to volume drives. For NAND Flash-Integrated Compute Engines, enterprise engineers compare execution isolation, service-level latency, application portability, endurance and data integrity against conventional NVMe alternatives. For NAND Flash-Integrated Compute Engines, this technical assessment determines whether the compute location creates a durable advantage for production storage fleets. | The smallest type holds an estimated ~9% share and ~22.5% CAGR. Its higher modeled growth reflects a low initial base, not a claim of current production dominance. Commercialization depends on repeatable yields, standards-compatible interfaces and workloads that justify silicon complexity. The NAND Flash-Integrated Compute Engines business case requires a documented application interface and a reproducible performance-per-watt gain under realistic data distributions. For NAND Flash-Integrated Compute Engines, supplier qualification, firmware recovery and customer-specific operating costs remain principal barriers to volume deployment. |
Computational Storage Devices (CSDs)
For Computational Storage Devices (CSDs), the engineering trade-off is the balance between available compute resources and reliable storage service. Its estimated ~44% share and ~18.7% CAGR reflect the difficulty of moving from device demonstration to repeatable deployment. A Computational Storage Devices (CSDs) buyer evaluates firmware continuity, workload fit, failure recovery, data movement and power budgets before committing to this product class.
The commercial path for Computational Storage Devices (CSDs) depends on whether suppliers can reduce total host work without making application operation more fragile. For Computational Storage Devices (CSDs), controller and software teams must establish a stable boundary for program execution, storage management and telemetry. Contracts for Computational Storage Devices (CSDs) favor verified mixed-I/O performance and support plans spanning several enterprise drive generations.
SSD Controller-Embedded Processors
For SSD Controller-Embedded Processors, the engineering trade-off is the balance between available compute resources and reliable storage service. Its estimated ~31% share and ~19.5% CAGR reflect the difficulty of moving from device demonstration to repeatable deployment. A SSD Controller-Embedded Processors buyer evaluates firmware continuity, workload fit, failure recovery, data movement and power budgets before committing to this product class.
The commercial path for SSD Controller-Embedded Processors depends on whether suppliers can reduce total host work without making application operation more fragile. For SSD Controller-Embedded Processors, controller and software teams must establish a stable boundary for program execution, storage management and telemetry. Contracts for SSD Controller-Embedded Processors favor verified mixed-I/O performance and support plans spanning several enterprise drive generations.
Near-Storage Processing Units (NSPUs)
For Near-Storage Processing Units (NSPUs), the engineering trade-off is the balance between available compute resources and reliable storage service. Its estimated ~16% share and ~21.5% CAGR reflect the difficulty of moving from device demonstration to repeatable deployment. A Near-Storage Processing Units (NSPUs) buyer evaluates firmware continuity, workload fit, failure recovery, data movement and power budgets before committing to this product class.
The commercial path for Near-Storage Processing Units (NSPUs) depends on whether suppliers can reduce total host work without making application operation more fragile. For Near-Storage Processing Units (NSPUs), controller and software teams must establish a stable boundary for program execution, storage management and telemetry. Contracts for Near-Storage Processing Units (NSPUs) favor verified mixed-I/O performance and support plans spanning several enterprise drive generations.
NAND Flash-Integrated Compute Engines
For NAND Flash-Integrated Compute Engines, the engineering trade-off is the balance between available compute resources and reliable storage service. Its estimated ~9% share and ~22.5% CAGR reflect the difficulty of moving from device demonstration to repeatable deployment. A NAND Flash-Integrated Compute Engines buyer evaluates firmware continuity, workload fit, failure recovery, data movement and power budgets before committing to this product class.
The commercial path for NAND Flash-Integrated Compute Engines depends on whether suppliers can reduce total host work without making application operation more fragile. For NAND Flash-Integrated Compute Engines, controller and software teams must establish a stable boundary for program execution, storage management and telemetry. Contracts for NAND Flash-Integrated Compute Engines favor verified mixed-I/O performance and support plans spanning several enterprise drive generations.
