Computex 2026 reinforces the boost from Personal AI, AI CPU, and Physical AI to memory demand
AI summary card
Computex 2026 reinforces the boost from Personal AI, AI CPU, and Physical AI to memory demand
Citigroup believes that RTX Spark, DGX Station, Vera CPU, and Cosmos 3 showcased by NVIDIA at Computex 2026 will drive AI inference from centralized GPU computing toward CPU and distributed computing, thereby benefiting DRAM, server DDR5, SoCAMM2, and Korean memory suppliers.
- NVIDIA's RTX Spark is equipped with 128GB LPDDR5X and can run models with about 120 billion parameters, with DRAM capacity more than 10 times that of a traditional roughly 12GB notebook.
- DGX Station for Windows can be configured with up to 748GB of memory and 20 petaFLOPS FP4 compute, enabling it to run AI models with up to 1 trillion parameters, with memory capacity above the average level of about 600GB DRAM in typical servers.
- Vera CPU has entered full-scale mass production and targets workloads such as agentic AI, reinforcement learning, and data processing, with task completion speed about 1.8 times that of x86 processors, and is expected to drive demand for server DDR5 and SoCAMM2.
- As an open foundation model for Physical AI, Cosmos 3 can understand and generate text, images, video, sound, and actions, shortening training and evaluation cycles for robotics and autonomous driving.
- Citigroup assigns an NVDA target price of US$300, based on about 28x C27E P/E; Samsung Electronics has a 12-month target price of W460,000, using the SOTP method.
Report interpretation
Overview
This report is Citigroup's global semiconductor flash note following Computex 2026, focusing on NVIDIA's product updates in Personal AI, AI CPU, and Physical AI, as well as the potential impact of these updates on memory demand and Korean memory suppliers. The report argues that AI inference demand is expanding from centralized cloud GPU computing to local devices, CPUs, and edge physical AI scenarios, and that demand for DRAM, server DDR5, and SoCAMM2 is likely to receive structural support.
Core views
Citigroup's core view is that NVIDIA's new-generation Personal AI PCs and desktop AI systems significantly increase per-device memory capacity, Vera CPU brings AI inference demand to the CPU side, and Cosmos 3 strengthens Physical AI scenarios such as robotics, autonomous driving, and edge inference. As AI memory demand expands from centralized computing to distributed computing, Korean memory suppliers such as Samsung Electronics are set to benefit more clearly.
Analysis framework
Using the product launches at Computex 2026 as the starting point, the report maps changes in NVIDIA hardware specifications to downstream memory demand and then evaluates NVDA and Samsung Electronics within their respective valuation frameworks. NVDA valuation uses about 28x C27E P/E and discounts back to present value; Samsung Electronics uses a 2026E EBITDA-based SOTP approach, assigning peer-comparable EV/EBITDA multiples separately to its five main divisions: Memory, Foundry, Display Panel, Mobile, and Consumer Electronics.
Methodology notes
Infers the demand elasticity for DRAM, server DDR5, and SoCAMM2 from NVIDIA's Personal AI, AI CPU, and Physical AI product specifications.
The report compares the memory capacity differences between RTX Spark, DGX Station, and traditional PCs and general-purpose servers, and views the expansion of Vera CPU and Physical AI inference demand as a source of structural growth for the memory industry.
NVDA target price is based on about 28x C27E earnings power.
Citigroup states that NVDA's US$300 target price is based on about 28x C27E EPS of about US$11.9 (including SBC) and discounted back to present value, with the multiple in line with the three-year average.
Samsung Electronics target price uses a sum-of-the-parts valuation.
Samsung Electronics' 12-month target price of W460,000 is based on 2026E EBITDA, applying EV/EBITDA multiples of 7.9x, 4.1x, 0.5x, 4.8x, and 2.0x to Memory, Foundry, Display Panel, Mobile, and Consumer Electronics, respectively.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- NVIDIA Corp (NVDA.O)Core product launch provider and leader in AI hardware demand.
- Strengths
- RTX Spark, DGX Station, Vera CPU, and Cosmos 3 span Personal AI, AI CPU, and Physical AI, driving AI inference from the cloud into local and edge scenarios.
- Weaknesses
- Valuation realization depends on the pace of adoption of new platforms, continued growth in data center and gaming businesses, and a stable competitive landscape.
- Comparison
- RTX Spark's 128GB DRAM is more than 10 times that of a traditional notebook with about 12GB; DGX Station's 748GB memory is above the average capacity of about 600GB DRAM in typical servers.
- Risks
- Gaming competition leading to share loss, slower-than-expected adoption of new platforms, volatility in automotive and data center markets, and the impact of crypto mining on gaming sales.
- Samsung Electronics (005930.KS)Korean memory supplier and potential beneficiary of AI memory demand.
