Nodvolt / Reports / Logic & Memory / Socamm Socamm2 Memory Module
Logic & Memory Socamm Socamm2 Memory Module Logic & Memory

SOCAMM & SOCAMM2 Memory Module Market - By Generation (SOCAMM, SOCAMM2), By Capacity (128GB, 192GB, 256GB and Above), By Platform (NVIDIA Grace Blackwell, NVIDIA Vera Rubin, Custom AI CPU), By Application (AI Inference Server, AI Training Node, Edge AI System), By Region

Published Date
Jul, 2026
Report Id
Nod-94
Base Value
USD 1.08 Billion
CAGR
40.6%
Forecast Period
USD 32.54 Billion
Market Synopsis

The global SOCAMM and SOCAMM2 memory module market size was USD 1.08 Billion in 2025 and is expected to register a revenue CAGR of 40.6% during the forecast period. SOCAMM, the Small Outline Compression Attached Memory Module, packages LPDDR5X low power DRAM into a detachable, compression-attached server module, bringing mobile-class power efficiency to data centre CPUs without the permanent soldering that previously made LPDDR unserviceable in servers. Micron Technology shipped the first generation SOCAMM in volume for NVIDIA GB300 Grace Blackwell systems in 2025, and the second generation SOCAMM2, aligned with the JEDEC compression attached memory module standardisation track, delivers more than double the bandwidth of DDR5 RDIMM at over 70 percent better power efficiency. SK Hynix quantifies the SOCAMM2 advantage at more than twice RDIMM bandwidth with over 75 percent improved energy efficiency, and Micron measures rack-scale low power DRAM densities exceeding 50 terabytes per rack.

NVIDIA platform adoption is the primary revenue growth driver, because the Vera Rubin generation specifies SOCAMM2 as the CPU-attached memory across its rack-scale systems, converting every Rubin rack shipment into module demand across three qualified suppliers. Industry reporting in September 2025 indicated NVIDIA placed 2026 SOCAMM module orders with Samsung, SK Hynix, and Micron with allocations distributed across all three, and Micron disclosed that offloading inference KV cache from HBM to high-capacity LPDDR modules improves time to first token economics for long context workloads. Samsung began supplying SOCAMM2 customer samples to NVIDIA in December 2025 and resolved module warpage through low temperature solder processing below 150 degrees Celsius. For instance, in April 2026, SK Hynix Inc., South Korea, began mass production of 192GB SOCAMM2 modules built on its sixth generation 1c LPDDR5X process, positioning the product as the first module optimised for the NVIDIA Vera Rubin platform. These are some of the key factors driving revenue growth of the market.

However, the SOCAMM2 module structure is mechanically fragile in ways RDIMM never was, because LPDDR5X packages secured through bolt-based compression are susceptible to warpage from coefficient of thermal expansion mismatches, and a warped module produces intermittent connector failures that are difficult to screen at test. LPDDR5X also carries structurally higher access latency than DDR5, a design trade that confines SOCAMM to bandwidth-bound AI workloads and keeps general purpose server platforms on RDIMM, capping the addressable share of server memory. The market is also captive to a single platform decision maker, since NVIDIA defines the SOCAMM specification and allocates orders across suppliers, and the broader DRAM shortage that suppliers warn could persist until 2027 diverts LPDDR5X wafer supply toward SOCAMM at the direct expense of smartphone customers, creating allocation conflict inside each supplier. These factors substantially limit SOCAMM and SOCAMM2 memory module market growth over the forecast period.

Market Data
SOCAMM2 vs DDR5 RDIMM - Supplier-Disclosed Advantage
Source: Nodvolt Intelligence primary research, SK Hynix and Samsung technical disclosures
SOCAMM2 vs DDR5 RDIMM - Supplier-Disclosed Advantage
SOCAMM2 Supplier Qualification Timeline to Mass Production
Source: Nodvolt Intelligence primary research, company announcements
SOCAMM2 Supplier Qualification Timeline to Mass Production
Questions before purchase?
Get a preview or speak with an analyst
See the exec summary, scope, and sample data before you commit.