Segment Analysis: By Application
Applications are led by AI and machine-learning inference at an estimated ~36% of 2026 revenue, followed by data analytics at ~25%, database query processing at ~19%, compression and encryption at ~14%, and other uses at ~6%. Inference has the fastest modeled expansion at ~23.0% CAGR, while the simpler transparent offload functions offer a more immediate deployment path. The value case is measured through data reduction and host resource savings.
| Application | Demand characteristics |
|---|---|
| AI & Machine Learning Inference | An estimated ~36% of 2026 revenue and ~23.0% CAGR. Vector-index scans, embedding lookups, retrieval pipelines and inference context handling can create repeated movement of storage-resident data into host memory. In-SSD filtering is most valuable when a large fraction of candidate records can be discarded before crossing PCIe, and when the execution model does not disrupt GPU scheduling. |
| Data Analytics & Business Intelligence | An estimated ~25% of 2026 revenue and ~17.0% CAGR. Columnar scans, aggregation and selection predicates can move close to flash when operators access large datasets with relatively simple operations. The business case depends on reducing bytes moved rather than performing complex joins in a constrained drive environment, and results vary with data compression and query selectivity. |
| Database Query Processing | An estimated ~19% of 2026 revenue and ~19.0% CAGR. Database engines can push eligible filters, scans or data-placement hints toward the storage layer, but transaction consistency and optimizer integration remain critical. Kioxia reported laboratory RocksDB benefits from Flexible Data Placement in 2024; that host-managed feature is related evidence for SSD intelligence, not identical to general-purpose in-drive execution. |
| Data Compression & Encryption | An estimated ~14% of 2026 revenue and ~16.0% CAGR. Transparent compression, decompression and cryptographic functions have a direct operating-cost rationale because they reduce write amplification, data transfer or host cycles without changing application semantics. ScaleFlux describes CSD3000 compression and endurance improvements, although realized benefits depend strongly on dataset compressibility and the storage workload. A deployment decision depends on benchmarked application throughput, fleet observability and predictable recovery behavior when an offload function fails or firmware is updated. |
| Others | An estimated ~6% of 2026 revenue and ~16.5% CAGR. Specialized uses include genomic data preprocessing, telemetry reduction, media transcoding and selected scientific computing kernels. The value is highly workload-dependent because computation may be constrained by drive power, flash bandwidth, execution memory or software integration; small projects do not automatically translate into a repeatable hardware market. |
AI & Machine Learning Inference represents an estimated ~36% of 2026 revenue and is projected to grow at ~23.0% annually. For AI & Machine Learning Inference, adoption centers on whether SSD-resident functions reduce data traffic or host processing without disrupting established reliability controls. Buyers of AI & Machine Learning Inference test production inputs, host integration, secure firmware, failure handling and full-system economics before adopting a specialized drive configuration.
Data Analytics & Business Intelligence represents an estimated ~25% of 2026 revenue and is projected to grow at ~17.0% annually. For Data Analytics & Business Intelligence, adoption centers on whether SSD-resident functions reduce data traffic or host processing without disrupting established reliability controls. Buyers of Data Analytics & Business Intelligence test production inputs, host integration, secure firmware, failure handling and full-system economics before adopting a specialized drive configuration.
Database Query Processing represents an estimated ~19% of 2026 revenue and is projected to grow at ~19.0% annually. For Database Query Processing, adoption centers on whether SSD-resident functions reduce data traffic or host processing without disrupting established reliability controls. Buyers of Database Query Processing test production inputs, host integration, secure firmware, failure handling and full-system economics before adopting a specialized drive configuration.
Data Compression & Encryption represents an estimated ~14% of 2026 revenue and is projected to grow at ~16.0% annually. For Data Compression & Encryption, adoption centers on whether SSD-resident functions reduce data traffic or host processing without disrupting established reliability controls. Buyers of Data Compression & Encryption test production inputs, host integration, secure firmware, failure handling and full-system economics before adopting a specialized drive configuration.
Others represents an estimated ~6% of 2026 revenue and is projected to grow at ~16.5% annually. For Others, adoption centers on whether SSD-resident functions reduce data traffic or host processing without disrupting established reliability controls. Buyers of Others test production inputs, host integration, secure firmware, failure handling and full-system economics before adopting a specialized drive configuration.