- Strengths
- Expansion in demand for server DDR5, SoCAMM2, DRAM, and HBM is expected to support the memory business, and the SOTP valuation target price indicates upside potential.
- Weaknesses
- HBM customer shipment approvals, PC and NAND demand, competition in memory and foundry, and mobile margins remain key constraints.
- Comparison
- In the SOTP, the Memory division is assigned 7.9x EV/EBITDA, higher than Foundry at 4.1x, Display Panel at 0.5x, Mobile at 4.8x, and Consumer Electronics at 2.0x.
- Risks
- Delays in shipment approvals of HBM to key customers, weaker-than-expected PC sales, NAND demand missing expectations, aggressive investment by competitors depressing prices, intensified competition in the handset market, and a sharp appreciation of the Korean won.
- DRAM / server DDR5 / SoCAMM2Memory categories that directly benefit from the expansion of AI PCs, AI CPUs, and distributed AI inference.
- Strengths
- Higher per-device memory capacity, growth in CPU-side AI inference, and edge Physical AI deployment together support demand.
- Weaknesses
- Demand realization depends on the commercialization speed of AI PCs and new CPU platforms, as well as end-market PC and server procurement cycles.
- Comparison
- The report compares the memory capacity of RTX Spark and DGX Station with traditional notebooks and general-purpose servers, showing the step-up in DRAM content in AI devices.
- Risks
- If AI platform adoption is slower than expected, or if PC and NAND demand remain weak, memory demand and price elasticity may be lower than expected.
Key data
- Report date2026-06-02 06:29:35 ETThe publication time shown on the report cover page.
- RTX Spark memory128GB LPDDR5XMore than 10 times higher than a traditional notebook with about 12GB DRAM, and can run AI models with about 120 billion parameters.
- RTX Spark performanceUp to 1 petaflop FP4; 600GB/s NVLink-C2C bandwidthUsed to run large AI models and agents locally.
- DGX Station memoryUp to 748GB (252GB HBM3e + 496GB LPDDR5X)Higher than the average capacity of about 600GB DRAM in typical servers, and can run models with up to 1 trillion parameters.
- DGX Station computeUp to 20 petaFLOPS FP4Designed for Windows desktop AI workstations.
- Vera CPU specifications88-core Olympus; up to 1.2TB/s LPDDR5X memory bandwidth; up to 1.8TB/s CPU-GPU bandwidthThe report says its task completion speed is about 1.8 times that of x86 processors across multiple AI and data-processing workloads.
- NVDA valuationCurrent price US$224.36; target price US$300; rating Buy (1)The target price is based on about 28x C27E P/E, implying about 33.7% upside.
- Samsung Electronics valuationCurrent price W360,500; target price W460,000; rating Buy (1)The 12-month target price uses the SOTP method, implying about 27.6% upside.
Impact & implications
The investment implication of the report is that AI inference demand is no longer concentrated only in GPUs and cloud data centers, but is expanding to local PCs, desktop workstations, CPU hosts, and edge Physical AI devices. The significant increase in memory capacity per device will lift DRAM demand, Vera CPU will drive demand for server DDR5 and SoCAMM2, and Physical AI will extend the structural growth cycle of the memory industry. Citigroup believes this will make Korean memory suppliers, especially Samsung Electronics, more notable beneficiaries.
Risks
- Intensifying competition in NVIDIA's gaming business may pressure market share and stock price.
- Slower-than-expected adoption of new AI platforms may drag on data center and gaming sales.
- Volatility in automotive and data center market orders may increase stock and valuation multiple volatility.
- Changes in crypto mining-related demand may affect NVIDIA gaming sales.
- If Samsung Electronics' HBM shipment approvals to key customers are delayed, target price realization may be constrained.
- Weaker-than-forecast PC sales or NAND demand below expectations would weaken the strength of the memory recovery.
- Aggressive investment by competitors in memory semiconductors and foundry may depress prices.
- Intensified competition in the handset market may compress Samsung Electronics' mobile business margins.
- A sharp appreciation of the Korean won may negatively affect Samsung Electronics earnings.
What to watch
- The actual adoption pace of RTX Spark and DGX Station in local AI model and AI agent scenarios.
- The pull on server DDR5 and SoCAMM2 orders after Vera CPU mass production.
- The rollout pace of Cosmos 3 in robotics, autonomous driving, and edge inference.
- Whether AI memory demand continues to expand from centralized cloud computing to distributed device-side computing.
- Progress in Samsung Electronics' HBM key customer approvals and shipments.
- Price trends for PC, NAND, and DRAM, as well as ASP momentum for Korean memory suppliers.
- Whether NVDA's roughly 28x C27E P/E valuation assumption and Samsung Electronics' segment EV/EBITDA multiples remain aligned with peer trading multiples.