2026 Computex Focuses on Agentic AI and Edge Computing

Overview

This report is a summary of key points from J.P. Morgan’s coverage of the 2026 Taipei International Computer Exhibition (Computex) and concurrent NVIDIA GTC conference, focusing on the development path of agentic AI technology and its profound impact on chips, PCs, and data center infrastructure. The report indicates that AI computing is evolving from pure cloud to a collaborative 'edge + cloud' model, with NVIDIA and Qualcomm each leveraging new hardware and ecosystem strategies to gain competitive advantage.

Core views

At the event, NVIDIA unveiled its first AI processor for Windows PCs, the N1X (also known as RTX Spark), jointly developed with MediaTek. This chip delivers 1 petaflop of computing power and can run 120 billion parameter large language models (LLMs) locally, supporting million-level context windows. This move aims to bring agentic AI capabilities down to client devices. However, the report believes its success hinges on two critical factors: first, whether the application ecosystem restructured for agentic AI can mature quickly, and second, compatibility with traditional x86 applications — a major reason why Qualcomm’s ARM PC initiative previously faced challenges. Although Qualcomm did not launch new products during the keynote session, its CEO emphasized that agentic AI will reshape personal device application development paradigms, forecasting a 40-fold increase in AI token consumption from 2026 to 2030. Qualcomm expects AI workloads to dynamically distribute between device and cloud, potentially triggering a new cycle of edge device replacements. Notably, certain tasks such as programming or web generation could reduce token usage by 30%-60% and improve speed under a device-cloud hybrid architecture. On the data center side, NVIDIA confirmed that the Vera Rubin platform has entered full mass production, with Microsoft and Dell/CoreWeave already deploying engineering sample cabinets. Thanks to liquid cooling and mid-board PCB design, cabinet assembly time has been reduced from 2 hours in the Blackwell era to just 5 minutes. In 2026, limited by HBM4 supply and CoWoS-L packaging capacity, initial shipments are expected to reach about 2 million chips, corresponding to approximately 10,000 VR200 NVL 72 racks. Additionally, NVIDIA is aggressively promoting its Vera CPU specifically designed for agentic AI, claiming it offers 1.8x better performance in AI sandbox tasks compared to x86 and 3x faster SQL database processing. The company plans to sell standalone Vera CPUs as a new growth engine, expecting 600,000 units shipped in 2026 and over 3 million in 2027. TSMC, Amkor, and ASE will provide packaging support, while Hon Hai and Quanta are primary cabinet suppliers.

Analysis framework

The report adopts a three-tier analytical framework: 'Technology Trends – Product Implementation – Supply Chain Impact'. First, it identifies agentic AI as the core paradigm of next-generation AI applications. Second, it breaks down NVIDIA and Qualcomm’s hardware and software strategies across edge (PC/mobile) and cloud (CPU/GPU/cabinets). Finally, it evaluates the pull effect on upstream semiconductor manufacturing (TSMC), packaging/test (ASE), thermal management (AVC, Fositek), and power supply (Delta) sectors in Asia. The report particularly focuses on key bottlenecks in technology implementation, such as x86 compatibility, HBM4 supply, and CoWoS packaging capacity, supported by supply chain validations (e.g., ByteDance and a U.S. hyperscale cloud provider as Qualcomm ASIC customers) to strengthen credibility. It also highlights Apple WWDC as the next important observation window, given its potential to further fuel the resurgence of edge AI computing.

Methodology notes

• Industry/Industrial Analysis FrameworkSupply-demand framework

  Supply and Demand Analysis of AI Chip Industry

  The report analyzes the surge in demand for AI accelerators (GPUs) and new data center CPUs, combined with key supply constraints like HBM memory and advanced packaging (CoWoS), to predict that supply chain tightness will persist through 2027, demonstrating typical application of the supply-demand framework.

• Industry/Industrial Analysis FrameworkUpstream-Midstream-Downstream Transmission

  Transmission Effect of AI Technological Evolution on Hardware Industry Chain

  Starting from the top-level demand for agentic AI, the report systematically drills down to chip design (NVIDIA, Qualcomm), manufacturing and packaging (TSMC, ASE), and thermal/power components (AVC, Delta), clearly illustrating how technological change propagates through the industrial chain and creates investment opportunities.

• Industry/Industrial Analysis FrameworkVolume-price decomposition

  Potential Pull on PC Shipments from Edge AI

  The report decomposes the potential recovery of the PC market into whether edge AI adoption brings incremental shipments, i.e., whether AI can serve as a new catalyst for device upgrades—a classic volume (shipment) dimension analysis rather than price or ASP changes.

Asset mapping & comparison

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

• MediaTek Inc. (2454.TW)

  Benefiting from NVIDIA N1X Collaboration Project

  Strengths

  Generating royalty income via ARM CPU and connectivity chips, and entering the PC computing market

  Weaknesses

  Current contribution limited, only 1-2% of revenue

  Comparison

  Compared to Qualcomm's early ARM PC efforts, this collaboration with NVIDIA offers stronger ecosystem synergy

  Risks

  Product success depends on x86 compatibility and application ecosystem reconstruction

• ASE Technology Holding Co Ltd (3711.TW)

  Benefiting from Vera CPU Packaging Demand

  Strengths

  Potential OSAT supplier participating in high-end CPU packaging

  Comparison

  Sharing new TAM with TSMC and Amkor

• Asia Vital Components (3017.TW) / Fositek (6805.TW)

  Benefiting from CPU Cabinet Liquid Cooling Demand

  Strengths

  CPU cabinets require more cold plates and quick-connect fittings (QDs)

• Delta Electronics, Inc. (2308.TW)

  Benefiting from DSX Platform Power Optimization Needs

  Strengths

  Supplier of power solutions in AI factory infrastructure construction

Key data

• N1X Processor AI Compute Power1 petaflopsSupports local execution of 120B parameter LLM
• Vera CPU vs x86 Performance1.8xAI sandbox performance improvement
• Vera Rubin Cabinet Assembly Time5 minutesSignificantly shortened from 2 hours of Blackwell
• 2026 Vera Rubin Initial Shipments~2M unitsCorresponding to ~10k VR200 NVL 72 racks
• 2026–2030 Token Consumption Growth40xPrediction by Qualcomm CEO

Impact & implications

The report suggests that the rise of agentic AI will reshape computing architectures, pushing device-cloud collaboration into mainstream, thereby creating new upgrade cycles for edge devices such as PCs and smartphones. On the data center side, dedicated CPUs (like Vera CPU) will become a new TAM (Total Addressable Market) following AI accelerators, driving overall semiconductor ecosystem growth. Infrastructure suppliers specializing in liquid cooling and efficient power solutions will also benefit from the sharp rise in AI factory capital expenditure (estimated at $800–$1000 per GW).

Risks

• The N1X processor faces challenges in x86 application compatibility, which may delay the adoption of agentic AI on Windows PCs
• Vera Rubin production constrained by HBM4 supply and CoWoS-L packaging capacity, potentially resulting in lower-than-expected deliveries in 2026
• The pace of application ecosystem restructuring for agentic AI lags behind expectations, delaying the edge AI device replacement cycle

What to watch

• Apple WWDC announcement regarding device-side AI
• Qualcomm Dragonfly AI server rack official release at the analyst day in late June
• Actual shipment ramp-up of Vera CPU in H2 2026
• Adoption progress of Qualcomm inference ASIC by ByteDance and a U.S. hyperscale cloud provider