Regional Analysis
North America is the largest NMC in-SSD processor market with an estimated ~43% share in 2026, ahead of Asia-Pacific (~32%), Europe (~16%), Middle East & Africa (~5%) and South America (~4%). Asia-Pacific has the fastest modeled regional CAGR at ~21.0%, supported by SSD manufacturing and controller integration, while the smaller markets are tied to selective enterprise and cloud deployments. These percentages describe the narrow computational-storage revenue pool, not total SSD or data-center equipment sales.
How does adoption differ across storage production and cloud demand regions?
Cloud procurement and semiconductor production are geographically different activities. North America concentrates early software qualification and fleet-scale cloud evaluation, while Asia-Pacific hosts much of the NAND manufacturing and SSD engineering needed to implement data-local processing. Europe places greater emphasis on regulated enterprise environments and standards portability. Regional revenue is therefore shaped by the location of the buying decision and the incremental compute functionality purchased, not simply the location where a NAND wafer is fabricated.
The regional model assigns North America ~43% share with ~20.0% CAGR, Asia-Pacific ~32% with ~21.0%, Europe ~16% with ~18.0%, Middle East & Africa ~5% with ~17.0%, and South America ~4% with ~15.0%. Each geography faces different qualification economics, but all must establish whether offload improves total system throughput and energy use after software and operational costs are included. These modeled growth rates are not independently published country-level statistics.
In the United States, Samsung SmartSSD and ScaleFlux drive products give customers contrasting options for programmable and transparent functions. In Japan, Kioxia has demonstrated data-placement optimization for RocksDB, a related but distinct approach. SNIA and NVM Express standards support interoperability across these architectures. The next commercialization hurdle is reliable host integration and service tooling, including monitoring, firmware rollback, security isolation and workload-specific evidence of cost reduction.
| Region | Position | Growth outlook | Demand profile | What decides supplier selection |
|---|---|---|---|---|
| North America | ~43% share in 2026 | ~20.0% CAGR forecast | Largest demand and software qualification | In North America, the commercial priority is largest demand and software qualification. Customers in North America require secure firmware, workload benchmarking, documented integration and reliable support contracts. Qualification in North America must distinguish computational SSD silicon from conventional drives and external DPUs, because ordinary storage-capacity investment is not incremental NMC processing revenue. The procurement path in North America also depends on vendor service coverage and software maintenance economics. |
| Asia-Pacific | ~32% share in 2026 | ~21.0% CAGR forecast | NAND supply and SSD manufacturing | In Asia-Pacific, the commercial priority is nand supply and ssd manufacturing. Customers in Asia-Pacific require secure firmware, workload benchmarking, documented integration and reliable support contracts. Qualification in Asia-Pacific must distinguish computational SSD silicon from conventional drives and external DPUs, because ordinary storage-capacity investment is not incremental NMC processing revenue. The procurement path in Asia-Pacific also depends on vendor service coverage and software maintenance economics. |
| Europe | ~16% share in 2026 | ~18.0% CAGR forecast | Enterprise data sovereignty and standards | In Europe, the commercial priority is enterprise data sovereignty and standards. Customers in Europe require secure firmware, workload benchmarking, documented integration and reliable support contracts. Qualification in Europe must distinguish computational SSD silicon from conventional drives and external DPUs, because ordinary storage-capacity investment is not incremental NMC processing revenue. The procurement path in Europe also depends on vendor service coverage and software maintenance economics. |
| Middle East & Africa | ~5% share in 2026 | ~17.0% CAGR forecast | Sovereign AI and selective edge systems | In Middle East & Africa, the commercial priority is sovereign ai and selective edge systems. Customers in Middle East & Africa require secure firmware, workload benchmarking, documented integration and reliable support contracts. Qualification in Middle East & Africa must distinguish computational SSD silicon from conventional drives and external DPUs, because ordinary storage-capacity investment is not incremental NMC processing revenue. The procurement path in Middle East & Africa also depends on vendor service coverage and software maintenance economics. |
| South America | ~4% share in 2026 | ~15.0% CAGR forecast | Enterprise storage optimization | In South America, the commercial priority is enterprise storage optimization. Customers in South America require secure firmware, workload benchmarking, documented integration and reliable support contracts. Qualification in South America must distinguish computational SSD silicon from conventional drives and external DPUs, because ordinary storage-capacity investment is not incremental NMC processing revenue. The procurement path in South America also depends on vendor service coverage and software maintenance economics. |
Competitive Landscape and Technology Positioning
The competitive field spans SSD manufacturers, controller designers, computational-storage specialists and standards organizations. Direct in-SSD products are distinguishable from adjacent AI storage infrastructure: Samsung has demonstrated programmable SmartSSD drives, and ScaleFlux offers computational-storage SSDs with transparent data reduction. Broadcom and Solidigm illustrate the importance of controller supply relationships, while SNIA and NVMe define interoperability conditions that influence purchasing confidence.
