Computex 2026 reinforces the themes of personal AI, AI CPU, and physical AI, benefiting AI memory demand

Overview

This report is Citigroup's Flash commentary on the global semiconductor theme at Computex 2026, focusing on NVIDIA's product updates in personal AI, AI CPU, and physical AI. The report believes that NVIDIA's Windows AI PC products, Vera CPU, and Cosmos 3 show that AI inference is expanding from cloud GPUs to local devices, CPUs, the edge, and the physical world, which will significantly increase demand for memory such as DRAM, Server DDR5, and SoCAMM2.

Core views

The core views are: first, localized inference on personal AI devices will significantly increase PC memory capacity; second, Vera CPU will extend AI inference demand from GPUs to CPUs, strengthening demand for server DDR5 and SoCAMM2; third, physical AI will push robotics, autonomous driving, and edge inference into a new stage; fourth, as AI memory demand shifts from centralized computing to distributed computing, Korean memory suppliers such as Samsung Electronics stand out as greater beneficiaries.

Analysis framework

The report uses event-driven analysis and supply-chain demand transmission analysis: it first reviews the key specifications of RTX Spark, DGX Station, Vera CPU, and Cosmos 3 announced by NVIDIA at Computex 2026, then maps these hardware changes to demand for memory capacity, bandwidth, and product form factors, and finally explains the impact on related assets by combining the target prices, ratings, and valuation methods for NVIDIA and Samsung Electronics.

Methodology notes

• 事件驱动研究Computex 2026产品发布解读

  Infer changes in supply-chain demand from key product specifications

  Through NVIDIA's personal AI PC, AI CPU, and physical AI models launched at Computex 2026, the report concludes that AI inference will spread to local, CPU, and edge environments, thereby increasing demand for memory capacity and bandwidth.

• 股票估值P/E估值法

  NVIDIA target price is based on approximately 28x P/E

  NVIDIA Corp's US$300 target price is based on C27E earnings power of approximately US$11.9 and discounted using approximately 28x P/E, in line with its three-year average multiple.

• 股票估值SOTP分部估值法

  Samsung Electronics target price is based on sum-of-the-parts valuation using 2026E EBITDA

  Samsung Electronics' W460,000 12-month target price uses the SOTP method, assigning different EV/EBITDA multiples to the five major divisions of Memory, Foundry, Display Panel, Mobile, and Consumer Electronics with reference to global peers.

• 评级框架Citi Research Fundamental Research Investment Ratings

  Buy (1), Neutral (2), and Sell (3) are based on 12-month expected total return and risk

  Citigroup defines Buy (1) as typically corresponding to an expected total return of no less than 15% over the next 12 months, with the threshold for high-risk stocks being no less than 25%; target prices usually correspond to a 12-month horizon.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• NVIDIA Corp (NVDA.O)

  Core product launch provider and driver of AI hardware demand

  Strengths

  RTX Spark, DGX Station, Vera CPU, and Cosmos 3 cover multiple directions including personal AI, AI CPU, and physical AI, reinforcing NVIDIA's platform position in the AI hardware ecosystem.

  Weaknesses

  The pace of adoption for new platforms still needs to be validated, and demand in data centers, gaming, and automotive remains volatile.

  Comparison

  Compared with traditional PCs, RTX Spark has more than 10 times the memory capacity; compared with x86 processors, Vera CPU completes tasks about 1.8 times faster in relevant workloads.

  Risks

  Competition in gaming leading to share loss, slower-than-expected adoption of new platforms, volatility in automotive and data center markets, and the impact of cryptomining on gaming sales.

• Samsung Electronics (005930.KS)

  Korean memory supplier and potential beneficiary of AI memory demand

  Strengths

  As AI inference expands from GPUs to CPUs and distributed devices, demand related to Server DDR5, SoCAMM2, DRAM, and HBM is expected to strengthen.

  Weaknesses

  Affected by HBM customer certification, PC and NAND demand, memory pricing cycles, competition in the handset business, and exchange rates.

  Comparison

  The report believes that as AI memory demand expands from centralized computing to distributed computing, Korean memory suppliers will benefit more clearly.

