NVDA and Qualcomm keynotes point to a new cycle for agentic AI, physical AI, and data centers

Overview

This report is Citi's summary of views on the NVIDIA GTC and Qualcomm Computex keynotes, focusing on the two companies' technology roadmaps in data center AI infrastructure, agentic AI, physical AI, edge devices, robotics, and 6G, as well as the potential implications of these changes for China's technology and communications industry chain. The report believes the industry is shifting from the foundational LLM stage driven by static prompts to the stage of agentic AI and physical AI.

Core views

The core views include: first, Vera Rubin entering full commercial mass production is a positive signal for NVDA supply chain ramp-up in 2H26; second, NVDA is transforming from a GPU acceleration company into a full-stack AI infrastructure company and may expand its control over racks, clusters, power, and liquid cooling chains; third, Vera CPU and RTX Spark strengthen agentic AI capabilities in local devices and data centers, with positive implications for ecosystem partners such as Lenovo; fourth, Qualcomm Dragonfly may become an incremental source for China ASIC data center solutions; fifth, the SOCAMM/LPDDR5X route may create long-term pressure on demand for memory interface chips tied to Montage's RDIMM-related opportunities.

Analysis framework

The report adopts an event interpretation and supply chain mapping approach, starting from the products, architectures, and ecosystem collaboration information disclosed in the NVDA and Qualcomm keynotes, and inferring the potential impacts on servers, AI PCs, data center ASICs, automotive, robotics, 6G, and China's semiconductor supply chain.

Methodology notes

• Event catalyst interpretationKeynote takeaway mapping

  Extract product cadence and industry chain implications from the GTC and Computex keynotes.

  The report maps public keynote information such as Vera Rubin mass production, Vera CPU, RTX Spark, Dragonfly, and the Snapdragon AI ecosystem to supply chain ramp-up, product ASP, edge AI proliferation, and potential demand substitution.

• Supply chain look-throughAI infrastructure supply chain mapping

  Infer the benefiting or pressured links from changes in chip platforms and system architecture.

  The report connects NVDA's extension from GPU systems to full-stack AI infrastructure to opportunities in L10 servers, L11 racks, L12 clusters, power, cooling, memory architecture, and related Chinese suppliers.

• Technology architecture analysisAgentic AI and physical AI framework

  Use changes in AI workload forms to explain demand for data centers and edge devices.

  The report believes the industry is moving from the static Q&A stage of foundational LLMs into the stages of agentic AI and physical AI, so CPUs, GPUs, local AI PCs, automobiles, robots, and 6G networks may all become carriers of incremental computing demand.

Asset mapping & comparison

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

• NVIDIA CORP / NVDA.US

  Core keynote source and leader of AI infrastructure platforms.

  Strengths

  Vera Rubin mass production, full-stack AI infrastructure positioning, Vera CPU, RTX Spark, and AI factory reference designs strengthen its data center and edge AI ecosystem.

  Weaknesses

  The report does not provide a rating or target price, and the investment conclusion mainly comes from keynote look-through rather than a complete valuation framework.

  Comparison

  Upgrading from a GPU-accelerated systems company to a full-stack AI infrastructure company spanning chips, racks, clusters, networking, storage, power, and cooling.

  Risks

  The 2H26 ramp-up pace, supply chain execution, power and cooling requirements, architecture transitions, and customer capex timing may affect realization.

• QUALCOMM INC / QCOM.US

  Keynote source, expanding from edge chips into data centers and the AI agent execution layer.

  Strengths

  Dragonfly enters the data center, and Snapdragon becomes a native execution platform for third-party AI agents, Android, Windows, automotive, robotics, and 6G scenarios.

  Weaknesses

  Further details on Dragonfly are still pending disclosure on June 24, and its data center competitiveness and commercialization path still require validation.

  Comparison

  Compared with NVDA's focus on AI factories and data center infrastructure, Qualcomm places more emphasis on a continuous compute pipeline from low-power edge devices to back-end infrastructure.

  Risks

  There is uncertainty around data center product rollout, ecosystem partner adoption, competition with existing GPU/ASIC solutions, and the monetization pace of edge AI.

• Lenovo

  NVDA ecosystem partner and potential beneficiary.

  Strengths

  It may benefit from the new GPU server platform and the higher ASP contribution from Nvidia Windows-based laptops.

  Weaknesses

  The degree of benefit depends on specific product shipments, channel demand, and the implementation of NVDA ecosystem cooperation.

  Comparison

  Compared with pure chip supply chain players, Lenovo's opportunity is more tilted toward complete servers, AI PCs, and ecosystem end devices.

  Risks

  AI PC demand realization, server competition, sustainability of ASP improvement, and supply chain delivery capability need to be tracked.

• FII; VGT; WUS; TFC

  NVDA supply chain-related beneficiary links mentioned in the report.

  Strengths

  Vera Rubin full commercial mass production is viewed as a positive signal confirming supply chain ramp-up in 2H26, and TFC was prominently displayed twice in the video presentation.