Samsung Electronics described a second-generation SmartSSD in July 2022 with in-drive computation and workload-specific reductions in CPU use and processing time. The commercial implication is a programmable offload architecture that can execute eligible functions near flash. A buyer must nevertheless test the application stack, FPGA or processor configuration, data consistency and deployment cost before translating a benchmark into an annual procurement budget.
ScaleFlux markets CSD3000 and CSD5000 products with controller-level write reduction, NVMe interfaces and enterprise SSD form factors. The company positions transparent compression as a way to improve effective capacity, endurance and performance without requiring application changes. This is a materially different adoption proposition from arbitrary device programs: software integration is lighter, but realized performance remains dependent on dataset characteristics and workload patterns.
Solidigm and Broadcom extended a multi-year controller agreement in January 2025, underscoring the importance of reliable controller supply, qualification and firmware continuity for enterprise storage. Kioxia demonstrated Flexible Data Placement with RocksDB in October 2024. Both activities show how SSD intelligence is expanding, but controller procurement and host-directed placement should not be misclassified as sales of programmable in-SSD compute processors.
NVMe Computational Programs and SNIA architectural standards provide a common vocabulary for device discovery, program loading and execution. They may reduce integration friction over time, but a standard alone does not establish market share for any hardware vendor. Buyers also compare in-drive processing with CPUs, GPUs and DPUs; NVIDIA BlueField-4 is relevant as an alternative storage architecture, not as proof of direct in-SSD processor supply.
Tier structure
| Tier | Companies | Basis of competition |
|---|---|---|
| Direct computational SSD suppliers | Samsung Electronics; ScaleFlux | Companies with published products or demonstrations that place processing functions inside SSD devices. Their differentiation concerns programmable versus transparent functions, controller integration, endurance and host software requirements. Product claims should be validated against the buyer’s own application workloads before an adoption decision. This tier carries direct device-level functional evidence and must be separated from ordinary SSD capacity announcements and wider storage-controller contracts. |
| Controller and NAND ecosystem | Broadcom; Solidigm; Kioxia; SK hynix | Suppliers and technology contributors that influence controller design, NAND characteristics, SSD firmware and data placement. Their wider storage products are adjacent to NMC, and the full revenue of conventional enterprise drives is not included in the narrow computational-storage processor category. Procurement analysis needs a separate allocation for processing features and should not use total company storage revenue as a proxy for in-SSD computation. |
| Adjacent acceleration and standards | AMD; NVIDIA; SNIA; NVM Express | AMD documents FPGA-assisted SmartSSD hardware, NVIDIA provides DPU-centered storage acceleration, and industry bodies publish device-interface standards. These roles shape the addressable workload and interoperability environment without implying that each participant independently sells the same in-SSD processor product. These contributors influence software portability and alternative architectures, but the narrower revenue boundary requires an actual qualifying drive-resident processing product or directly attributable component. |
Key companies profiled
- Samsung Electronics
- ScaleFlux
- Broadcom
- Solidigm
- Kioxia
- SK hynix
- AMD
- NVIDIA
Technology Architecture and Product Qualification
An NMC processor is commercially useful only when its location in the data path matches the workload. Transparent compression can sit inside controller firmware, while programmable computational storage needs a device execution environment, memory protection and a stable application interface. Hardware designers must protect NAND management tasks from unpredictable user workloads so that latency, endurance and recoverability remain within enterprise service requirements.