Speak with Analyst Request Preview Request Custom Scope
Segment Insights
NVIDIA's decision to standardise LPDDR-based modules across Grace Blackwell and Vera Rubin CPU memory converts every rack shipment into three-supplier SOCAMM demand
The Grace CPU pairs with LPDDR because a rack drawing 190 to 230 kilowatts cannot afford DDR5 RDIMM power on the CPU side, and SOCAMM makes that LPDDR serviceable. Each Vera Rubin NVL144 rack carries dozens of SOCAMM2 modules across its compute trays, so module demand scales linearly with the rack shipments that hyperscalers have contracted through 2027. Industry reporting indicates NVIDIA distributed 2026 SOCAMM2 orders across Samsung, SK Hynix, and Micron at comparable allocations, giving all three suppliers a qualified revenue stream tied to the fastest growing server platform in the industry.
KV cache offload from HBM to high-capacity LPDDR modules gives SOCAMM2 a defined role in inference economics, not just a power efficiency story
Long context inference workloads hold key-value caches that exceed HBM capacity on the accelerator package, and Micron's published testing with a 70 billion parameter model at 500,000 token context lengths demonstrates LPDRAM as a second memory tier that improves time to first token while holding power flat. This gives SOCAMM2 a workload-anchored value proposition, because the module tier sits between HBM and SSD in the inference memory hierarchy and grows with context length. As reasoning models push context windows longer, the capacity demanded from the SOCAMM tier grows independently of CPU count.
The detachable compression-attached design removes the serviceability penalty that kept LPDDR out of servers for a decade
Soldered LPDDR meant a single failed DRAM package scrapped an entire motherboard, an unacceptable failure economics for hyperscale fleets. SOCAMM2's bolt-secured compression connector lets operators replace or upgrade memory without board rework, aligning LPDDR with the field replaceable unit model that hyperscale operations require. Samsung also highlights the horizontal module orientation, which improves heat sink placement and airflow design and remains compatible with both air and liquid cooling, easing integration in dense trays where vertical RDIMM slots obstruct cold plate routing.
JEDEC standardisation of compression attached memory modules opens SOCAMM-class products to non-NVIDIA platforms from 2027
Samsung's SOCAMM2 is expected to comply with the JEDEC JESD328 compression attached module standard under the CAMM2 umbrella, and formal standardisation work for LPDDR-based server modules is underway with participation across the memory industry. A published standard lets AMD, Arm server CPU vendors, and hyperscaler custom CPU programmes specify SOCAMM-class memory without NVIDIA-specific engineering, expanding the addressable platform base beyond the single customer that created the category. Custom AI CPU programmes at hyperscalers are the most likely second wave of adoption given their power density constraints mirror Grace.
Warpage and connector reliability in bolt-compressed LPDDR modules remain the dominant quality risk in volume production
SOCAMM2 secures LPDDR5X packages through bolt-based compression, and coefficient of thermal expansion mismatches between package, PCB, and fastening hardware bend modules enough to produce intermittent connector failures. Samsung addressed the issue by lowering soldering process temperatures from above 260 degrees Celsius to below 150 degrees using low temperature solder, but every supplier carries warpage-related yield and screening cost, and field failure rates across million-module fleets remain unproven. A reliability event in a hyperscaler fleet would slow the platform transition regardless of the power efficiency case. These factors substantially limit SOCAMM and SOCAMM2 memory module market growth over the forecast period.
LPDDR5X latency penalties confine SOCAMM to bandwidth-bound AI workloads and exclude the general purpose server memory base
LPDDR5X achieves its bandwidth and power characteristics through design choices that raise access latency versus DDR5, which is why it was historically limited to tightly controlled system designs. Latency-sensitive databases, transactional systems, and general purpose virtualised workloads will remain on DDR5 and MRDIMM roadmaps, capping SOCAMM's addressable share of total server memory. The module competes for AI-attached CPU sockets only, and its growth depends on that socket base expanding rather than on displacing the installed RDIMM estate. These factors substantially limit SOCAMM and SOCAMM2 memory module market growth over the forecast period.