  Weaknesses

  The report does not elaborate on each company's orders, capacity, or earnings elasticity, and is more of a supply chain look-through.

  Comparison

  These assets are more directly exposed to NVDA hardware platform ramp-up rather than Qualcomm's edge ecosystem expansion.

  Risks

  NVDA platform timing, customer qualification, capacity allocation, pricing pressure, and technical specification changes may affect revenue realization.

• Montage

  A memory interface chip-related company potentially under pressure.

  Strengths

  It still holds a position in the server memory interface chip-related industry chain.

  Weaknesses

  The report notes that the Vera CPU and Spark platform use SOCAMM/LPDDR5X rather than RDIMM, which may weaken opportunities for DIMM chips.

  Comparison

  Unlike companies benefiting from the NVDA supply chain, Montage's risk comes from memory architecture migration potentially changing the original demand structure.

  Risks

  If Nvidia CPUs gain a larger share in server processors, RDIMM-related memory interface demand may be eroded.

• China ASIC data center solutions

  May benefit from the incremental source brought by Qualcomm Dragonfly.

  Strengths

  Dragonfly may supplement the supply options for China data center ASIC solutions and form a continuous compute path with Qualcomm's edge AI ecosystem.

  Weaknesses

  Current disclosure is limited, and specific specifications, customers, supply capability, and business model remain unclear.

  Comparison

  Compared with NVDA's high-performance AI factory route, Qualcomm's solution may place more emphasis on coordination between low-power edge devices and back-end infrastructure.

  Risks

  Product details, performance validation, ecosystem compatibility, and the pace of adoption in the China market still require follow-up confirmation.

Key data

• Report date2026-06-01The page shows the publication time as 01 Jun 2026 13:05:39 ET.
• Report length10 pagesThe main investment views are concentrated on pages 1 to 3, with the remainder mainly consisting of disclosures and regulatory statements.
• Vera Rubin statusFull commercial mass productionThe report believes this likely confirms NVDA supply chain ramp-up in 2H26.
• Vera Rubin technology node3nm, seven dedicated chips, HBM4Used for the next-generation multi-rack AI infrastructure platform, with upgrades targeted at memory bandwidth bottlenecks.
• System interconnect and scale-outNVLink 72, CX9 SuperNICs, BlueField-4 DPUsSupports large-scale scale-out and scale-up capabilities.
• Power supply and coolingMore than 5,000 amps, 45°C liquid cooling loopReflects the Rubin architecture's requirements for high-power AI factory infrastructure.
• Vera CPU specifications88 cores, LPDDR5, 1.2TB/s memory bandwidth, 10 instructions decoded per cycleThe report says this CPU is designed for autonomous AI workloads and agentic tool-calling pipelines.
• Vera CPU performance description3x faster on complex SQL tasks, 6x faster on real-time data stream processingThe comparison is against corresponding x86 solutions and reflects architecture benchmark descriptions cited by the report.
• RTX Spark platform1 petaflop local AI computing power, 128GB unified memoryLaunched in cooperation with Microsoft for Windows-compatible, CUDA-enabled local agentic devices.
• Qualcomm DragonflyA new product brand for data center environmentsMore details are expected to be disclosed on June 24.

Impact & implications

In terms of investment implications, the NVDA supply chain, AI servers, liquid cooling, power, Lenovo AI servers, and the AI PC ecosystem may benefit from Vera Rubin and RTX Spark; Qualcomm's Dragonfly and Snapdragon AI ecosystem expansion may increase the range of options for China ASIC data center solutions and edge AI; however, the shift in memory architecture from RDIMM to SOCAMM/LPDDR5X may create long-term risk for demand for Montage-related memory interface chips.

Risks

• Keynote disclosures are not equivalent to order or revenue confirmation, and supply chain benefits require subsequent validation through shipments, capacity, and customer adoption.
• If Vera Rubin's 2H26 ramp-up pace is delayed, the elasticity of related supply chains may be lower than expected.
• NVDA's extension of control to L11 racks and L12 clusters may change supply chain bargaining power and profit distribution.
• The trend of SOCAMM/LPDDR5X replacing RDIMM may compress opportunities for some memory interface chips.
• Qualcomm Dragonfly disclosure is still incomplete, and there is uncertainty around the competitiveness and commercialization path of its data center products.
• The proliferation of agentic AI and physical AI in China's edge devices, automotive, robotics, and 6G scenarios may be slower than expected.

What to watch

• Vera Rubin's mass-production ramp-up in 2H26, supply chain orders, and capacity allocation.
• Whether NVDA further extends supply chain control from L10 servers to L11 racks and L12 clusters.
• Lenovo's shipments and ASP contribution in GPU server platforms and Nvidia Windows-based laptops.
• The adoption pace of SOCAMM/LPDDR5X and its impact on RDIMM and Montage memory interface chip opportunities.
• The Dragonfly specifications, customers, mass-production plan, and China data center application scenarios to be disclosed by Qualcomm on June 24.
• The progress of Snapdragon's native execution ecosystem in third-party AI agents, Android, Windows, automotive, robotics, and 6G.