NVMe defines the transport and command architecture that makes enterprise SSDs interoperable across host platforms. The Computational Programs Command Set provides mechanisms to discover, download and execute programs in a storage subsystem. This does not mean that every NVMe 2.4 drive implements programmable computation; product capability must be verified separately. Device makers face a trade-off between allocating silicon area to acceleration and preserving power and thermal margins for flash operations.
Samsung SmartSSD combines SSD media with programmable hardware, and AMD documentation describes a 3.84 TB U.2 configuration with a Kintex UltraScale+ FPGA. These published specifications demonstrate how offload can be packaged in a recognizable enterprise drive envelope. The software still needs to define what executes locally, how intermediate results are returned and what happens when a drive is replaced or firmware is upgraded.
Kioxia demonstrated Flexible Data Placement for RocksDB with laboratory performance and lifespan gains in October 2024. FDP is not general-purpose in-SSD computation, but it highlights how host-directed placement can change the storage workload without moving an application kernel into flash silicon. Enterprise buyers should compare such simpler features with programmable offload to determine whether additional device compute complexity produces a meaningful net advantage.
Market Dynamics, Adoption Barriers and Opportunities
Growth is driven by data-intensive AI and database workloads, controller-level improvements and standardization. The constraints are equally material: software integration, the limited thermal and compute budgets of SSDs, and uncertain economics outside carefully selected workloads. The direction of demand depends on whether suppliers can convert benchmark improvements into repeatable fleet-wide savings while maintaining data integrity and secure operation.
MARKET DRIVERS
Drivers Impact Analysis*
| Driver | (~) % impact on CAGR forecast | Geographic relevance | Impact timeline |
|---|---|---|---|
| AI inference data movement | +3.1% | North America and Asia-Pacific | 2027–2033 |
| Transparent controller offload | +2.2% | Global enterprise SSDs | 2027–2036 |
| Standards interoperability | +1.7% | Global | 2027–2032 |
AI retrieval and context processing
Vector search and inference context management increase the amount of data repeatedly read from SSDs. Some scans can be filtered near storage, reducing host transfers when selectivity is high and execution is simple. The benefit is not universal: GPU scheduling, index layout and cache behavior can dominate total latency. Commercial demand grows when customers can show that in-drive processing lowers full-system cost on production retrieval workloads.
Controller-level data reduction
ScaleFlux CSD3000 and CSD5000 demonstrate an enterprise route through transparent compression and write reduction. These functions can lower the physical writes and host overhead for compressible datasets without requiring a new application programming model. Buyers still need to benchmark incompressible data, mixed read-write traffic and failure recovery. Demand therefore depends on measurable endurance and capacity benefits rather than headline compression ratios alone.
Common execution and management interfaces
SNIA released Computational Storage API v1.1 in May 2025 and an updated architecture model in April 2026. NVMe specifications in August 2026 include computational program operations. Together these standards clarify the roles of host software and storage devices. A common interface can lower vendor lock-in and application maintenance costs, although suppliers must implement the relevant features before buyers obtain those benefits.
MARKET RESTRAINTS
Restraints Impact Analysis*
| Restraint | (~) % impact on CAGR forecast | Geographic relevance | Impact timeline |
|---|---|---|---|
| Software integration complexity | −2.3% | All regions | 2027–2036 |
| Power and thermal budgets | −1.6% | Dense SSD racks | 2027–2034 |
| Alternative CPU/DPU architectures | −1.4% | Cloud and enterprise | 2027–2036 |
Application portability and security
Moving execution into an SSD requires a safe interface for code loading, data access, error handling and resource accounting. Proprietary firmware can impose testing and maintenance costs that exceed a narrow benchmark gain. Operators also need to know how programs are isolated from sensitive data and flash management functions. These requirements slow adoption where standard host-side processing is already operationally reliable.