Single-customer specification control concentrates pricing power and roadmap risk with NVIDIA
NVIDIA defines the SOCAMM specification, qualifies the suppliers, and allocates order volumes, a structure that gives the platform owner leverage over module pricing that memory suppliers do not face in JEDEC-governed RDIMM markets. A platform decision to change memory architecture, integrate memory differently, or rebalance allocations between suppliers moves revenue in ways individual suppliers cannot hedge. The JEDEC standardisation track partially mitigates this over time, but through at least 2027 the market operates at the discretion of one buyer. These factors substantially limit SOCAMM and SOCAMM2 memory module market growth over the forecast period.
LPDDR5X wafer allocation conflict between SOCAMM and smartphone customers tightens supply and inflates input costs during the memory shortage
SOCAMM2 consumes advanced 1c and 1-gamma node LPDDR5X wafers that would otherwise supply flagship smartphones, and industry reporting indicates the diversion is tightening LPDDR supply for Chinese handset makers. Samsung and SK Hynix have warned that AI-driven memory shortages could persist until 2027 and beyond, and in a shortage the opportunity cost of SOCAMM allocation is priced into module contracts. Sustained LPDDR price inflation erodes the total cost of ownership case that justifies the platform transition. These factors substantially limit SOCAMM and SOCAMM2 memory module market growth over the forecast period.
SOCAMM2 generation segment is expected to account for a significantly large revenue share in the global SOCAMM and SOCAMM2 memory module market during the forecast period.
Based on generation, the global SOCAMM and SOCAMM2 memory module market is segmented into SOCAMM and SOCAMM2. SOCAMM2 leads because it is the generation specified for the Vera Rubin platform ramp, carries JEDEC standardisation alignment, and is the first generation produced by all three memory suppliers rather than a single source. First generation SOCAMM served GB300 Grace Blackwell deployments primarily through Micron and transitions to a legacy position as Rubin volume replaces Blackwell in new builds.
256GB and above capacity segment is expected to account for a significantly large revenue share in the global SOCAMM and SOCAMM2 memory module market during the forecast period.
Based on capacity, the global SOCAMM and SOCAMM2 memory module market is segmented into 128GB, 192GB, and 256GB and above modules. The 192GB class leads early volume through Samsung and SK Hynix mass production, while the 256GB and above segment is expected to register a rapid revenue growth rate over the forecast period following Micron's March 2026 shipment of 256GB customer samples built on monolithic 32Gb LPDDR5X die, which raises capacity 33 percent in the same footprint and carries the highest revenue per module in the category.
NVIDIA Vera Rubin platform segment is expected to account for a significantly large revenue share in the global SOCAMM and SOCAMM2 memory module market during the forecast period.
Based on platform, the global SOCAMM and SOCAMM2 memory module market is segmented into NVIDIA Grace Blackwell, NVIDIA Vera Rubin, and custom AI CPU platforms. Vera Rubin leads because SOCAMM2 is specified across its rack systems entering volume in the second half of 2026 with orders placed across all three suppliers. The custom AI CPU segment is expected to register a rapid revenue growth rate over the forecast period as JEDEC standardisation lets hyperscaler CPU programmes adopt compression attached LPDDR without NVIDIA-specific engineering.
AI inference server application segment is expected to account for a significantly large revenue share in the global SOCAMM and SOCAMM2 memory module market during the forecast period.
Based on application, the global SOCAMM and SOCAMM2 memory module market is segmented into AI inference servers, AI training nodes, and edge AI systems. AI inference servers lead because the KV cache offload role gives SOCAMM capacity demand that scales with context length, and because inference fleets are the volume deployment tier of the accelerator installed base. The edge AI system segment is expected to register a rapid revenue growth rate over the forecast period as compact inference systems adopt the module for its power envelope and serviceability.
Regional Insights
Asia Pacific market accounted for largest revenue share in the global SOCAMM and SOCAMM2 memory module market in 2025.
Asia Pacific dominates because all module production sits in the region, with SK Hynix producing SOCAMM2 on 1c LPDDR5X in Korea, Samsung producing 192GB modules with low temperature solder processing, and module assembly, connector supply, and test concentrated across Korea, Taiwan, and Japan. Taiwan ODMs including Foxconn, Quanta, and Wistron integrate the modules into Rubin compute trays, placing rack-level consumption in the region as well.
North America market is expected to register significant growth driven by platform specification control and Micron's domestic LPDDR module production.