  Risks

  Delays in shipment approval of HBM to key customers, weaker-than-expected PC sales, NAND demand below expectations, aggressive investment by competitors in memory/foundry, intensifying smartphone competition, and sharp appreciation of the Korean won.

• DRAM / Server DDR5 / SoCAMM2

  Direct demand categories from the spread of AI CPU and personal AI devices

  Strengths

  Local AI model execution, high-speed CPU-GPU interconnects, and edge inference all require higher memory capacity and bandwidth.

  Weaknesses

  Demand realization depends on the rollout speed of endpoint AI PCs, server CPU platforms, and physical AI applications.

  Comparison

  A traditional laptop has about 12GB of DRAM, while RTX Spark has 128GB; DGX Station's 748GB of memory is above the average DRAM capacity of a typical server at about 600GB.

  Risks

  Endpoint device shipments falling short of expectations, slower-than-expected AI application adoption, and overly rapid expansion of memory supply suppressing prices.

Key data

• Report date2026-06-02 06:29:35 ETTime disclosed on the report cover page.
• RTX Spark memory and compute128GB LPDDR5X; up to 1 petaflop FP4; 600 GB/s NVLink-C2CCan run AI models with approximately 120 billion parameters, with DRAM capacity more than 10 times higher than a traditional roughly 12GB laptop.
• DGX Station memory and computeUp to 748GB memory; up to 20 petaFLOPS FP4Memory includes 252GB HBM3e and 496GB LPDDR5X, capable of running models with up to 1 trillion parameters.
• Vera CPU specifications and performance88 cores; up to 1.2TB/s memory bandwidth; up to 1.8TB/s CPU-GPU bandwidth; task completion speed about 1.8x that of x86Vera CPU is already in full mass production, and Citigroup expects it to drive demand for Server DDR5 and SoCAMM2.
• Cosmos 3 positioningOpen physical AI foundation model, available in Super, Nano, and Edge tiersCan process text, images, video, sound, and actions, with the goal of shortening physical AI training and evaluation cycles from months to days.
• NVIDIA Corp valuationTarget price US$300; previous close US$224.36; implied upside approximately 33.7%Based on approximately 28x P/E and C27E earnings power of approximately US$11.9.
• Samsung Electronics valuationTarget price W460,000; previous close W360,500; implied upside approximately 27.6%Uses an SOTP valuation method based on 2026E EBITDA.
• Samsung Electronics SOTP multiplesMemory 7.9x; Foundry 4.1x; Display Panel 0.5x; Mobile 4.8x; Consumer Electronics 2.0xMultiples are referenced to trading multiples of relevant global peers.

Impact & implications

The investment implication is that AI hardware demand is no longer concentrated only in cloud GPU clusters, but is spreading to local PCs, desktop workstations, CPU hosts, edge inference, and physical AI systems. As models run across different endpoint and server layers, demand for memory capacity, bandwidth, and advanced packaging rises in parallel, providing stronger structural support for the supply chains of DRAM, Server DDR5, SoCAMM2, and HBM.

Risks

• NVIDIA faces competition in gaming, and loss of market share could weigh on its stock price.
• Slower-than-expected adoption of new platforms could lead to weaker-than-expected data center and gaming sales.
• Order volatility in the automotive and data center markets could increase valuation multiple volatility.
• The impact of cryptomining on gaming sales may create uncertainty.
• If Samsung Electronics' HBM shipment approvals to key customers are delayed, target price realization may be affected.
• Weaker-than-expected PC sales or NAND demand below expectations would weaken memory demand and earnings elasticity.
• Aggressive investment by competitors in memory semiconductors and foundry may negatively affect pricing.
• Intensifying competition in the smartphone market and sharp appreciation of the Korean won may compress Samsung Electronics' earnings.

What to watch

• The actual adoption pace of RTX Spark and DGX Station in personal AI and local AI model execution scenarios.
• The strength of demand pull for Server DDR5 and SoCAMM2 after Vera CPU enters mass production.
• Developer ecosystem and commercialization progress of Cosmos 3 in robotics, autonomous driving, and edge inference.
• Whether AI memory demand continues to expand from centralized cloud computing to distributed endpoints and edge devices.
• The certification and shipment pace of Samsung Electronics HBM to key customers.
• Price and inventory cycle changes for DRAM, Server DDR5, SoCAMM2, and HBM.