Device power and endurance constraints
An enterprise SSD has a fixed power and cooling envelope shared by the controller, NAND media, DRAM and any accelerator. Adding compute can increase thermal stress or reduce bandwidth available to normal I/O. Compression may improve endurance for some datasets, but user-programmable workloads can create unpredictable contention. Qualification must cover sustained mixed traffic, firmware updates, drive failures and data recovery under realistic rack conditions.
Competition from external accelerators
CPU instruction sets, GPUs, SmartNICs and DPUs can perform data-path acceleration without embedding user programs inside each drive. NVIDIA announced BlueField-4-based AI context storage in January 2026, illustrating this adjacent approach. NMC vendors must show why executing at the SSD is better than consolidating acceleration at the server or network level, especially when workloads require shared memory and complex coordination.
MARKET OPPORTUNITIES
Selective database pushdown
Database query engines can offload simple scans or predicates to reduce bytes returned to the host. The strongest candidates have large sequential reads, high filtering selectivity and stable query patterns. Integrators can begin with narrowly scoped kernels, compare them with RocksDB or NVMe placement optimizations and expand only where the end-to-end performance and support costs remain favorable.
AI storage fleet efficiency
High-capacity drives are becoming central to AI data pipelines, as Solidigm and ScaleFlux product disclosures demonstrate. NMC adds a separate potential value layer by processing or reducing data before transfer. Operators with thousands of drives can measure power, CPU utilization and storage traffic at fleet scale, providing a stronger basis for procurement than standalone peak-throughput tests.
Edge and regulated workloads
Edge systems and regulated enterprises may value local filtering when moving data over a network is expensive or restricted. The opportunity is strongest for deterministic compression, telemetry reduction or small analytics kernels that fit within drive power budgets. Buyers still need standardized software, secure updates and vendor support, which can be harder to obtain outside major data-center markets.
Supply Chain, Controllers and Commercial Integration
The NMC value chain starts with NAND wafers and enterprise SSD controllers, adds firmware or programmable acceleration, and ends with application software qualification. The incremental computational-storage value must be separated from ordinary NAND capacity and generic SSD sales. Supply resilience matters because a controller redesign or firmware change can force a new qualification cycle across an entire storage fleet.
Memory and controller sourcing. The NAND manufacturing base is concentrated in Asia-Pacific, while software qualification and system integration span North America, Europe and other data-center markets. The SK hynix 321-layer QLC client SSD supply announcement in April 2026 demonstrates continuing density improvements, but a standard client drive does not become an NMC product merely because it serves an AI PC. The market boundary follows the computation function and the price paid for it.
Enterprise firmware lifecycle. Solidigm and Broadcom extended a controller supply agreement in January 2025, showing the importance of multiyear silicon and firmware continuity. A computational SSD adds program management and execution behavior to the conventional controller stack. Operators need consistent telemetry, stable host drivers and support for security patches across multiple drive generations, or the operational cost of the feature can erase performance gains.
Application integration. A storage-side operation only produces value when software can use it correctly. SNIA API v1.1 defines an application-facing interface, while the NVMe Computational Programs Command Set specifies device program operations. Developers must still design query plans, data layout and failure recovery. The commercial bottleneck is often the cost of maintaining these integrations rather than the raw ability of silicon to execute a function.
Validation and service. Enterprise qualification covers endurance, sustained latency, power-loss behavior, controller fault isolation and device replacement. Kioxia’s 2024 RocksDB laboratory tests illustrate how performance and lifespan can be measured for a specific workload. Procurement teams should repeat comparable tests on their own data, including low-compressibility and mixed-transaction scenarios, before attributing fleet-level savings to a storage feature.