North America anchors demand because NVIDIA defines the SOCAMM specification and hyperscaler deployment of Grace and Rubin systems concentrates in US data centres. Micron Technology, the only US-headquartered supplier and the first to ship both SOCAMM and 256GB SOCAMM2 samples, produces on its 1-gamma DRAM process with CHIPS Act supported fab expansion in Idaho and New York adding domestic DRAM capacity behind the module roadmap.
Europe market is expected to register steady growth driven by sovereign AI infrastructure and regional cloud deployment of Rubin-class systems.
European SOCAMM demand arrives through EuroHPC AI factory programmes and hyperscaler cloud regions deploying Grace and Rubin racks in Germany, the Netherlands, and the Nordics. The EU AI Continent action plan's gigafactory initiative specifies accelerator systems whose CPU memory tier is SOCAMM-based, and European industrial AI adopters consume the modules indirectly through those deployments rather than through direct procurement.
Middle East market has growing SOCAMM demand through gigawatt-scale AI infrastructure programmes deploying Rubin rack systems.
Saudi Arabia's HUMAIN and UAE's G42 accelerator campuses deploy rack systems that embed SOCAMM2 modules on every Grace and Vera CPU procured, making Gulf sovereign AI programmes a fast-growing consumption channel. The Iran-US conflict and the March 2026 Strait of Hormuz disruption inflated Gulf logistics and insurance costs, but memory modules ship by air freight, so the conflict's effect on SOCAMM supply into regional deployments has been schedule caution and cost inflation rather than physical interruption.
Latin America market has an emerging SOCAMM position through hyperscaler cloud region expansion in Brazil and Mexico.
Latin American exposure is indirect, arriving through cloud regions deploying Grace-based and Rubin-based instances in Sao Paulo and Queretaro. Brazil's data centre buildout, including Scala Data Centers' AI campus programme, creates regional rack demand that embeds SOCAMM modules, while no module production or LPDDR fabrication capacity exists in the region.
Analyst Voice - Field Interview Excerpts
"We spent ten years telling server architects that LPDDR was a mobile technology you could not service in the field. SOCAMM killed that objection with a connector and eight bolts. The thing people underestimate is what it does to our wafer planning, because every terabyte we bolt into a Rubin rack is a terabyte that does not go into a flagship phone, and the phone customers are already calling about allocation."
Nodvolt Analysts
Memory manufacturer, South Korea
Nodvolt analyst note based on the report methodology and supporting source review.
"The warpage problem is the whole ballgame for the next twelve months. Compression attach works beautifully until a module bows eighty microns and a corner row of contacts starts making intermittent contact at temperature. Whoever proves a million-module fleet with boring reliability numbers wins the 2027 allocation conversation, and low temperature solder is how you get there."
Nodvolt Analysts
US hyperscaler infrastructure team
Nodvolt analyst note based on the report methodology and supporting source review.
Strategic Developments
Apr 2026
In April 2026, SK Hynix Inc., South Korea, began mass production of 192GB SOCAMM2 modules built on its sixth generation 1c LPDDR5X process, quantifying more than double the bandwidth of conventional RDIMM with over 75 percent improved energy efficiency and positioning the module as the first optimised for the NVIDIA Vera Rubin platform.
Apr 2026
In April 2026, Samsung Electronics Co. Ltd., South Korea, was reported to have resolved SOCAMM2 module warpage by applying internally developed low temperature solder processing below 150 degrees Celsius, addressing thermal expansion mismatch in the bolt-compressed module structure ahead of its volume ramp.
Mar 2026
In March 2026, Micron Technology Inc., USA, shipped the world's first 256GB SOCAMM2 customer samples built on monolithic 32Gb LPDDR5X die using its 1-gamma process, raising module capacity 33 percent above the 192GB flagship products of its Korean competitors.
Jan 2026
In January 2026, Samsung Electronics Co. Ltd., South Korea, detailed its SOCAMM2 architecture publicly, disclosing more than 70 percent better power efficiency than DDR5 RDIMM, horizontal module orientation for improved cooling integration, and alignment with JEDEC standardisation for compression attached server modules.