Recent Developments in Computational Storage
- August 2026 STANDARDS
NVM Express released its 2.4 specification set with the Computational Programs Command Set, describing mechanisms for discovering, downloading and executing programs on supported storage devices. The development matters because interoperable commands can reduce software fragmentation between controller families. Actual adoption still requires vendor implementations and application testing, so the publication is a standards milestone rather than proof of production shipments. Source - 21 April 2026 STANDARDS
SNIA published Computational Storage Architecture and Programming Model version 1.2, updating the common model for computational-storage hardware and software behavior. This gives developers and buyers a clearer framework for device capabilities, interfaces and management expectations. Commercial impact depends on whether products implement compatible functions and whether customers can verify real workload improvements with acceptable operational risk. Source - 8 April 2026 SSD SUPPLY
SK hynix announced supply of its 321-layer QLC NAND client SSD line with 1 TB and 2 TB options, beginning with Dell. This illustrates continuing improvements in conventional flash capacity and efficiency, an important baseline against which NMC products must compete. The announcement does not itself establish sales of in-SSD computational processors, and its value is principally storage supply-chain context. Source - 5 January 2026 AI STORAGE
NVIDIA introduced a BlueField-4-based inference context memory storage platform designed for AI workloads that require high-bandwidth context access. The initiative illustrates growing demand for specialized storage data paths and data movement optimization. Because the processing platform is DPU-centered rather than necessarily located inside an SSD, its spending cannot be counted automatically as direct NMC processor revenue. Source - 31 May 2025 STANDARDS
SNIA published Computational Storage API version 1.1, defining an interface between applications and computational-storage devices. The release is commercially relevant because software portability is a key adoption barrier for programmable storage acceleration. Device vendors still need compatible implementations, and buyers must validate error handling, security isolation and integration with production storage stacks before relying on the interface. Source - 23 January 2025 CONTROLLERS
Solidigm extended its multiyear agreement with Broadcom for high-capacity SSD controllers supporting AI and data-intensive storage. The collaboration underlines the importance of controller continuity and qualification in enterprise drive supply. It is an adjacent storage-controller development rather than evidence that every controller under the agreement includes programmable NMC functionality; this distinction is material to the market boundary. Source - 15 October 2024 STORAGE SOFTWARE
Kioxia demonstrated a Flexible Data Placement-enabled SSD running RocksDB at the OCP Global Summit, reporting laboratory gains in performance and flash lifespan. The result shows that software-aware storage behavior can produce measurable application benefits. FDP is not the same as arbitrary in-drive program execution, so the development is best interpreted as a related storage-intelligence use case. Source
REPORT SCOPE & SEGMENTATION
| Attribute | Details |
|---|---|
| Report title | Near Memory Compute (NMC) In-SSD Processor Market, Trends, Business Strategies 2026–2036 |
| Base year | 2026: USD 211 million |
| Estimated year | 2027: USD 252 million |
| Forecast year | 2036: USD 1,280 million |
| Forecast CAGR | 19.8% during 2027–2036 |
| Revenue boundary | Incremental revenue attributable to SSD-resident or directly coupled near-storage processing silicon, firmware and qualifying computational-storage drives; ordinary SSD media sales and external GPU/DPU systems are excluded. |
| Study period | 2021–2036, with 2021–2025 historical context and 2027–2036 forecasts |
| By Type | Computational Storage Devices (CSDs) · SSD Controller-Embedded Processors · Near-Storage Processing Units (NSPUs) · NAND Flash-Integrated Compute Engines |
| By Application | AI & Machine Learning Inference · Data Analytics & Business Intelligence · Database Query Processing · Data Compression & Encryption · Others |
| By End User | Cloud Service Providers & Hyperscalers · Enterprise Data Centers · Telecommunications & Network Operators · Research & Academic Institutions |
| By Interface Standard | NVMe-based NMC Processors · PCIe-attached Computational Storage · SATA-based Near-Storage Compute |
| By Region | Five-region revenue allocation for 2026; all percentages are estimated shares of qualifying NMC processor revenue. North AmericaUnited States, Canada, Mexico Asia-PacificSouth Korea, Japan, China EuropeGermany, United Kingdom, France Middle East & AfricaUnited Arab Emirates, Saudi Arabia, South Africa South AmericaBrazil, Chile, Argentina |
| Key Companies Profiled | Samsung Electronics · ScaleFlux · Broadcom · Solidigm · Kioxia · SK hynix · AMD · NVIDIA |
| Customization Scope | Country-level application analysis, enterprise storage architecture, controller and NAND value-chain scope, software compatibility, workload-specific ROI and procurement assessment. |
Frequently Asked Questions
How large is the near-memory compute in-SSD processor market?
The global market was USD 211 million in 2026 and is projected to reach USD 1,280 million in 2036, implying 19.8% CAGR over 2027–2036. The scope covers qualifying processing functions sold in or directly with SSD architectures rather than the entire value of standard NAND flash storage. Commercial growth depends on measurable reductions in data movement, host processing or operating cost.
Which NMC processor type leads revenue?
Computational Storage Devices hold the largest estimated 2026 share at ~44%, followed by controller-embedded processors at ~31%. CSDs fit enterprise procurement because they can be installed in recognizable SSD form factors. The categories differ in whether processing is transparent, programmable or closely integrated with NAND management, and that distinction affects software integration and qualification cost.
Which application has the largest NMC demand?
AI and machine-learning inference accounts for an estimated ~36% of 2026 revenue, followed by analytics at ~25% and database processing at ~19%. Inference can benefit when storage-side filtering reduces repeated movement of large indexes or context data. However, a storage accelerator is useful only if end-to-end system performance improves after software and GPU coordination are considered.
Which region is the largest market?
North America leads with an estimated ~43% of 2026 revenue, while Asia-Pacific represents ~32% and Europe ~16%. North America has a strong concentration of cloud software qualification and controller design, whereas Asia-Pacific anchors NAND and SSD production. These are modeled shares of the narrow computational-storage segment, not the distribution of all enterprise SSD revenue.
What is the role of SNIA standards in computational storage?
SNIA Computational Storage API version 1.1, published in May 2025, and Architecture and Programming Model version 1.2, published in April 2026, establish common interface and behavior concepts. They help developers reason about device capabilities and application portability. Their presence does not guarantee implementation, so enterprise buyers must confirm that a particular drive supports the relevant features.
How does NVMe computational storage differ from a conventional SSD?
A conventional NVMe SSD moves blocks and performs internal flash-management operations. A computational-storage implementation additionally exposes processing or program execution related to stored data. NVMe 2.4 includes a Computational Programs Command Set, but not every compliant drive implements it. Buyers should verify the command set, isolation model, firmware behavior and application interface before treating a drive as programmable computational storage.
Are high-capacity AI SSDs automatically part of this market?
No. A 122 TB enterprise SSD or a 256 TB effective-capacity configuration is not automatically an in-SSD compute processor sale. Capacity, endurance and transparent data reduction are separate product attributes. Only the qualifying incremental processing capability belongs in the narrow NMC market, and adjacent DPU or GPU systems must also be excluded unless they are directly within the defined storage-processing boundary.
What are the most important deployment barriers?
Software portability, workload variability, thermal constraints and data security are the principal barriers. A function that performs well on a selective database scan may add little value to random transactional workloads or incompressible data. Enterprise operators also need recovery procedures, monitoring, secure firmware updates and predictable device replacement, making lifecycle cost as important as a peak benchmark.
What is the difference between Flexible Data Placement and in-SSD computation?
Flexible Data Placement lets host software influence where data is stored within an SSD, potentially reducing write amplification and improving performance. Kioxia demonstrated this approach with RocksDB in 2024. It is related to intelligent storage but does not by itself execute arbitrary application code inside the drive, so its revenue should not be treated as synonymous with programmable computational storage.
How should enterprises assess NMC return on investment?
Enterprises should benchmark real datasets and applications against conventional NVMe drives and alternative CPU or DPU acceleration. Useful metrics include bytes moved, CPU time, sustained latency, endurance, energy use and the cost of software integration. A supplier claim becomes commercially meaningful when these benefits persist across realistic mixed workloads and can be supported through normal firmware and hardware lifecycle operations.
Research Sources & Evidence Base
Get Sample Report PDF for Exclusive Insights
Report Sample Includes
- Table of Contents
- List of Tables & Figures
- Charts, Research Methodology, and more...
Download Sample Report PDF