Dec 2025
In December 2025, Samsung Electronics Co. Ltd., South Korea, delivered SOCAMM2 samples to NVIDIA Corp., USA, ahead of an early 2026 production ramp, entering the qualification race alongside Micron and SK Hynix.
Sep 2025
In September 2025, NVIDIA Corp., USA, was reported to have placed 2026 SOCAMM module orders with Samsung Electronics, SK Hynix, and Micron Technology with allocations distributed across all three suppliers, aligned with Vera Rubin platform deployment schedules.
Jun 2025
In June 2025, Micron Technology Inc., USA, disclosed volume shipment of first generation SOCAMM modules for NVIDIA GB300 Grace Blackwell systems, establishing the first data centre class modular LPDDR form factor and measuring rack-scale low power DRAM densities exceeding 50 terabytes.
Major Companies
Micron Technology Inc. Samsung Electronics Co. Ltd. SK Hynix Inc. NVIDIA Corp. Foxconn (Hon Hai Precision Industry) Quanta Computer Inc. Wistron Corp. Dell Technologies Inc. Hewlett Packard Enterprise Co. Super Micro Computer Inc. Lenovo Group Ltd. TE Connectivity Ltd. Amphenol Corp. Lotes Co. Ltd. Simmtech Co. Ltd.
Key Questions Answered
What is the SOCAMM and SOCAMM2 memory module market size and forecast through 2035?
The market was USD 1.08 Billion in 2025 and is forecast to reach USD 32.54 Billion by 2035 at a CAGR of 40.6%.
What is SOCAMM2 and how does it differ from RDIMM?
SOCAMM2 packages LPDDR5X into a detachable compression-attached server module, delivering more than double RDIMM bandwidth at over 70 to 75 percent better power efficiency, in a horizontal form factor suited to dense liquid-cooled trays.
Which suppliers are qualified and at what capacities?
Samsung and SK Hynix produce 192GB SOCAMM2 modules, with SK Hynix in mass production from April 2026 on 1c LPDDR5X, and Micron sampled 256GB modules in March 2026 on monolithic 32Gb 1-gamma die.
What platform demand anchors the market?
NVIDIA specifies SOCAMM2 across Vera Rubin rack systems and reportedly split 2026 orders across all three suppliers, with first generation SOCAMM already deployed in GB300 Grace Blackwell systems.
What is the main technical risk?
Module warpage from thermal expansion mismatch in the bolt-compressed structure, which produces intermittent connector failures; Samsung addressed it with low temperature solder processing below 150 degrees Celsius.
Does SOCAMM displace general purpose server memory?
No. LPDDR5X carries higher access latency than DDR5, confining SOCAMM to bandwidth-bound AI workloads while databases and general purpose platforms remain on DDR5 RDIMM and MRDIMM roadmaps.
Scope of Research
Generation
SOCAMM (Gen1, GB300)
SOCAMM2 (Rubin platform)
JEDEC CAMM2-aligned
Capacity
128GB
192GB (Samsung, SK Hynix)
256GB and above (Micron)
Platform
NVIDIA Grace Blackwell
NVIDIA Vera Rubin
Custom AI CPU
Geography
North America
Europe
Asia Pacific
Latin America
Middle East & Africa
Table of Contents
Ch. 1 Executive Summary
  • Market overview and LPDDR-to-server transition analysis
  • Three-supplier qualification race
Ch. 2 Market Sizing & Forecast
  • 2025 baseline and 2026-2035 projections
  • Revenue by generation, capacity, and platform
Ch. 3 Technology Analysis
  • Compression attach architecture and warpage engineering
  • LPDDR5X latency trade-offs vs DDR5
Ch. 4 Platform Deep Dive
  • Vera Rubin CPU memory architecture
  • KV cache offload and inference memory tiering
Ch. 5 Segment Analysis
  • By generation, capacity, and application
  • Smartphone LPDDR allocation conflict
Ch. 6 Regional Analysis
  • Asia Pacific production concentration
  • Gulf sovereign AI deployment demand
Ch. 7 Competitive Analysis
  • 15 company profiles and module roadmaps
  • Micron vs Samsung vs SK Hynix positioning
Ch. 8 Primary Research
  • Interview panel - 20 memory and platform executives
  • Methodology and data validation